 

 

 

 

 

 

 

 

 

.‘A.

”4.111;
4'

.‘(H

 

A
,'. '1'“,

$.33??? '5’

. .

, .

1‘5"» 1". ‘g ’JI‘Ch' L.‘
m #433: 3.33::

('3;

I 'vk-ni.’

 

V. .
‘ \ ‘-
J!‘ 1‘.

5., 9.1.4

4

~ wlvi‘lm“ "" V ‘
2‘3? 5.: 4
.4 . «1.4 .
.7 21“ $5; . ”$23.?
. u 5"! ‘ "~ “~“
3&1?“ mmh’ ’W 9'33“};
:\ ahv§127 IL} ”(wuﬁiq
‘ i" “mu.
‘5‘i ‘1::j~”,:..O1:V\-I:wh-~“”-?“.
'1 -‘ .- I‘ . ' “~ ‘~ 45‘~
3.63125:- (“L‘W‘W'a- 2"“
'w' A ‘ 333v} 3. “ ‘u‘ 'JVOQ‘V‘Q'A
.: j. “UR-u 4 .. ~ 5». s K's)" :33:-
.\ Emu» “ "4.7" .-
'v.‘ .44- . “2:33:

_ x
‘ pvﬁw v.‘ .

V
,4
“'I I
4. ‘ M‘i‘
\“<
“‘V'Q~§ld
' 17,-?
“.343
~ .

4‘34

‘ ‘r
.
=11. "
_ z- ‘w
1-1A~_‘o'j‘~}
w; 1&7“.
“8.... ﬁnk}-
43.2%;
.v 1. . 4
VV
r €?{,‘J,'f’.:43;._
‘ 4 ..

avig if???“ "I u,

a {23$}; 5“: ’ .
‘3? Kuwyx; 11‘ "3"
62:}: .0.” .7

a 1 4‘7

, gﬁhf‘i ‘}4

O a I 4;!

1'7 ‘3‘
- .2.
‘I I

.
t4
. 3:
.3.

.
-‘.
~

" m-IL‘

4 1"
“u M (a;
ﬁd‘§$
, ‘ ‘
«a

v
y: . «
b-g’u‘u$"<
04A ‘l‘I ‘7’ ’5

"$3“; If» ‘ ‘ A'
.J'Jqu“.-- 5" ' ' ‘
V ,4 " ' "“14-
‘ «‘4‘ 1"”
1’?!“ a" .3

.1 ""-":’4" W: "

'52,
‘1':

'21th ’4"
,4 "a
«A; .4’

'4"
""1"."
q '1
1’3

J‘Jyl

.y
M,"
new;
44', {f‘II'Ct wt;

f‘K/“e'
ﬂ.

j,“,’:(2‘ ‘3 ‘ .4

- ,r 4 " :‘:.""

ﬁlﬂ-(fo'v'l'urf 9“)?“ K 3 . ' 5% {I’VE}
‘Vpﬂﬁ

y-‘H‘lvdra'g ('6'?
'ﬁ "
_‘ .w
”it"{ngl'f "$112: -’?r;;"""'&

(12...,

”52:7!

5“"
W" ‘ ‘
.4' '1
'31""'

~14 £5

4
#1.” ‘5‘.)
IV :(‘41 [:4
w\’f.‘,{ a;
13""! ‘rf,

............................... 0 51670
IIIIIIIIIII IIIIIIIIIIIIIIIIIIII IIIIIIIIIIIIIIIIII Lab-4AM

_ 313005741172 chigan SI I
University

 

 

This is to certify that the

dissertation entitled

0n Interleaving Syntax and Semantics in Parsing

presented by
Dongyul Ra

has been accepted towards fulﬁllment
of the requirements for
PhD Computer Science
(kgnxin

 

 

I Major professor

17 Oct 1989
Ihne

 

MS U is an Aﬂinnativc Action/Equal Opportunity Institution 0- 12771

 

 

 

 

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

MSU In An Affirmative Action/Equal Opportunhy lnetilution

ON INTERLEAVING SYNTAX
AND SEMANTICS IN PARSING

By

Dongyul Ra

A DISSERTATION

Submitted to
Michigan State University
in partial fulﬁllment of the requirements

for the degree of

DOCTOR OF PHILOSOPHY

Department of Computer Science

1989

Liz-{‘3 4 ’2. 2. I I

ABSTRACT

ON INTERLEAVING SYNTAX
AND SEMANTICS IN PARSING

By

Dongyul Ra

It is agreed that natural language parsers should use both syntactic and semantic
knowledge for better analysis. Several approaches for using these distinct knowledge
have been proposed: non-inte grated parsing and integrated parsing. Under non-
integrated parsing, there are two kinds of approaches: the old extreme one which is
represented by the serial stage parsing and the recent one called interleaved semantic
processing. Under integrated parsing, there are also two views: the old extreme one

represented by conceptual analyzers and the recent moderate one shown in Lytinen’s

MOPTRANS parser.

This thesis attempts to update the interleaved semantic processing to achieve the
power of integrated parsing. An integrated parser, called SYNSEM, has been developed
for this purpose. In SYNSEM, each alternative syntactic analysis becomes a branch of
parsing. Multiple branches can run in parallel but they are ﬁltered, as soon as possible,
according to their semantic processing results at the periodic synchronization points.
SYNSEM’s approach is able to utilize semantics(as well as syntax) as much as the

recent integrated parsing approach.

To provide the necessary semantics for comparing the branches of parsing, a
knowledge base(KB) is built using the recent knowledge representation language,
KODIAK. A marker passing paradigm is used to utilize the knowledge base. The seman-

tic processing result for each branch is realized as the paths in the knowledge base. To

solve the hard problem of ﬁnding the best paths between two concepts, a mechanism
called secondary marker passing has been developed which efﬁciently cuts down the
number of paths to be considered. Several heuristics have been developed for selecting
only the best paths. The comparison of branches to ﬁlter those with inferior semantic
results is done by comparing the KB paths. A heuristic for this comparison is also

developed.

The approach described so far aims at structural disambiguation. Word sense
ambiguity is another big source of ambiguities in natural language. Several methods for
word sense disambiguation are developed in SYNSEM. By manipulating a set of the
KB paths, an easily conceived and simply implementable method for word sense disam-
biguation is introduced. The path adjustment operation after the concretion of a sense to

the more speciﬁc sense is a new source of information utilized in the SYNSEM parser.

To my wife and daughter

for their love and support

iv

ACKNOWLEDGEMENTS

I would like to express my sincere gratitude to my advisor, Dr. George C. Stock-
man, for his guidance over the years at Michigan State University. Dr. Stockman
advised me at every phase of my research. He showed a great deal of patience while I
was going through the long period of reading literature and preparing for the research.
He patiently listened to my ideas and plans and provided valuable comments and
suggestions even when my ideas were sometimes vague or seemed strange. He also pro-
vided encouragement for me when I needed it to continue this arduous thesis work.

Without that, my program might never have been completed.

I am grateful to my committee members, Dr. Anil K. Jain, Dr. Carl V. Page, Dr.
Jon Sticklen, Dr. J. Kathryn Bock, and Dr. Barbara K. Abbott for giving me kindness,

valuable suggestions and comments and for teaching me during my graduate program.

I am grateful to Barbara Czemy for reading the draft of this thesis and correcting
numerous errors. I also would like to thank my friends I’ve made during my stay in this
university for helping me in one or more ways: Chaur-Chin Chen, Sci-Wang Chen, Xian
He Sun, Jayashree Ramanathan, Arun Nanda, Rob Jaksa, Mahmoud Pegah, John Kern,

Joe Miller, Paul Wolberg and many others.

Finally I would like to thank my wife and my daughter for their constant support,

patience, and love.

TABLE OF CONTENTS

List of Tables ............................................................................................................ ix
List of Figures .......................................................................................................... x
Chapter 1. Introduction .......................................................................................... 1
1.1. Serial Stage Parsing ........................................................................................ 3
1.2. Integrated Parsing ........................................................................................... 4
1.3. Interleaved Semantic Processing .................................................................... 7
1.4. Improving Interleaved Semantic Processing: a New Proposal ....................... 10
1.5. Outline of the Thesis ....................................................................................... 13
Chapter 2. Parsing Models in Psycholinguistics ................................................... 15
2.1. Serial Models .................................................................................................. 15
2.1.1. Structural Preference Model .................................................................... 16
2.1.2. Lexical Preference Model ........................................................................ 20

2.2. Parallel Models ............................................................................................... 24
2.2.1. Discourse Model ...................................................................................... 25
2.2.2. Ranked Parallel Model ............................................................................ 28
2.2.3. A Parallel Model with Conceptual Selection Hypothesis ....................... 31

2.3. Summary ......................................................................................................... 36
Chapter 3. Framework of New Approach ............................................................. 37
3.1. General Framework of the Control ................................................................. 38
3.2. Conﬁguration of SYNSEM ............................................................................. 42

vi

3.3. Operating Principle of SYNSEM ................................................................... 44

3.4. Conclusion ...................................................................................................... 48
Chapter 4. Syntactic Analysis ................................................................................. 50
4.1. Syntactic Analysis .......................................................................................... 50
4.2. Structure of Rules ........................................................................................... 53
4.3 Delay Used for Increasing Syntactic Context ................................................ 59
4.4. Rules and State of Parsing .............................................................................. 64
4.5. Rules and Ambiguity ...................................................................................... 67
4.5.1. Delay and Parallelism .............................................................................. 70

4.6. Syntactic Processing in Operation .................................................................. 72
4.7. Conclusion ...................................................................................................... 80
Chapter 5. Knowledge Base and Marker Passing ................................................ 82
5.1. KODIAK ......................................................................................................... 83
5.1.1. Modeling Knowledge .................................................................................. 87
5.2. Marker Passing in SYNSEM .......................................................................... 89
5.2.1. Links and Flow of Markers ...................................................................... 92
5.2.2. Basic Considerations on Marker Passing ................................................ 94
5.2.3. Follow-on Collisions ............................................................................... 97
5.2.4. Overlapping Collisions ............................................................................ 99
5.2.5. Marker Passing Algorithm ....................................................................... 100

5.3. Conclusion ...................................................................................................... 102
Chapter 6. Providing Semantics to Parsing .......................................................... 103
6.1. Interfacing Syntax and Semantics .................................................................. 104
6.2. Finding a Path for a Suggestion ...................................................................... 108
6.2.1. Secondary Marker Passing: Motivation and Basic Concept ................... 110
6.2.2. Consideration on the Polarity of Paths .................................................... 115
6.2.3. Heuristics for Finding the Best Path ........................................................ 119

vii

6.2.3.1. Hierarchy: H—heuristic ..................................................................... 120

6.2.3.2. Length: L-heuristic .......................................................................... 121

6.2.3.3. Speciﬁcity: S-heuristic ..................................................................... 122

6.2.4. Operating Example of Finding the Best Path .......................................... 123

6.3. Semantic Comparison of Branches ................................................................. 129
6.4. Semantic Processing Component in Action ................................................... 132
6.4.1. Example 1 ................................................................................................ 133
6.4.2. Example 2 ................................................................................................ 136

6.5. Conclusion ...................................................................................................... 143
Chapter 7. Word Sense Disambiguation ............................................................... 145
7.1. Use of Semantic Paths for Word Sense Elimination ...................................... 147
7.2. Use of Concretion ........................................................................................... 155
7.3. Use of Semantic Association .......................................................................... 158
7.4. Use of Information about Syntactic Category ................................................ 161
7.5. Conclusion ...................................................................................................... 162
Chapter 8. Conclusion ............................................................................................. 164
8.1. Proposed Approach of Parsing ....................................................................... 164
8.2. Argument for the Proposed Approach ............................................................ 165
8.3. Observations on Implementations and Test ................................................... 168
8.4. What Has Been Achieved ............................................................................... 170
8.5. Future Research Items .................................................................................... 172
Appendix A Structure of Noun Phrase(NP) ............................................................. 174
Appendix B Determining Dead-end Situation .......................................................... 17 5
Appendix C Test Sentences ...................................................................................... 177
Appendix D System Output of Parser ....................................................................... 182
Appendix E List of Sample Rules ............................................................................ 201
List of References ..................................................................................................... 206

viii

LIST OF TABLES

Table 4.1 Some Actions ............................................................................................ 57
Table 5.1 Links in KODIAK .................................................................................... 85
Table 6.1 Data of Finding Paths ............................................................................... 134

ix

LIST OF FIGURES

Figure 1.] Stages of Parsing ..................................................................................... 3
Figure 1.2 Integrated Stage of Parsing ..................................................................... 4
Figure 1.3 Generalized Syntactic Rules in MOPTRAN S ......................................... 5
Figure 1.4 Arcs of an ATN State .............................................................................. 8
Figure 2.1 Alternative Analyses Tried Serially ........................................................ 16
Figure 2.2 Minimal Attachment ............................................................................... 18
Figure 2.3 Failure of Local Attachment ................................................................... 19
Figure 2.4 Ambiguous Region .................................................................................. 20
Figure 2.5 Lexical Forms of Verbs ........................................................................... 22
Figure 2.6 An example of Final Argument Principle ............................................... 23
Figure 3.1 Interleaved Semantic Processing - 1 ....................................................... 38
Figure 3.2 Interleaved Semantic Processing - 2 ....................................................... 39
Figure 3.3 Interleaved Semantic Processing - 3 ....................................................... 40
Figure 3.4 Interleaved Semantic Processing - 4 ....................................................... 41
Figure 3.5 Conﬁguration of SYNSEM ..................................................................... 42
Figure 3.6 "the" Processing ...................................................................................... 44
Figure 3.7 Lives of Branches .................................................................................... 46
Figure 3.8 Partial Parallelism ................................................................................... 47
Figure 3.9 Flow of the Control ................................................................................. 49
Figure 4.1 Syntactic Working Memory .................................................................... 51
Figure 4.2 Syntactic Node ........................................................................................ 51
Figure 4.3 A Tree in SWM ....................................................................................... 52
Figure 4.4 Format of Rules ....................................................................................... 53
Figure 4.5 Speciﬁcation of the Base Pattern ............................................................ 55
Figure 4.6 Main-verb2 Rule ..................................................................................... 56
Figure 4.7 Mapping Patterns to Trees ...................................................................... 58
Figure 4.8 Parsifal’s Pattern Matching ..................................................................... 58
Figure 4.9 Delaying in the build-PP Rule ................................................................. 59
Figure 4.10 Steps of Building a PP ........................................................................... 60

Figure 4.11 Recursive Embedding of Attention Shifting ......................................... 61

Figure 4.12 Matching of Lowest and Rightrnost NP ................................................ 62
Figure 4.13 Implicit Attachment of Secondary Clause ............................................ 63
Figure 4.14 PP Attachment Ambiguity .................................................................... 67
Figure 4.15 PP Attachment and the Corresponding Rules ....................................... 68
Figure 4.16 Verb Form Ambiguity ........................................................................... 69
Figure 4.17 CLIST and Categorical Ambiguity ....................................................... 69
Figure 4.18(a,b) Snapshots of SWM during Parsing ................................................ 73
Figure 4.18(c,d,e) Continued .................................................................................... 74
Figure 4.18(f,g,h) Continued .................................................................................... 74
Figure 4.18(i.j,k) Continued ..................................................................................... 75
Figure 4.18(l,m,n,o) Continued ................................................................................ 76
Figure 4.18(p.q,1’.s,t) Continued ............................................................................... 77
Figure 4.18(u,v,w,x) Continued ................................................................................ 78
Figure 4.18(y,z,zl,22) Continued ............................................................................. 79
Figure 4.18(z3,z4,25) Continued .............................................................................. 80
Figure 4.19 Creation of Multiple of Branches .......................................................... 75
Figure 5.1 Elements of KODIAK ............................................................................. 84
Figure 5.2 Frame Representation .............................................................................. 84
Figure 5.3 Instance Links ......................................................................................... 86
Figure 5.4 An Example of KB .................................................................................. 88
Figure 5.5 Memory of TLC ...................................................................................... 90
Figure 5.6 Origin of Marker Passing ........................................................................ 93
Figure 5.7 Links and Nodes Passing Markers .......................................................... 93
Figure 5.8 Split of Markers ....................................................................................... 95
Figure 5.9 Loop in Marker Passing .......................................................................... 96
Figure 5.10 Merging of Markers .............................................................................. 96
Figure 5.11 Follow-on Collision .............................................................................. 98
Figure 5.12 Regular and Follow—on Collision .......................................................... 98
Figure 5.13 Overlapping Collisions .......................................................................... 99
Figure 5.14 Marker Passing Algorithm .................................................................... 101
Figure 6.1 Flow of Information from Syntax to Semantics ...................................... 104
Figure 6.2 PP Attachment Depending on Semantics ................................................ 106
Figure 6.3 Inﬂuence of Semantics on Parsing .......................................................... 107
Figure 6.4 Conversion of Suggestions ...................................................................... 108

xi

Figure 6.5 Part of KB ................................................................................................ 109

Figure 6.6 Hierarchies of Relations .......................................................................... 111
Figure 6.7 A Three-way Collision ............................................................................ 113
Figure 6.8 Cut-down of Marker Passing Space ........................................................ 114
Figure 6.9 Two Collisions with One to Be Removed ............................................... 115
Figure 6.10 Role and Play of Aspectuals of Relations ............................................. 117
Figure 6.11 An Example of Role and Play ............................................................... 117
Figure 6.12 Stages of Finding Best Paths ................................................................. 119
Figure 6.13 Removal of a Collision Using Ancestor Information ........................... 120
Figure 6.14 Number of Nodes in the D—hierarchy .................................................... 122
Figure 6.15 Possible Forms of Paths ........................................................................ 122
Figure 6.16 End Absolutes of Two Paths to Be Compared ...................................... 123
Figure 6.17 Rank-Ordering the Branches ................................................................. 129
Figure 6.18 Comparison of Two Rules(vp-pp & np-pp) .......................................... 130
Figure 6.19 Instances and Two Paths to Be Compared ............................................ 130
Figure 6.20 Two Paths in "cat" Example ................................................................. 132
Figure 6.21 Competition of Three Branches ............................................................ 135
Figure 6.22 Removal of a Branch with Bad Semantic Processing ........................... 135
Figure 6.23 A Snapshot of CLIST ............................................................................ 136
Figure 6.24 Flow of Branch for Noun-parse3 .......................................................... 137
Figure 6.25 Flow of Branch for Noun-parse4 .......................................................... 138
Figure 6.26 Flow of Branch for Noun-parseS .......................................................... 139
Figure 6.27 Comparison of Three Branches ............................................................. 140
Figure 6.28 A Path without a Relation Node ............................................................ 140
Figure 6.29 Parallel Running of two Branches ......................................................... 142
Figure 6.30 Final Branch for the Parse ..................................................................... 143
Figure 7.1 Knowledge Base for Examples ............................................................... 147
Figure 7.2 Representation of Multiple Word Senses ................................................ 148
Figure 7.3 Different Word Senses for Different Branches ....................................... 149
Figure 7.4 Paths between Two Instance Nodes ........................................................ 150
Figure 7.5 Effect of Reading a New Word ............................................................... 151
Figure 7.6 Paths after Removing a Sense and a Path ............................................... 151
Figure 7.7 RWSR Operation .................................................................................... 152
Figure 7.8 Consideration for the Set of Non-removable Senses .............................. 153
Figure 7.9 Starting the removal Operation ............................................................... 154

xii

Figure 7.10 RWSR Operation Routine ..................................................................... 154

Figure 7.11 Shape of Concretion Path ...................................................................... 155
Figure 7.12 Concretion for the Example .................................................................. 156
Figure 7.13 Success of Path Adjustment .................................................................. 157
Figure 7.14 Failure of Path Adjustment ................................................................... 158
Figure 7.15 Use of Semantic Association ................................................................ 160
Figure 7.17 Use of Categorical Information ............................................................. 161
Figure A-l Structure of an NP ................................................................................. 174

xiii

CHAPTER 1

INTRODUCTION

For humans natural language is an easy, ﬂexible, and powerful means of communi-
cation. The goal of natural language processing, an area of Artiﬁcial Intelligence, is to
allow humans to use natural language to communicate with computers. One of the most
important advantages of achieving this goal is that humans can use their language, i.e.,
natural language, to communicate with computers. They can issue commands to the
computer with, for example, English and receive responses in English. Computers could

read or write documents written in natural language.

The global goal of natural language parsing is to make computers understand input
expressed in natural language. But this deﬁnition needs more elaboration because
"understand" is not speciﬁc enough. The fact that "a man understood a sentence" means
that he got the meaning of the sentence. This means: ﬁrst he translated the input sen—
tence into a chunk of some biological medium that is used by the brain processes for
doing things such as thinking, storing, reasoning, speaking, etc.; second, he identiﬁed the
status of the chunk by relating it to the existing knowledge (for example, it is true, false,
new, etc.). An interesting thing is that these two steps of understanding are not separate
steps. It seems that they are interwined and mixed. A similar analogy can be applied to
the case of computers. The fact that a computer understands an input sentence means
that the input sentence should be translated into some form of a representation language
that can represent the meaning of the sentence. This representation language is usually
called the internal representation language or knowledge representation language. This
internal representation language is the universal medium that is used by the intelligent

processes in the system. This might be used for further processing such as reasoning or
1

answering.

Building a natural language parser is a very complex and difﬁcult task. The
detailed processing principles and steps have been studied much but there is no uniﬁed
theory yet. One of the difﬁculties of building a natural language parser is because
natural language is full of ambiguities. To resolve the ambiguities, various kinds of
knowledge should be extensively used. The use of syntactic knowledge usually enables
the parser to understand the grammatical structure of a sentence. Syntactic analysis is
one of the areas that is most clearly understood according to the research of linguists.
But as we go from syntactic knowledge to the higher level knowledge, we do not have
clear understanding of the exact way of representing the knowledge or using the
knowledge during parsing. A big problem arises from the fact that many syntactic ambi-
guities require the use of higher level knowledge (let’s call this semantic information

broadly) for resolution.

In addition to this problem, there is not much agreement on the role that the syntac-
tic and the semantic knowledge should play in parsing and there is not much agreement
on what kinds of structures should be built during the course of parsing. Some argue
that syntactic knowledge need not be extensively used and syntactic structure of the
input sentence need not be completely built and used because the ﬁnal goal of parsing is
to compute its meaning (semantic interpretation) instead of its syntactic
structure(Schank and Bimbaum, 1980). Some argued that the syntactic structure of the
sentence should be fully computed ﬁrst and then semantic information should be applied
to resolve the ambiguities and compute the semantic interpretation(Woods, 1972). Oth-
ers argued that the parsing should involve the use of syntactic and semantic information
at the same time(Charniak, 1981c). The means of how to use semantic information is not

completely known.

The research to be reported in this thesis investigates the use of semantics in pars—

ing. What should the parsers’ control strategy be for the efﬁcient use of syntactic and

semantic information? What kind of method should be used to glean the semantic infor-
mation that is necessary for the disambiguation? How are they applied to inﬂuence the
computation of the syntactic structure? In other words, this thesis is dedicated to ﬁnding
a parsing model that has a better answer for these questions. A parsing model that is new
and advanced will be adopted and an actual parser based on this model will be
developed. The problems and techniques that arise during the development of the parser

will be examined in this thesis.

In this chapter, several parsing approaches that have been used in previous parsers
will be introduced and then a new parsing approach will be introduced and motivated.

This chapter will conclude with the outline of the thesis.

1.1 Serial Stage Parsing

In the serial stage parsing approach, the syntactic analysis and semantic analysis
are done separately in two stages which are ordered serially. As shown in Figure 1.1, the

syntactic analysis stage comes

 

 

Input :> Syn‘awc j > / > Semantic It a1
Sentence Analysis Ananysis 11 em

Representation

 

 

 

 

 

 

Syntactic
Tree

Figure 1.1 Stages of Parsing

ﬁrst and then the semantic analysis stage follows. In the syntactic analysis stage, gram-
matical knowledge is used to build the syntactic tree for the input sentence. After the full
syntactic tree for the whole input sentence has been constructed, the semantic analysis is
done in the next stage. The syntactic tree from the ﬁrst stage is input to the semantic
analysis stage. Non-syntactic knowledge is the major information used in the semantic
analysis stage. One of the advocates of this parsing approach was Woods(l972). He

argued that the parsing will be more efﬁcient and faster if syntactic analysis is done ﬁrst

without worrying about the semantic content of the sentence and then semantic process-
ing (which requires lots of computation) is based on the result of the syntactic analysis
which is a more appropriate form to process. The following quotation from

Woods(l972:145) shows his argument:

We have not gotten any conclusive answers yet regarding the effectiveness of this
type of semantic screening, but it looks as if it takes longer to do the parsing and
semantic interpretation overall if the interpretation is done during the parsing than
it does if the parsing is done ﬁrst and the interpretation afterwords. This seems to
be because the semantic interpretation process is at least as expensive as following
out the syntactic consequences without semantic guidance, and the syntax is as
likely to rule out an alternative and keep the interpreter from having to interpret it
as the semantics is likely to rule out an alternative and keep the parser from having
to operate further on it. Clearly there is no a priori reason to expect that semantic
screening will save more of the parser’s effort than syntactic screening saves of the
interpreter’s effort.

But this approach has been opposed by many researchers. For example,
Chamiak(1981c) refuted Woods by saying that the serial stage approach has no psycho-
logical reality and it is rather inefﬁcient in the face of lots of ambiguities in the human
language. The approach of serial stage parsing became an out—of-date parsing approach
which is the typical example of non-integrated parsing. The non-integrated parsing
approach is directly opposite to the integrated parsing approach that will be explained in

the next section.

1.2 Integrated Parsing

In the integrated parsing approach, there is no separation between the syntactic and

semantic processing stages as shown in Figure 1.2.

 

Input : Integrated Internal
Sentence Stage Representation

 

 

 

Figure 1.21ntegrated Stage of Parsing

In addition to the fact that two stages are integrated into one stage, this approach has the

following features: semantics is the major driving force of parsing; syntactic trees are
not built, which results in using only local syntactic cues and not global syntactic infor-
mation. Semantic representation of the input sentence is directly built. In this model,
syntax is viewed as less important information for parsing. This view has been strongly
advocated in Schank and Birnbaum(1980). A series of parsers have been developed fol-
lowing this approach: the Margie parser in Riesbeck(1975), ELI in Riesbeck and
Schank(1976), IPP in Lebowitz(1980), and Conceptual Analyzer in Birnbaum and Sel-
fridge(1981). Another parser that belongs to this model is the semantic grammar shown

in Burton(l976).

But one recent parser attempts to provide more syntactic analysis to integrated
parsing while it still tries to maintain the characteristics of integrated parsing. This
parser is the MOPTRANS parser developed by Lytinen(1984). Instead of using the lexi-
cal request-based rules used in the conceptual analyzers, MOPTRAN S uses the general-

ized syntactic rules. Figure 1.3 shows two generalized syntactic rules, "subject" rule and

"object" rule.
Subject Rule
Syntactic pattern: NP, V (active)
Additional restriction: NP is not attached syntactically
Syntactic assignment: NP is SUBJECT of V, V is a MAIN CLAUSE
Semantic action: NP is actor of V(or another slot,
if specified by V)
Result: V (changed to 8)

Object Rule

Syntactic pattern: S, NP

Additional restriction: NP is not attached syntactically

Syntactic assignment: NP is (syntactic) OBJECT of S

Semantic action: NP is (semantic) OBJECT of S(or
another slot, if specified by S)

Result: 8, NP

Figure 1.3 Generalized Syntactic Rules in MOPTRANS

The generalized syntactic rules are indexed in a way that semantic knowledge is fully
used before the use of syntactic knowledge. At each step of parsing, all possible seman-

tic connections between the semantic objects in the memory are found. Then the

generalized syntactic rule which satisﬁes the constraint of its syntactic pattern and builds
the best semantic connection is selected and executed. Thus semantic knowledge is
used ﬁrst and then syntactic knowledge is consulted at each step of rule indexing and
execution. This strategy guarantees the utilization of the predictive power of semantics.
The important point in the rule indexing strategy of MOPTRANS is that the parser will
never consider a rule that results in incorrect semantic effect. This can be shown using

the following two sentences:
(1-1)The man wrapped the present.
(1-2) The present wrapped by the man was expensive.

After "wrapped" is input in the ﬁrst sentence, the memory contains two elements,
HUMAN for "the man" and COVER for "wrapped". The possible semantic connections
between these two memory elements are sought by the parser and it is found out that
there is only one possible connection: HUMAN is connected to COVER via the agent
relationship. Then any rule that matches the syntactic pattern and builds this semantic
connection is sought. In this case, only one rule, the Subject rule, is found. Then this
rule is executed and the parsing continues. Now, let’s consider the situation after read-
ing "wrapped” in sentence (1-2). The memory contains two elements: GIFI‘ and
COVER. Only one semantic connection is found between these two elements: GIFT is
the (semantic) OBJECT of COVER. All syntactic rules which match the syntactic pat-
tern and can build this semantic connection are found. Note that the Subject rule
matches the syntactic pattern of the current situation but it can not build the required
semantic connection. Thus the Subject rule is not found. The Object rule is not found
because its syntactic pattern does not match the input pattern. But the Unmarked-Passive

rule is found by the rule matching process.

Unmarked-Passive Rule

Syntactic pattern: NP, VPP
Additional restriction: none
Syntactic assignment: NP is (syntactic) SUBJECT of VPP, VPP
is passive, VPP is a RELATIVE CLAUSE of NP
Semantic action: NP is (semantic) OBJECT of S(or another slot
if specified by VPP)
Result: NP, VPP(changed to S)

This rule’s syntactic pattern matches the pattern of the input. According to the semantic
action of this rule, the connection of semantic OBJECT can be built by this rule. There-
fore this rule is found and executed. The approach of MOPT'RAN S greatly improves on
the weakness of the conceptual analyzers in the analysis of sentences that have complex
syntactic structure. MOPTRANS used one stage of parsing but both syntactic and

semantic knowledge is used in a more balanced way titan the conceptual analyzers.

1.3 Interleaved Semantic Processing

It has been agreed that the serial stage parsing model is not the appropriate model
for natural language parsing. An updated version of the serial approach has been popu-
lar. This is called interleaved semantic processing(Ritchie, 1983). In this approach the
syntactic analysis is still the major driving force of parsing but semantic knowledge is
consulted during syntactic analysis. Semantic processing is interleaved with the syntac-
tic decision making points so that the decision can be made after consulting semantic
information. Semantic interpretation is built in parallel with the construction of the syn-
tactic representation. The possibility of doing the wrong syntactic analysis (which does
not make sense semantically) for the whole sentence in the serial stage parsing model
can be removed in the interleaved semantic processing model while the parser can still
take advantage of the clear-cut information that can be provided by the syntactic struc-

ture of the sentence.

One of the typical parsers incorporating interleaved semantics is Psi-Klone(Bobrow
and Webber, 1980). Psi-Klone’s syntactic analyzer is based on the ATN parser. When it

executes actions of an arc of a state of ATN, messages are passed to the semantic

component. The semantic component builds the semantic structure corresponding to the
messages. If the attempt at building the semantic structure succeeds, the semantic com-
ponent returns the signal of success to the ATN analyzer (otherwise, a signal of failure is

returned). Let’s consider Figure 1.4 which shows states of the ATN.

 

 

Figure 1.4 Arcs of an ATN State

At state A, arc l is tried ﬁrst. If the semantic component returns the signal of success for
the messages of this arc, the parser moves to state B without considering the other
arcs(arc 2 and are 3). But if the failure signal is returned, the parser tries the next arc,
are 2. This arrangement of control takes advantage of separating syntax and semantics
but allows a large amount of semantic checking. This is a modular approach and the

syntactic rules are independent of the domain.

The messages passed to the semantic component are of the form (M, L, C), where
M(matrix) is some half-built constituent, C is a constituent of some kind and L is a label
that indicates some syntactic relation. What (M, L, C) suggests is that C is attached to M
by a link labeled with L. The semantic component checks the semantic constraint related
to this attachment to decide if the attachment can be done. When the attachment is
done, a symbolic name that points to the result of this structure building operation is
returned to the parser. Upon the receipt of a message (M, L, C), the semantic com-
ponent uses relation-mapping rules(RM rules) to determine into which semantic relation
L can be mapped. The semantic relation here is something like cases or roles in other

systems. The attachment is done based on this semantic relation. For each semantic

relation, there are interpretation rules(l rules) which are applied to convert the case-

frame-like structure into a full semantic item for the updated matrix.

Note in Figure 1.4 that arcs are tried one after another, serially, until one succeeds.
The ﬁrst arc that gets the signal of success is the one that is selected by the parser to
update the parsing. But the problem is that an are ordered later may have better seman-
tics. This arc with best semantics is missed without being detected because the are
appearing before it has just enough reasonable semantic effect to pass the semantic
check. For example, the semantic effect related to arc 3 may be better than that of arc 1
but the are 1 is reasonable enough to pass the semantic check. In this case, the parser
never notice the fact that arc 3 is better than are l and it just chooses are 1 to update the

parsing and move to the next state.

Another parser that is an example of the interleaved semantic processing approach
is Paragram and Absity system(Charniak,1983b; Hirst, 1984, 1986). Paragram is based
on the Marcus-style parser. At each stage of parsing, the state of stack and buffer is
matched against the rules in the rule base (actually the rules in a packet of the rule base).
A rule that is found is executed. During the execution of the rule, the corresponding
semantic processing is done in parallel. The parser attaches one element to another ele-
ment only when the semantic effect related to the attachement is proved to be good. Oth-
erwise, another attachement in the rule will be tried. Because the parser is a determinis-
tic parser, only one rule should be found by the pattern matcher. It is not clear what will
happen when all attachment attempts in the rule fail because of semantic reasons. This
problem is beyond the scope of Paragram. Thus the selection of one out of several pos-
sible attachments in the same rule is based on the semantic checking and this is the way

of achieving "semantic interleaving" used in the Paragram and Absity parsing system.

Another early parsing system that used interleaved semantic processing is the
SHRDLU parser developed by Winograd(1972). When this parser builds a semantic

structure, a list of semantic markers is associated with the structure. The parser checks

10

the internal consistency of the semantic marker list when it builds a new structure and
the checking is used as the feedback to the parser. This semantic feedback enables the
parser to avoid the unnecessary construction of structures. The parser is based on pro-
cedures instead of declarative rules. But it is clear that the syntactic analysis routine con-
sults the semantic component before it decides to take any action of syntactic analysis.
At an ambiguous point, the ﬁrst analysis with the reasonable semantic feedback is
selected. The advocates of integrated parsing argued that the interleaved semantic pro-
cessing approach also belongs to non-integrated parsing where the serial stage parsers
are the most extreme examples. Lytinen(1984:4) described non-integrated parsing as

follows:

Syntax plays an important role in the process of understanding the meaning of a
natural language text. Syntactic and semantic processing are largely separate, with
syntactic processing performed ﬁrst (although semantic processing can be inter-
leaved with syntactic processing; i.e. once syntax has produced a partial analysis,
semantic interpretation of that portion of the text can proceed before other portions
of the text are syntactically analyzed). Syntax and semantics interact with each
other in limited ways, if at all. Syntax might be allowed to ask certain types of
questions of semantics at particular times, but communication between syntactic
and semantic processing is not unlimited. Knowledge about syntax and semantics
is also largely separate. Syntactic knowledge can be expressed without much refer-
ence to semantics.

1.4 Improving Interleaved Semantic Processing: a New Proposal

It cannot be denied that the popular approach to parsing is to make syntactic
analysis rather than the semantic analysis the major driving force of parsing. The reason
is that syntax can provide the ﬁrst clear information about the structure of a sentence.
This structural information is useful to understand the meaning that the sentence con-
veys. If the meaning can be understood easily without the use of this syntactic informa-
tion as the integrated parsing advocates argued, why does the syntax exist? The follow-
ing quotation from Chamiak(1981c) argues for the whole idea behind interleaved

semantic processing:

sq: 30 med uopse sq: u: rsq:o sq: rsAo uonse :usurqsene suo Sunsslss .10} &[uo pssn
sq ues Suppsqs snuemss voyeurism: spueurss ﬂue :noqnm uoneuuowr susmuﬁs Kluo
sszqnn uonsslss sInr stql 'suss rsynq ssrq: o: dn pesqe Suplool Kq puno} Apssrros
sq ues slur suo Kluo :eq: psurnsse st 1! ‘sldunrxs rod ussred sq: jo uonse sq: o: rsmod
snueurss Kldde o: preq st 1! srsqm uonmnﬁguos :2 sum srssmd ssoql 'suom-tsd q:rm

ueq: ssrom s:s3 urslqord sq: ‘umrﬁmd .to (0861 ‘snsrewnejtsred jo sses sq: 1:1

'srssnzd palerﬁsmt
sq: u: pspromr sq ues stql 'srsﬁleue rs::sq sq: Supper morj rssmd sq: :usAsrd lieu:
rszﬁleue snsmuﬁs :eq: suesur STILL (xmuﬁs Kq sspueurss go [onuos srsAss sq: :sn[ ‘srsq
:utod :umloduq ue sq 0: areas :,ussop rspro sq: :eq: awn) 'Sutssssord suuetuss uo psseq
stsﬁleue rs::sq sq: sq Ken: 9 are Kq ps:usssrds.t stsﬁleue sq: :eq: sses sq: sq Ken: 1! :ng
'psrsptsuos rsAsu era g are pure z are 50 ssuqtqrssod sq: ‘pooB Alqeuosesr 8; :eq: ﬂutssss
-o.rd spueurss sq: u: sqnssr 1 are 51 '1 suns u: Sutsred jo sAem slqrssod sm sore om: .Isq:o
sq: 5: mm 1 are 10; Sutssssord spueurss sq: s:ss88ns 1819 :e rssred my sq: ‘7'1 smﬂtd
u: ‘sIdtuexs 10:1 'rszriprue snsmuﬁs sq: Kq ps:ss88ns st :eq: stsﬁleue sq: ro; Kluo suop
s: Sutssssord suuetuss :eq: st mtels srq: ro; uosesr sq; 'qseordde Butsred psterﬁsmt
-uou e se psgtssels sq plnoqs an: pue ssnueurss jo rsmod {In} sq: ssn :ou ssop

Surssssord snueurss psAeslrsrur :eq: psurtep Butsred psmrﬁswt jo ssmsonpe sql

°1spour ssnueurss psurqtuos s ro ‘Ispour ”muﬁs

on“ e rsqns om; smrsusﬁsp Alsremnln [um [spom preoquselq e ‘sprom rsq:o u1
(3111219 xmuﬁs qua: .Isqroq qu usq: ‘stq: op pmos :1 J1 'suestu .tsqro Kq uopeunoy:
sures sq: qsqqms-sr 0: mm pmom 1! :eq: uesur pIROM stql ':usuoduros snsmuﬁs
sq: Kq psqsqqmss sm:snns [estﬁol sq: :noqnm op o: pssroj sq plnom ssnueurss sq:
[spom preoqxselq 12s.: e :11 °ssus:uss sq: jo smrsnns [estﬂol sq: :12 :ns my a :98 0:
Arm snot/\qo :sour sq: st xmuﬁs ‘ps:ou Kp'esrle sAeq 9M sv 'sses sq: sq pInoqs stq:
qu sss o: preq :r s; .toN 'rsqnzs pssodord ([spou: Surssssord snueurss psAeslrsmt
sq: ro) [spour supnorqns spueurss sq: o: stqdrourosr [spout sq: Suntan ‘:usuod
-tnos snsmuﬁs sq: jo mdrno sq: uo xrom Kluo ptnom sspuemss sq: uBrssp [emote
u: ‘Isuered u: psxrom SUIGISKS om: sq: ﬁnestuqss: anm '1; ssn lines: :ou prp 5198:;
AVSHVEIH :eq: :02; sq: ,(q pmssﬁﬁns st [spout stq: qum squnon sre srsq: :eqL
'qutsusqsrduros st ssusruss sq: J; was] :2 ‘pssssns nus mm :usuoduros

snuetuss sq: ‘SIIBJ :usuodnros spsmuﬁs sq: j: mm mm s: esp: sqL 'preoqxstqq
uoururos e no suorsnpuos nsq: ButAesI ‘pnered ur SIJOM srusuoduros snueurss

pun xmuﬁs 9m ‘(199001 pmoqxovtq savsuvamppow sun 30 awaqos 9m III

II

12

same rule. Thus the use of semantics for guiding the parser is limited compared to Psi-
Klone. There is also an unsettled argument on the number of buffer cells to look ahead.
For example, Milne used the lookahead of up to only two buffer cells, arguing that the
two buffer lookahead is enough. Milne( 1982) showed that a parsing decision should
sometimes be made based on the semantic bias instead of using the lookahead of all

three buffer cells.

The major goal of the research reported in this thesis is to design a parser that can
fully utilize semantics but is still based on interleaved semantic processing. The previous
interleaved semantic processing approach will be updated to achieve better use of
semantics. Thus the major driving(controlling) force of the parser will be syntactic
analysis but the method of utilizing the semantics will be modiﬁed. The parsing system
that has been developed for the purpose is called SYNSEM. SYNSEM’s syntactic
analyzer is based on Parsifal. It has been signiﬁcantly modiﬁed in several ways to
achieve the aim of the thesis. The semantic processing is based on world knowledge that
is written in the knowledge representation language called KODIAK developed by
Wilensky(l987). This means that the internal representation language of SYNSEM is
KODIAK.

The control of SYNSEM is designed in a way that the full use of semantics can be
possible even if the parser is based on syntactic analysis (note that the base of inter-
leaved semantic processing is syntactic analysis). This approach resulted in the partial-
parallelism in the parser. The partial-parallelism means that the parser mostly runs in
serial mode but it can run in parallel for the period when there exists ambiguity. The
parser tries to reduce down the number of parallel branches to one as soon as possible

using any information available (either semantics or syntax).

To use the knowledge base efﬁciently and conveniently, considering that the
KODIAK knowledge base is a kind of semantic network, the marker passing paradigm

has been adopted as the major means that the semantic component uses to obtain

13

semantic information. This strategy allows easy use of the knowledge base. Thus the
framework of the parser is to use a rule-based syntactic analyzer as the syntactic analysis
component and use the marker passing paradigm as the major means of getting semantic
information. Some schemes are developed that make this framework achieve the global

aim of designing a parser that utilizes semantics as fully as possible.

One of the most important problems related to marker passing is how to ﬁnd the
best paths between two concepts. This problem has been extensively studied in this
research. The parser is designed in a way that the goodness of a syntactic analysis can be
represented by the goodness of the paths related to the analysis. This creates a problem
of comparing the paths that represent the possible analyses that compete. The path com-

parison problem is also addressed in this research.

The word sense disambiguation problem is one of the problems that any signiﬁcant
natural language understanding system should address because it is another major source
of ambiguity. The marker passing-based parsers have some advantages for handling this
problem(Charniak,1986). Computational techniques on how to handle this problem

under the framework of SYNSEM have also been studied.

1.5 Outline of the Thesis

In Chapter 2, the parsing models in psycholinguistics are reviewed for the purpose
of providing some support to the framework of SYNSEM. The partial—parallel approach
used by SYNSEM is supported by some recent parsing models. The overall framework
of SYNSEM will be designed in Chapter 3. The discussion in that chapter is mostly
related to the issue of the control strategy of parsing. The partial-parallelism is adopted
as a means of improving interleaved semantic processing to use semantics as fully as
possible. Chapter 4 is related to the description of the syntactic component of SYN-
SEM. The rule-based Marcus-style syntactic analyzer is introduced. Several new

features of SYN SEM are explained such as the format of rules, rule-indexing, and the

14

relation between ambiguity and rules. How the analyzer works is explained using some
detailed examples. Chapter 5 is a prelude to the description of the semantic processing
component. First KODIAK knowledge representation language is explained in detail but
only as much as necessary in this research. Secondly, the marker passing method used
in SYN SEM is described in detail. How the markers are passed and how the collisions
are detected will be explained. The core problems of the semantic processing com-
ponent are touched on in Chapter 6. The two problems addressed in that chapter are:
how the best path(s) are searched between two concepts in the knowledge base; how the
branches that run in parallel are compared based on the semantic processing result and
how the best ones are ﬁltered to reduce the ambiguity. In Chapter 7, word sense disam-
biguation is attacked. The major weapon for this comes from the framework of SYN-
SEM that uses the semantic network and the marker passing paradigm. The use of syn-
tactic information for this problem is also considered. Chapter 8 concludes the thesis by
discussing why SYNSEM can utilize semantics as much as integrated parsers and sum-

marizes the research results and introduces the future research problems.

The brief summary of the accomplishments in this thesis is as follows. A new con-
uol strategy for using syntactic and semantic information in parsing has been proposed.
The syntactic and semantic analysis components to achieve this control strategy have
been developed. For the syntactic analysis component, a rule-based analyzer with new
features has been developed. To provide semantics to the parser based on the marker
passing paradigm, several heuristics for ﬁnding the best path between two concept nodes
have been introduced along with the powerful mechanism called secondary marker pass-
ing. We introduce a heuristic that is used for comparing the semantic interpretation
results of the competing branches of parsing. The word sense disambiguation problem

has been studied and partly solved in the context of the parser developed.

CHAPTER 2

PARSING MODELS IN PSYCHOLINGUISTICS

There has been a great deal of research in psycholinguistics for parsing models for
the human language processing system. It is worthwhile to review these parsing models
and compare them because this can shed some light on the design of natural language
processing systems. The human parsing system shows mysterious high performance in
spite of the abundant ambiguities in natural language. It does not necessarily mean that
the human parsing system has the most efﬁcient processing mechanism. But, no natural
language parsing systems that have been developed so far can compete with humans.
Thus there might be a good chance that following the strategy of the human parsing sys-

tem may result in the design of a good automatic natural language parser.

A parsing model can be characterized by considering the following two important
questions: (1) whether all possibilities of an ambiguity are considered in serial or in
parallel; (2) what kind of information is used to select one alternative out of the multiple
analyses of the ambiguity. Some important parsing models will be reviewed with these

points in mind.

2.1 Serial Models

A parsing model that belongs to this group considers only one possible analysis at a
time when an ambiguity is encountered. After selecting one possibility according to
some selection criteria, the parsing continues as if there has been no ambiguity. If this
selection proves to be wrong by some disambiguating material appearing later in the
sentence, the parser goes back to the ambiguous point and the next possibility is tried.

Thus the alternatives of an ambiguity are tried in serial as illustrated in Figure 2.1.
15

Figure 2.1 Alternative Analyses Tried Serially

At ﬁrst, alternative (i) is tried. If the parser encounters a dead end later, the next altema-
tive, (ii), is tried. Then if it fails again, the third alternative, (iii), is tried, and so on.
When a parser encounters a dead end, the parser should retreat to a previous point of the
sentence to reanalyze it. There hasn’t been much research on this backtrack and
reanalysis strategy. It has been known that there can be three kinds of reanalysis stra~
tegy. The ﬁrst is to start the reanalysis from the ﬁrst word of the sentence. The second is
the blind chronological backtrack, which means that the parser backs up to the most
recent ambiguous point and the next alternative is used to start the analysis from that
point. The last is the one which is called selective reanalysis. In this strategy, the parser
backs up to the most appropriate ambiguous point in the sentence which is determined
by considering all information available to the parser. Let’s consider some serial models
which are classiﬁed according to what kind of information is used to determine the order

of considering the alternatives of ambiguities.

2.1.1 Structural Preference Model

This model is also known as the local and minimal attachment principle which was
advocated by Frazier and Fodor(1978). They proposed that the human parsing model
consists of two stages. The ﬁrst stage is called the Preliminary Phrase Packager(PPP)
and the second stage the Sentence Structure Supervisor(SSS). The PPP assigns lexical
and phrasal nodes to groups of words within the lexical suing that is received. The out-
put of the PPP is shunted to the SSS which combines these structural phrases into a com-
plete phrase marker by adding higher nonterrninal nodes. The PPP is a shortsighted dev-

ice which looks at the input string using a narrow window (of the size of 6 words). Thus

17

the PPP analyzes the sentence locally and the global processing is the job of the SSS. It
is not clear how the PPP and the SSS are synchronized in detail except that the PPP goes
ahead of the SSS. From the fact that the PPP is shortsighted, there comes the local

attachment principle which can be described as:

An attachment can not be made to a structure that is out of sight of the PPP. If
there are more than one possible attachment points inside the sight of the PPP,

then they are equally ﬁne as the attachment point.

We can show the example of local attachment in sentence (2-1). In (2-1),
(2-1)John read the postcard, the memo, and the letter to Mary.
(2-2) John read the letter to Mary.

"to Mary" is preferably attached to "the letter" because another possible attachment
point "read" is out of sight of the PPP. But in (2—2), both "read" and "the letter" are
within the window of the PPP. Thus "read" and "the letter" are equally ﬁne as the attach-
ment point of "to Mary" (but actually "rea " is preferred by the minimal attachment
principle which will be explained below). Another example of the local attachment prin-

ciple can be found in (2—3) and (2-4):
(2-3)Ellen got the idea that the coast guard was going to send a life raft across.

(2-4) Though Martha claimed that she would be the ﬁrst woman president yesterday

she announced that she’d rather be an astronaut.

In (2-3), "across" is associated with "send" instead of "got" because "got" is too far
away. In (2-4), "yesterday" is associated with "announced" because "claimed" is not

available within the window of the PPP when "yesterday" is being processed.

After the local attachment principle is applied, the minimal attachment principle is
used if there is more than one possible attachment that is still left. The minimal attach-

ment principle can be described as follows:

18

Attach the incoming material into the phrase-marker being constructed using the

fewest number of nonterminal nodes.

This principle can be illustrated by looking at the processing of (2—5) and (2-6):
(2-5) The city council argued the mayor’s position forcefully.
(2-6) The city council argued the mayor’s position was incorrect.

When the noun phrase "the mayor’s position" is analyzed, there are two possible attach-

ment points as shown in Figure 2.2. It is clear that

‘\“
~

The city council V
arguedA AP

 

f f 11
the mayor’s orce u y

position
(a)

/\
A v/K

The city council
argued A /VP\

the mayor’s
posrytion

 

 

 

was lllCOI'l'CCI

(b)
Figure 2.2 Minimal Attachment

19

the attachment shown in (a) of the ﬁgure uses a smaller number of nonterminal nodes
than that in (b). According to the minimal attachment principle, the attachment in (a) is
chosen ﬁrst by the parser. In the processing of (2-5), this decision will prove to be
correct by the input that follows. But "was incorrect" in (2-6) will prove that the decision
of the minimal attachment principle is wrong. Then the parser should attempt reanalyz-

ing the sentence by ﬂying the other attachment.

It was said that the local attachment principle provides no preference in (2-2). Then
the minimal attachment principle is used to attach "to Mary" to "read" rather than to "the

letter", which can be illustrated using Figure 2.3.

N P/S>P\ NP/\/'\
/V A /PP\ V A

John read the letter to Mary

 

Figure 2.3 Failure of Local Attachment

Thus the local attachment principle is used ﬁrst and then the minimal attachment princi-

ple is used if there is still more than one possible attachment point.

In (Frazier and Raynor,1982), some experimental data was reported which further
strengthens their argument. In their experiment, the eye movements were recorded as the
subject read a sentence containing the attachment ambiguity. They found that shorter
reading time is taken for the sentences whose ambiguity resolution conforms to the stra-
tegy of the local and minimal attachment principle than for the sentences whose ambi-

guity resolution does not conform to this strategy.

They extend the concept of garden-path by incorporating into the group of garden-

path sentences the sentences whose temporary ambiguity is resolved into the wrong

20

analysis and then revised later by the disambiguating material. In other words, the sen-
tences containing temporary ambiguities which require reanalysis but do not require
conscious effort are also considered to be garden-path sentences. Other signiﬁcant data
they provided is that the reading time and ﬁxation duration in area (b) in Figure 2.4 (the
region where the ambiguity remains) is not longer than (a), and the difference in reading

times is associated with the disambiguating region (c).

‘— (b) _’5‘— (C)

l
I

I I I I

(a) _"

sentence onset of disambiguating end of
start . ambiguity matenal sentence
Figure 2.4 Ambiguous Region

This data supports the view that the existence of the ambiguity is not detected if it does
not garden-path (even if there is ambiguity) and the view of the serial model in which

only one analysis of the ambiguous material is computed at ﬁrst.

2.1.2 Lexical Preference Model

Ford, Bresnan and Kaplan(1982) (henceforth FBK) introduced a theory which
differs from the structural preference model in the previous section in how an ambiguity
is resolved. FBK argued that the lexical property of the head of a phrase is what deter-
mines the order of choosing the alternatives of an ambiguity. The sentences from (2-7)

to (2-9) can be used to illustrate this lexical preference theory :
(2-7a)The woman wanted the dress on that rack.
(2-7b)The woman positioned the dress on that rack.

(2-8a)The tourists objected to the guide that they couldn’t hear.

21

(2-8b)The tourists signaled to the guide that they couldn’t hear.
(2-9a)Joe included the package for Susan.
(2-9b)Joe carried the package for Susan.

In (2-7a), the PP, "on that rack", can be attached to either the VP headed by "wanted" or
to the NP headed by "the dress". (2-7b) has ambiguity similar to (2-7a). The only differ-
ence in (2-7a) and (2-7b) is in the main verb. But the attachment to the NP is preferred
in (2-7a) while the VP is the preferred attachment point in (2—7b). A similar situation
can be seen in (2-8). The only difference in (2-8a) and (2-8b) is also the main verb. In
(2-8a), the preference is that the phrase "that they couldn’t hear" is analyzed as the rela-
tive clause which is attached to the NP "the guide". But the secondary clause "that they
couldn’t hear" is analyzed as the complement clause of the VP headed by "signaled" in
(2-8b). (2-9) shows a similar ambiguity resolution as in (2-7). The ambiguity resolution
phenomenon shown in (2-7) to (2-9) can not be explained using the parsing model of
Frazier and Fodor because the sentences have the same syntactic structure but the ambi-

guity is resolved in different ways.

FBK sought the explanation of these phenomenon from the lexical component of
the grammatical system in the human sentence processing system. They argued that each
verb has various lexical forms which have different strengths and the strongest forms
determine the preferred syntactic analysis. Figure 2.5 shows the lexical form of the verbs
appearing in the above sentences. The ﬁrst form is stronger than the second form in each
verb. They did not explain what determines the strength of the lexical forms but they
conjectured that the frequency of usage might contribute in part. But this theory can not

be applied if the ambiguity is not related to a verb as in (2-10):

(2-10)That silly old-fashioned ......

22

want: < (SUBJ), (OBJ) >
< (SUBJ), (OBJ), (PCOMP) >
position: < (SUBJ), (OBJ), (PCOMP) >
< (SUBJ), (OBJ) >
object: < (SUBJ), (OBJ) >
< (SUBJ), (OBJ), (COMP) >
signal: < (SUBJ), (OBJ), (COMP) >
< (SUBJ), (OBJ) >
include: < (SUBJ), (OBJ) >
< (SUBJ), (OBJ), (PCOMP) >
carry: < (SUBJ), (OBJ), (PCOMP) >
< (SUBJ), (OBJ) >

Figure 2.5 Lexical Forms of Verbs

FBK used the "syntactic preference" principle to solve this problem. It was argued that
the alternative phrase structure categories in the expansion of a phrase structure have
ordered priorities which determine the preference of the analyses. In (2-10), the subject
NP analysis as shown in "That silly old-fashioned joke is told too often" has the higher
priority than the sentential subject analysis as shown in "That silly old-fashioned jokes
are told too often is well-known." Therefore the NP analysis is preferred to the 8
analysis. The following sentence (2-11) provides another example related to the current

discussion:
(2-11)We discussed running.

"running" in (2-11) can be analyzed as either a noun as in (2-12) or a verb as shown in

(2-13).
(2-12)We discussed the excessive running of races.
(2-13)We discussed running races excessively.

Because the noun category is ranked higher than the VP category, "running" in (2-11) is

ﬁrst analyzed as a noun.

23

Based on the two principles discussed so far (i.e. the lexical preference principle
and the syntactic category preference principle), FBK formulated the ﬁnal argument

principle that is related to the closure problem as follows:

ﬁnal argument principle (Ford, Bresnan and Kaplan,1982): Give low priority to
attaching to a phrase the ﬁnal argument of the strongest lexical form of that phrase
and to attaching any elements subsequent to the ﬁnal argument. Low priority is
deﬁned here with respect to other options that arise at the end position of the ele-
ment whose attachment is to be delayed.
Let’s explain this principle using (2-7a) and (2-7b): In (2-7a), "the dress" is analyzed as
the NP and it corresponds to the ﬁnal argument of the strongest lexical form <(subj),
(obj)> of "want". So the attachment of "the dress" to the VP "want" is delayed. Rather
the other option that the NP2, "the dress", is attached to the partial complex NPl is given
the higher priority as shown in Figure 2.6.

/" ‘
/

want the dress

Figure 2.6 Example of Final Argument Principle

(The reason for this analysis is that if the NP, "the dress", is attached to the VP, then it
will be closed and the coming PP, "on that rack", can not be attached to the NP headed
by "the dress".) Then the PP "on that rack" is attached to the NPl and the NH is
attached to the VP. In the case of (2-7b), the ﬁnal argument principle can not come into
play to delay the attachment of the NP "the dress" to the VP because the object NP is not
the ﬁnal argument of the strongest lexical form of "position". In this situation, the NP
"the dress" is attached to the VP rather than to a complex NP (this tendency was called

the invoked attachmcnt principle). Then the PP, "on that rack", is analyzed as a ﬁnal

24

argument of the strongest lexical form. Thus its attachment to the VP is delayed. There
are no more words following it and the attachment of the PP to the VP is the only
choice. The model described in this section can be considered to be a serial model
because the ﬁrst alternative that conforms to the theory is chosen but the other altema-

tives are never considered.

2.2 Parallel Models

Parallel parsing models, which are directly opposite to the serial parsing models,
also received much attention in psycholinguistics. One of the oldest parallel parsing
model seems to be the model proposed by Fodor, Bever, and Garrett(1974) who sug-
gested that all possible analyses are computed when an ambiguity is encountered. All
possible analyses are canied out in parallel. An analysis that is inconsistent with the fol-
lowing input is dropped. When a clause boundary is reached, one analysis out of a set of

the parallel analyses is selected based on some decision principles.

Forster’s(l979) autonomous parsing model can also be considered to be a parallel
model. In this model, phonological, syntactic, and message-level processors compose
the language processing system. Each processor operates autonomously but the input
into it is the information that is available in its level and the information generated from
the next lower level. This deﬁnition implies that the syntactic processor computes all
possible syntactic analyses as far as it can and the result is shunted to the next upper
level (which is the message-level processor). The analyses in a level can be dropped or
accepted by the next level processor. This model can explain the dumb, fast, automatic
process of each linguistic processing level such as the lexical access. Recently there has
been serious research on parallel models that challenge serial models that have until now

prevailed. These parallel models will be explained in the following sections.

25

2.2.1 Discourse Model

Crain and Steedman(1985) proposed a model that directly argued against the struc-
tural preference and lexical preference models(Frazier and Fodor, 1978; Ford, Bresnan
and Kaplan, 1982). This model is called the discourse model (or context model) because
it is argued that the access to the contextual information is what disarnbiguates the local
syntactic ambiguity. It is obvious that Crain and Steedman’s original idea is that the
plausibility based on the broad sense of semantics is the thing that works for the disam-
biguation. This broad sense of semantics includes both the general world knowledge and
contextual information. But their paper concentrated on the explanation of using only
the contextual information. They mostly argued for the inﬂuence of speciﬁc conversa-

tional context upon the syntactic disambiguation.

Crain and Steedman described two possible interactions between syntax and
semantics: the weak and strong interaction. In weak interaction, the syntactic component
from time to time proposes several alternatives of syntactic analysis to the semantic
component either serially or in parallel. Then the semantic component selects one alter-
native by comparing the evaluations or referents of the alternatives. In this interaction
the syntactic processing level independently proposes the alternatives of the local struc-
tural ambiguity, while the semantic level disposes among these alternatives. On the
other hand, in sn'ong interaction, semantics and context inﬂuence which alternatives are
actually proposed by the syntactic level in the ﬁrst place. Semantics and context can
either set the order of alternatives being proposed serially or make some alternatives

entirely unavailable.

Crain and Steedman’s choice out of the two interactions was the weak interaction.
They also argued that the alternatives should be proposed in parallel. They opted for a
very intimate interaction between syntax and semantics, for example, an interaction for
every word. This means that syntax appeals to semantics frequently and immediately.

They called this interaction the radical weak interaction. One important point here is

26

that the alternatives are proposed in parallel in contrast to the serial models such as
minimal attachment. The reason they went for the parallel model is that their experi-
ment did not show any residual effect of the structure when semantics and context sup-
port the two alternatives equally. The most general principle they proposed is described

as follows:

The principle of a priori plausibility (Crain and Steedman,1985): If a reading is
more plausible in terms either of general knowledge about the world, or of speciﬁc
knowledge about the universe of discourse, then other things being equal, it will be

favored over one that is not.

One speciﬁc knowledge they most speciﬁcally pointed out was the knowledge
about the referents that are present in the listener’s mental model of the universe of

discourse. Consider (2-14a) and (2-14b) to see this point:
(2-14a) Put [the block in the box] [on the table].
(2-14b) Put [the block] [in the box on the table].

If there is actually a block that is in the box or that kind of block has been mentioned,
then (2-14a) would be the preferred analysis, because there is a referent in the mental
model. But if the (2-14a) analysis fails to produce a referent, but (2-14b) succeeds in
ﬁnding a referent for "the block", then (2—14b) will be the analysis that is chosen. This

heuristic was stated as follows:

The principle of referential success (Crain and Steedman,1985): If there is a read-
ing that succeeds in refening to an entity already established in the hearer’s mental
model of the domain of the discourse, then it is favored over one that does not.
This principle is a special case of the principle of a priori plausibility. In Crain and
Steedman’s model, the source of semantics is the mental representation of a speciﬁc
conversational context including things that have been mentioned and things that have

been implied in the conversation rather than the world itself. In the cooperative

27

conversation, the bearer will introduce the referents into his or her mental model as well
as recognizing the reference to items already introduced in the conversation. The reason
for this can be explained by the cooperativeness between the conversants(Grice, 1975).
In other words, the bearer will add the referents for the presuppositions of the speaker
that are not satisﬁed with his mental model. Based on this consideration, another heuris-
tic is suggested which is a more generalized version of the principle of referential suc-

06882

The principle of parsimony(Crain and Steedman,1985): If there is a reading that
carries fewer unsatisﬁed but consistent presuppositions or entailments than any
other, then, other criteria of plausibility being equal, that reading will be adopted as
the most plausible by the bearer, and the presuppositions in question will be incor-
porated in his or her mental model.
This principle is still a special case of the principle of a priori plausibility. The principle
of parsimony can be used to explain the garden path effect in (2-15):

(2-15)The horse raced past the barn fell.

In (2-15), local syntactic ambiguity exists in the main-verb analysis or the reduced rela-
tive clause analysis of "raced". In the case of reduced relative clause analysis, the fol-
lowing presuppositions are required in the so—called null context: (1) the set of horses (2)
some horses were raced past the barn. But the main verb analysis does not need the
presupposition of (2). Therefore it is argued that the main verb analysis is chosen over

the reduced relative clause analysis and the garden path occurs.
(2-16)A horse raced past the barn fell.

In (2-16), no presuppositions such as (1) and (2) above are needed and thus it is expected
that the possibility of the occurrence of garden pathing will be less than (2-15). This

argument was supported by their experiment.

28

One important point that was argued by Crain and Steedman is that there can not
be the null context in the sense of "no horses has been mentioned before (2—15) or (2-16)
is told to the hearer". The reason they provided was that a null context like this can not
be neutral to the two possible analyses because the two readings may differ in the ease
with which the hearer may have to set up the referents because of the different number
of presuppositions the readings invoke. They said that the argument for the structural
preference model is actually based on the null context which is not a real neutral context
and thus the experiment data provided by the structural proponents lose credibility. They
argued that the result that the structural proponents obtained is actually due to the effect

of context.

Crain and Steedman also argued that the source of semantics in the analysis of (2-
15) is not the prior plausibility based on the general world knowledge but the presuppo-
sitions in the context of discourse. According to their theory, the control of the context

can not only induce the garden path effects but also overcome them. Consider (2-17):
(2-17)It frightened the child that John wanted to visit the lab.

The usual preference is that "that lab" is analyzed as a complement clause rather than
as a relative clause attached to "the child". But if the statement, "There was an explo-
sion", was mentioned just before (2-17), it was found that the preference switched. In

(2-18), the context even induced the garden path effect:

(2-18)Context: Several children heard an explosion.
It frightened the child that John wanted him to visit the lab.

2.2.2 Ranked Parallel Model

Gorell(1987) attempted an experiment to investigate the issue of serial/parallel
model related to the resolution of the local syntactic ambiguity. He used an experiment
called the syntactic priming paradigm. He argued that the experiment used by the serial

model proponents(i.e. the acceptability judgement task or the eye movement

29

experiment) is not appropriate for investigating the issue. In the case of the experiment
which used the acceptability judgement task reported in Frazier(1978), the acceptability
judgement is made at the end of the sentence, which makes it hard to examine what is
actually going on during the time between the onset of ambiguity and the appearance of
the disambiguating material. Frazier argued that the poor and slow performance for the
non-preferred reading of an ambiguity is due to the reanalysis of the input after the ﬁrst
analysis has failed. But Gorell proposed another possible explanation that can replace
Frazier’s claim. He argued that the poor and slow performance might be due to the com-
plexity of the structure that exists in the sentences in which non-preferred reading is the

correct reading.
(2-19)John knew the very old woman was ill.

In (2-19), the appearance of "was" makes the parser construct a clause which is a com-
plex task. According to Gorrell, this complexity causes the parser to become slow. He
attributed the poor performance to the processing load due to the complexity of the

structure rather than the requirement for reanalysis.

Frazier and Raynor(1982) reported an experiment that provided a stronger support
for the serial model. The experiment measured the eye ﬁxation and movement during
the reading of the sentences. Even for this data, Gorell suggested that the data can be
accounted for by the increased processing load due to the structural complexity. Gorell
suggested an experiment in which "knew" in (2-19) is replaced by "thought" as shown in
(2-20):

(2-20)John thought the very old woman was ill.

(2-20) is an unambiguous sentence having the same syntactic structure as (2-19). Gorell
argued that if (2-20) shows the similar poor performance as shown in (2-19) this shows
that his argument is correct. If the serial model is assumed, (2-20) should show better
performance than (2-19) because (2-20) does not require reanalysis. But unfortunately

there is no data available on (2-20).

30

Gorell’s experiment which used the syntactic priming paradigm is similar to the
experiment for the problem of lexical access done by Swinney(1979). Swinney showed
that a semantically ambiguous word can prime a target related to only one of its various
meanings in any biasing contexts. This experiment made parallelism of lexical access
widely accepted. Gorell argued that the syntactic category can prime a target that is
associated to this category. In his experiment, a target word is presented during the read—
ing of a sentence with a local syntactic ambiguity before the disambiguating words
appear. The subjects are asked to decide if the target is an English word or not immedi-
ately after the appearance of the target word. He said that a signiﬁcant priming effect
was observed for targets belonging to the categories predicted by the structure associ-
ated with the non-preferred reading of the ambiguity. He interpreted the data in this way:
the non-preferred reading is actually computed and proposed in parallel with the pre-
ferred reading. He also used unambiguous counter parts of the ambiguous sentences in
the experiment and found that the response time and correct ratio is almost the same as
the ambiguous counter parts with non-preferred reading. The use of unambiguous con-
trol like this excluded the possibility that the priming effect for ambiguous contexts was

the result of the rapid reanalysis.

Using the result of his experiment, Gorell proposed a parsing model dubbed the
ranked parallel model. In this model, all possible alternative analyses of a structural
ambiguity are computed and proposed in parallel to the next processing level. To reach a
reconciliation with the serial model proponents, he suggested that the multiple altema-
tives are proposed in parallel but they are rank-ordered according to the complexity of
the structure and they are considered by the next level in that order. The major point is
that syntactic level computes and proposes the alternatives in parallel but they are exam-

ined serially in the next level.

31

2.2.3 A Parallel Model with Conceptual Selection Hypothesis

Kurtzman(1985) made an attempt which argues for the parallel parsing model.
Instead of the structural preference (i.e. the local and minimal attachment principles),
lexical preference, and discourse preference used in other models, Kurtzman suggested
that expectation for the conceptual information is what determines the resolution of the
ambiguity. This was called the conceptual selection hypothesis. Consider a situation in

sentence (2-21) where someone is "keeping" something:
(2-21)The woman kept the dogs on the beach.
(2-22)The woman discussed the dogs on the beach.

Kurtzman argued that the conceptual information expressed by "kept" expects to have
some conceptual information about "where" (something is kept) in addition to the infor—
mation about "who" and "what". Because of this expectation, "on the beach" is associ-
ated with "kept" instead of with "the dogs". This seems to be a reasonable argument. It
is interesting to see how he handled the case in (2-22). In the situation expressed by "dis-
cussed", the conceptual information for the location where the discussion occurred is not
so strong as in the situation of "keeping" in (2-21). Rather the more elaborated informa-
tion of the object of discussion is what is expected conceptually. He argued that this dif-
ferent expectation is what causes "on the beach" to be attached to "the dogs" instead of

"discussed".

The information sought by the expectations that affect the resolution of ambiguity
in Kurtzman’s model is of a fully conceptual sort at the level of real world knowledge.
It is quite speciﬁc information which can not be mapped one to one onto some type of
structural semantic representation. One of the parsing approaches that utilizes concep-
tual information is the conceptual analyzer developed by Schank’s group (for example,
Riesbeck and Schank, 1978; Birnbaum and Selfridge, 1981). But it seems that their
representation language, the Conceptual Dependency, is too limited to be used for

representing conceptual information discussed in Kurtzman’s model.

32

The use of conceptual information in Kurtzman’s model seems to be similar to the
use of mental representation in Crain and Steedman’s discourse model but the former
model emphasizes the conceptual expectation as it can be shown in the sentences (2-
23a) - (2-23c):

(2-23a) The horse raced past the barn fell.

(2-23b) A horse raced past the barn fell.

(2-23c) Horses raced past the barn fell.

In (2-23a), "the horse" has the determiner and it is already in focus. Thus "the horse"
expresses old information. The participants in the dialogue have some information about
this horse. But "a horse" in (2-23b) or "horses" in (2-23c) expresses new information
whose nature is not known yet. It seems that there will be more expectation of new
information about an entity that is not known well. For this reason, we can say that
"raced past the barn" is more likely to modify the subject in (2-23b) and (2-23c) rather
than in (2-23a).

But this conceptual selection hypothesis is not out of trouble. Consider an example

in (2-24) and (2-25):
(2-24)We discussed running.
(2-25)We considered running.

In (2-24), it is preferred that "running" is analyzed as the simple noun (running as a
sport) instead of as a VP dominating a simple verb (engaging in running). But in (2-25),
the preferred resolution is the VP. As Kurtzman acknowledged, the resolution does not
seem to be based on the expectations for particular conceptual information in these sen-
tences, because we can discuss "running" as a sport as much as "running" as "engaging
in running". What matters in this case is the subtle difference in relative plausibility. It
seems to be more plausible to discuss a sport of running than to discuss engaging in run-

ning. But Kurtzman argued that adding the general plausibility as a source of

33

disambiguation does not undermine his conceptual selection hypothesis model because
there are cases which the general plausibility can not handle as shown in (2-26) and (2-
27). In (2-26) "discussing on the beach" is as plausible as "dogs on the beach", which is

true in (2-27), too.
(2-26)The woman discussed dogs on the beach.
(2-27)The woman kept the dogs on the beach.

Another difﬁculty for the conceptual selection hypothesis can be found in (2-10).
The general preference is that "that" in (2-10) is analyzed as a determiner rather than as
a complementizer. His explanation for this case is based on the following observation:
Analyzing "that" as a determiner (which begins a main clause’s subject NP) invokes an
expectation for a term referent but analyzing it as a complementizer (which begins a
complement clause) invokes an expectation that the topic of the sentence is a proposi-
tion. He argued that the better conceptual expectation is to take a simple object or idea
rather than a whole proposition as the topic of a sentence. This explanation seems to be

at least as good as that given by other parsing models introduced earlier.

So far we examined the conceptual selection hypothesis. Now it is necessary to
consider how the conceptual information level works with the syntactic analysis level in
Kurtzman’s parsing model. In the Schankian conceptual analyzers, syntactic analysis is
held to a minimum and conceptual analysis is given the major role. But Kurtzman pro-
posed a model similar to the autonomous model. All possible syntactic analyses are
computed on the syntactic level and this result is input into the level that is in charge of
handling conceptual information. Then the conceptual level chooses one analysis which
is most appropriate based on the conceptual expectation. It was argued that it would be
hard to conceive a model in which the syntactic processing is interrupted by the concep-
tual level so that the parser can consider only one syntactic analysis (out of several) that
is most appropriate conceptually. The reason for this is that there are no systematic rela-

tions between the conceptual information and the syntactic information which could be

34

used in such a direct conceptual determination of the parsing process. Kurtzman also
objected to the scheme in which one analysis is proposed by the parser and the concep—
tual mechanism either rejects or accepts it (in the case of rejection, the parser proposes
the next possible analysis, and so on. Note that this is the approach used by the parsers
of interleaved semantic processing). His reason for the objection is that there can be so
many single analyses of the unambiguous and generally plausible sentences that fail to

express the conceptual expectation.

The architecture of the model he proposed is the so called, Immediate parallel
analysis(IPA) with strong parallelism. "Immediate parallel analysis" means that all pos-
sible analyses are computed from the onset of the ambiguity instead of delaying the
analysis (which is called the delayed parallel analysis as shown in (Marcus,1980)). The
strong parallelism means that each of the possible analyses is canied on until it is
removed based on further input material. The opposite of the strong parallelism is called
momentary parallelism. In momentary parallelism, the ambiguity is resolved at the onset
of the ambiguity. In other words, one analysis is chosen immediately out of the multiple

possible analyses based on some selection criteria.

Kurtzman’s parsing model can be described as follows. At an ambiguous point, all
possible analyses are computed. Out of these alternatives, those that are best according
to conceptual expectation are retained and the others are removed. If the conﬁdent deci-
sion can not be made because of the lack of conceptual information, the decision is
delayed until it is possible to. The analyses that have not been removed stay active as the
parsing is going on.

The reason that strong parallelism is adopted instead of the momentary parallelism

can be explained using sentences (2-28) and (2-29):
(2-28)The executives discussed the problems in the factory.

(2-29)The executives discussed the problems in the restaurant.

35

In the case of the momentary parallelism, the attachment point of the PP (even if it is not
complete yet) should be decided as soon as "in" is read. But this is not what actually
happens. In (2-28), the PP is attached to the NP "the problems", while it is attached to
the VP in (2-29). In other words, the attachment point is determined by the content of
the PP, which implies that the ambiguity can not be resolved at "in" but after completing
the PP. This example shows that strong parallelism should be adopted instead of the
momentary parallelism. The advocates of the serial model (for example, the Frazier and
Fodor model) might argue that a single analysis is chosen ﬁrst and backtrack is done if it
proves that the selection was wrong. But note that neither (2-28) nor (2-29) are garden-
path sentences which require backtracking (but the matter becomes complicated if we
consider that the slight garden-path sentences do not require the conscious backtrack-
ing).

The important question that needs to be investigated under strong parallelism is the
point at which the ambiguity is resolved (note that the source of disambiguation in
Kurtzman’s model is conceptual expectation). Kurtzman conjectures that the point of
disambiguation might be when the conceptual mechanism can conﬁdently evaluate
which analysis is an appropriate one. The resolution point can vary according to the lex-

ical content of the following input material. One example for this phenomenon can be

found in the following sentences.
(2-30)The horse raced past the barn fell.
(2-31)The aluminum strengthened to a great degree.
(2-32)The children observed here our plant ﬁelds.

In each of the above sentences, the ﬁrst verb can be analyzed as either the main verb of
the sentence or the past participle leading the reduced relative clause. After reading this
verb, the experimental data reveals the following result: in (2-30), the main verb
analysis is the prefered analysis and garden pathing occurs; No garden-path effect occurs

in (2-31); But in (2-32), garden-pathing occurs in favor of the reduced relative analysis.

36

This data suggests that the decision can be made prior to the reception of the disambi-

guating material.

2.3 Summary

In this chapter, the important parsing models proposed in the psycholinguistic
literature have been studied. The two major features that distinguish the models are
serial/parallel computation of the syntactic alternatives and the type of information that
determines the preferred analysis. In the serial model, the parser computes and proposes
the syntactic alternatives serially. In the parallel model, the alternatives are computed
and proposed in parallel and it is up to the next processing level to choose one out of the

multiple analyses of an ambiguity.

Information that is used to select an alternative is important in deﬁning a parsing
model. The complexity of the structure is used in Frazier’s minimal attachment, the lexi-
cal preference is used in Ford, Bresnan and Kaplan’s model, discourse information is
used in Crain and Steedman’s discourse model, and conceptual expectation is used in
Kurtzman’s parallel model. Note that the models that use semantics to resolve local syn-
tactic ambiguity go for parallelism. (Semantics here does not include the lexical infor-
mation used in the lexical preference model). One of the reasons for this tendency is the
autonomy of language processing. The input into the syntactic level consists of the out-
put of the next lower level(lexical level) and the syntactic knowledge available in the
syntactic level. The syntactic level does whatever processing it can and the output is
input into the next higher level (let us assume this is the semantic level). Because the
information available to the syntactic level and the semantic level is so distinct it is not

likely that the syntactic level’s processing is interrupted by the semantic level.

Gorell used an on-line experimental technique to show that parallel processing is
actually going on at the syntactic level. But it is too early to decide that one model is to

be chosen over another model.

CHAPTER 3

FRAMEWORK OF THE NEW APPROACH

We will describe the overall framework of the SYNSEM parser which was pro-
posed in Section 1.4. The objective of the parser is to update the approach of interleaved
semantic processing to achieve integrated parsing by approaching from syntax-oriented
parsing. The major attraction that integrated parsing has is to use semantic information
as much as possible. The previous integrated parsing advocates argued that the full use
of semantics can be achieved only by making the semantic analysis the major analysis
and restricting the amount of syntactic analysis. Integrated parsing could not be con-
sidered to be an approach that uses both full-ﬂedged syntax and semantics. The fact that
the integrated parsers are biased too much toward semantics was corrected by Lytinen’s
new integrated parsing approach which was shown in the MOPTRANS parser (Lytinen,
1984). But his approach still puts more emphasis on semantics and uses syntax only as

the ﬁlter that applies after the semantic analysis.

We take the position that the syntactic processing level is lower than the semantic
processing level and is an autonomous processing level which runs more or less
independently. We are not aware of any psycholinguistic research that claims that the
semantic processing is on the lower level than the syntactic processing. Thus it seems to
be natural to start the design of a parser from the syntax-oriented parser and then add
more on-line semantic processing to it. This approach resulted in the interleaved seman-
tic processing approach. But the use of semantics in interleaved semantic processing is

not enough to satisfy the advocates of integrated parsing.

37

38

Kurtzman’s conceptual selection hypothesis and the immediate parallel analysis with
strong parallelism provides the basic framework for the new parser that will be
developed in this thesis.

3.1 General Framework of the Control

The starting point to approach the desired parser is interleaved semantic process-

ing. The control loop for interleaved semantic processing is shown in Figure 3.1 (let’s

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

call this model ISPl).
l <
step- 1 :
ﬁnd syntactic rules by pattern matching
J
'I
step-2: Choose an untried rule update parsing
II
step-3: compute semantic processing
corresponding to this rule
step-4: semantically reasonable?
Ye ‘

   

 

 

Figure 3.1 Interleaved Semantic Processing - 1
(Loop terminated at end of sentence)

It is clear from this ﬁgure that the processing is compatible with the serial models dis-
cussed in the previous chapter. If the rules found at step-1 are sorted and tried in the
order of the simplicity of the structure at step-2 , then the approach shown in this ﬁgure
is the same as the structural preference model such as minimal attachment. If the rules
are tried in the order of lexical preference in step-2, this approach would be the same as
the lexical preference model. It is also conceivable that a parser adopts no speciﬁc selec-

tion strategy in step-2 but rather a next rule is chosen randomly from the rules that have

39

not been tried yet. This is the method that was usually used in natural language parsers

that used interleaved semantic processing.

The parallel parsing models suggested that all rules are proposed and processed in
parallel. We can take this suggestion as one way of changing the control strategy of
interleaved semantic processing. The control strategy updated according to this sugges-

tion is shown in ﬁgure 3.2 (let’s call this model ISP2).

 

 

 

 

 

 

 

 

 

processing processing processing

: . . . i
u do semantic do semantrc do semantrc :
I

t I
I I

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

step-3: select one with best semantic result

I

step-4: update parsing
l

 

 

 

 

 

 

Figure 3.2 Interleaved Semantic Processing - 2

All rules (or syntactic alternatives) found by the syntactic processor at step—1 are pro-
vided to the next higher level at step-2 and these alternatives are processed and com-
pared in parallel according to the information available in this level. This information
will be the discourse context in the discourse model and conceptual information in
Kurtzman’s model. In the case of Gorell’s model, this information would be the struc-
tural complexity (provided by the lower level, the syntactic level) rather than the sort of
semantic information. Based on the processing at step—2, one rule that has obtained the
best score at step-2 will be chosen at step-3. This rule will be used to update the parsing

'at step-4 and the parsing goes to the next round of the loop.

40

The improvement in going from ISPl to ISP2 is that the alternative with the best
semantic result can be selected. Thus the probability of making an error in selecting one
alternative can be reduced compared with ISPl. ISPl has the step of checking the
semantic effect of the chosen alternative but it is not an appropriate architecture for
selecting the best alternative. As far as the speed of the parsing is concerned, it is not
clear that ISP2 is better than ISPl. It is usually acknowledged that the semantic process-
ing takes more computation than the lower level processing such as the lexical or syn-

tactic processing. But ISP2 encourages more semantic processing than ISPl.

According to Kurtzman’s IPA with strong parallelism, the point of disambiguation
depends on the time that a conﬁdent decision can be made based on conceptual informa-
tion. This means that a decision may not be made just after ﬁnding the rules and consid-
ering the semantic effects of these rules (at step-2 and step-3 in Figure 3.2). It may be
several words later that one alternative can be determined to be better than the other. In
some cases two rules might have the same semantic processing result or in other cases
some rules might never invoke semantic processing but only the build-up of the syntac-
tic structure. These considerations indicate that the output of step-3 in Figure 3.2 may
have to be more than one rule. This implies that there might be more than one way of
parsing (called the branches of parsing) active at the same time. The control ﬂow shown
in Figure 3.3 (named to be ISP3) takes these considerations into account. Note that there
can be more than one branch at the beginning of the control loop. Finding more than
one rule at the pattern matching step(step-1) means that the number of branches will be
increased by that amount (if only one rule is found, the number of branches would not
change).

But there is a problem that needs more consideration in ISP3. The problem is
related to the question of when the branches are compared and ﬁltered. According to
Figure 3.3, the point of comparing and ﬁltering the branches comes always right after

the semantic processing is done for the rules found during the pattern matching step.

41

 

 

 

"-4 ........ r .................... + ....... ..

 

 

step-l"--

. ﬁnd rules ..... ﬁnd rules

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

I

! compute ..... compute ..... compute compute I

I semantrcs $61113 llthS semantrcs l

L - - - .. f1 ................................................ J
I v

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

i I
update ..... update ,,,,, update
brgrrrch

 

 

 

 

Figure 3.3 Interleaved Semantic Processing - 3

This might cause the system to compare the branches even if insufﬁcient semantic pro-
cessing results have been collected, which also means that the branches might be com-
pared too frequently. Based on this consideration, the control is modiﬁed as shown in

Figure 3.4 (this is named to be ISP4).

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

I—
: l ' I I r——
ﬁnd rules ﬁnd rules
compute compute compute compute
semantrcs semantics semantrcs semantrcs

 

 

 

 

 

 

 

 

 

 

 

   

 

 

semantic comparison stage

 

 

 

 

 

 

 

 

 

 

Figure 3.4 Interleaved Semantic Processing - 4

Looking at the ﬁgure, we can notice that another loop is added in the control. In the

42

inner loop, each branch goes through the following steps: (1) ﬁnd all rules whose pat-
terns match the working memory (called the pattern matching step); (2) for each rule
found in ( 1), a new branch is created and the semantic effect of the rule is computed and
the parsing state of the branch is updated; (3) check to see if the branch needs to read a
new input word and go into the semantic comparison stage if it does (if the branch does
not need a new word, the control returns to the beginning of the inner loop as shown in
the ﬁgure). The branches that reach the semantic comparison stage wait until all the
branches reach this stage. When all branches in the system have reached the comparison
stage, the actual processing of the stage begins. During this stage, the branches are com-
pared using the semantic processing result collected after the last semantic comparison
stage. Only those branches with the best semantic processing result are kept and others
are thrown away (they are actually stored in the backup stack). If there is a branch
which has no semantic processing result collected since the last semantic comparison
stage, this branch is also kept as an active branch. The detailed operations and imple-

mentation of ISP4 will be explained throughout the thesis.

3.2 Conﬁguration of the Parser

Figure 3.5 shows the conﬁguration of the SYN SEM parser. The program modules
are indicated by the ellipses and the data structures are enclosed in the boxes. The regu-
lar arrows indicate the ﬂow of control and the access of data structures is shown with the
dotted arrows. The monitor(MON) is in charge of reading each input word and calls the
other program modules to process the input. It supervises the control of the parser in a

synchronized way to successfully achieve the parsing.

The syntactic analysis component(SYN) is the program module that does all the
processing related to syntax. For example, matching the patterns of the syntactic rules
(henceforth, "rules" means the syntactic rules) against the state of parsing to ﬁnd the

rules to be triggered and building the syntactic trees are handled by SYN. The semantic

43

 
    

Other
Routines

 

 

 

SWM Rule Base KB

 

 

 

 

 

 

 

 

 

Figure 3.5 Conﬁguration of SYN SEM

processing component(SEM) builds a semantic interpretation for the input sentence
based on the analysis of SYN. The most important function of SEM to be discussed in
this thesis is to ﬁnd the paths in the knowledge base to get semantic information and
compare these paths to ﬁnd the best alternative out of multiple analyses proposed by
SYN. The marker passing module(MP) is the routine to pass the markers starting from
the concept representing input words. It detects the collisions so that SEM can use

them.

SYNSEM uses three major data structures: the syntactic working memory(SWM),
the rule base(grammar rules), and the knowledge base(KB) which is a semantic network.
The SWM consists of two lists called CLIST and CASH and some global variables.
CLIST contains the syntactic trees that are being built during parsing. CASH is the auxi-
liary list that stores syntactic trees temporarily. The rule base contains all grammar rules
which represent the grammatical knowledge of the parser. The grammer rules used in
SYNSEM are similar to those of Parsifal (Marcus, 1980).

The rules are written in a domain independent way so that the syntactic portion of
the system can be portable between different domains. The knowledge base(KB) is a

semantic network written in the KODIAK knowledge representation language (Norvig,

44

1986; Wilensky 1987; Wilensky & etc. 1986). The knowledge base encodes world
knowledge that the system knows about the world with which it is dealing. The semantic
power of the system depends on knowledge the knowledge base holds. The knowledge

base is the part of the system that is dependent on the domain.

Currently SYNSEM can only handle single sentence input. In other words, it can
not use the context available in a text consisting of multiple sentences. The output of
SYNSEM is a set of knowledge base paths which represents the meaning of the input

SCIltCI‘lCC.

3.3 Flow of Parsing in SYNSEM

In this section, the overall operation of the SYNSEM parser will be described so
that we can see how it operates globally. A word of the input suing is always read into
the system by MON(Monitor). If this word is an open word, MON calls the MP with the
word being passed. Open words are the words that can represent an entity in the world

such as objects, actions, or states. For example, "train , apple" and "eat" are the open

words while the prepositions, articles, etc. are examples of closed words.

After marker passing is done, the word goes through the morphological processing
and it is pushed into CLIST. This changes the content of CLIST, which might cause
some rules’ patterns to match CLIST. Then for each branch in the parser, MON calls
SYN to do syntactic analysis. SYN matches the patterns of the rules in the rule base
against CLIST. SYN returns the rules found during the pattern matching to MON. Note
that more than one rule can be found if there is ambiguity. For each rule returned by
SYN, a branch of parsing is created and the actions of the rule are executed. At this
point some action might call the semantic processing component to do semantic process-

ing corresponding to the rule.

Note that some rules may not have any action that invokes semantic processing.

These rules (thus the branches) have no semantic processing result after the execution of

45

the actions. For example, the rule Parse-dot which processes the word "the" has no

semantic processing. It only attaches the word "the" to a node NS which is newly

 

 

created (see Figure 3.6).
118
t
”I :>
the Rule: Parse-det I

the

 

 

Figure 3.6 "the" Processing

It should be noted that the execution of the actions of a rule does syntactic process-
ing such as updating the content of the SWM as well as calling the semantic component
to invoke some related semantic processing. The details of the actions of the rules will
be explained in the next chapter. The syntactic processing can be divided broadly into
two tasks: ﬁnding rules whose patterns match the SWM of the branch, and constructing

the syntactic trees in the SWM.

If there are some rules found during the pattern matching, the branches will be
created for these rules and each branch executes the actions of the rule and the process-

ing starts from the pattern matching again.

If there is no rule found during the pattern matching of a branch, the parsing has
encountered a blocked situation. A blocked situation can happen in two ways. One
blocked situation is not an actual blocked situation but it represents the situation where a
new input is required to be input into CLIST. Let us call this the read-again situation.
This happens when all possible processing has been done on the material in CLIST and
thus no rule can be found that matches CLIST. The other blocked situation happens
when the analysis corresponding to the branch is not compatible with the input string.
Because of the incompatibility of the input string, no grammar rule matches CLIST.
This is called the dead-end situation of the branch. MON can determine which blocked

situation it is when no rule matches CLIST (refer to Appendix B). If the blocked

46

situation is not the dead-end, MON puts the element in CASH into CLIST if CASH is
not empty. If CASH is empty, it is required to read the next input word. This situation is
the point where the branch reaches the semantic comparison stage. The branch is
pushed into the waiting pool so that the branches can be compared and ﬁltered when all

branches reach the stage.

When all branches get to the semantic comparison stage, the branches are com-
pared and ﬁltered according to the semantic processing results that have been collected
since the last semantic comparison stage. At this stage, the branches whose semantic
results are worse than other branches are removed (actually stored into the backup stack)
and the other branches are kept as active. Note that the branches with no semantic pro—
cessing result since the last semantic comparison stage are just retained without com-
parison. Then MON reads a new word from the input string and pushes it into CLIST of
each active branch after morphological processing. From this point, each branch ﬂows as

explained so far.

Let’s consider Figure 3.7 to see the creation and removal of the branches.

 

 

processing point processing point
p1 P2
Figure 3.7 Lives of Branches

The branch b in Figure 3.7 is removed because it encounters the dead-end situation at
p1. This is one way in which syntactic information is used for disambiguation. Branch a

forks into three new branches (a1, a2, and a3) because three rules are found during the

47

pattern matching of the rules while branch c gets only one rule and thus one branch is
forked. The rule matching and execution process is repeated until all branches request a
new input word. At processing point p2, let’s assume that all the branches reached the
semantic comparison stage. At this stage, branch a2 and branch c are removed because
their semantic processing results are worse than that of branch a3. But branch a1 is kept
without comparison because it has either the same semantic processing result as branch

a3 or no semantic processing result.

It is reiterated that a branch is completely thrown away if it proves to be a wrong
syntactic analysis (this appears as the situation in which no rules are found during the
pattern matching and it is determined to be the dead-end situation). This inactivation is
because of the syntactic reason. If the branch is not one of the branches with the best
semantic processing result during the semantic comparison stage, the branch is inac-
tivated but stored in the backup stack (if the semantic processing result is too bad, the
branch is immediately thrown away without being stacked). This inactivation is because
of the semantic reason. Thus, the disambiguation is done for both syntactic and semantic

reasons.

Usually all the branches except one will be removed during the semantic com-
parison stage, which means that the ambiguity is resolved completely. The period during
which there is more than one branch (i.e. there is ambiguity) is usually short. Thus, the
whole parsing consists of some periods during which there is only one active branch and
other periods which have multiple active branches. This phenomenon is called partial
parallelism and is illustrated in Figure 3.8. The branches that drop off during the
semantic comparison stage are stored in the backup stack so that they can be used when
it is necessary to do backtracking. A push down stack is used as the storage so that the

most recently stored branch can be used ﬁrst in the case of backtracking.

When all branches in the parser get to the dead-end, the parsing can not go any

further. This situation is the dead-end of total parsing (note that this dead end is more

48

 

R i/J

parallel
stages

Figure 3.8 Partial Parallelism
serious than that of a branch). In this situation, the parser should back up to some previ-
ous point. This is done by popping one image of a branch that was stored in the stack
and using it as the active branch. Figure 3.9 shows the ﬂow of the parser in a precise

form.

3.4 Conclusion

In this chapter, the general framework of the new parser with improved interleaved
semantic processing has been developed. The basic approach is to make the parser con-
sider all possible alternatives of syntactic processing at each step of parsing. This
resulted in a system that allows multiple branches of parsing to exist at the same time.
The disambiguation is done by removing the branches for one of two reasons. First, any
branch proven to have a semantic processing result inferior to any other branch is
suppressed during the synchronized semantic comparison stage. Second, any branch that
is not compatible with the following input material is removed. This is realized in the

way that the branch ﬁnds no rules during its pattern matching step.

The ﬁnal version of the parser studied in this chapter corresponds to Kurtzman’s
parsing model which adopts the IPA with strong parallelism with the conceptual selec-
tion hypothesis. The conﬁguration of the parser has been introduced and its detailed

operation has been explained.

49

 

step 1:

step 2:
step 3:

step 4:

read a word and push into SWM of each active branch;
(if all words in sentence has been processed, then stop.)
make a queue Q consisting of all active branches;
if Q is empty, goto step 4;
pop a branch from front of Q;
find rules that match SWM of the branch;
if no rules found,
then if read-again situation
then if CASH is nonempty
then read from CASH
else push this branch to All_done_queue
else throw away this branch

else
begin
for each rule found
do begin
execute actions of the rule;
fork a new branch for this rule:
insert this new branch into rear of Q;
end
end;

goto step 3;

(Semantic Comparison Stage)

compare the branches in All_done_queue and
choose the branches with best semantic results;
(All branches chosen become active ones and
others are thrown away.)

goto step 1;

Figure 3.9 Flow of the Control

 

CHAPTER 4

SYNTACTIC ANALYSIS

The design of the syntactic analysis component has been inﬂuenced by the three
considerations in SYNSEM. The ﬁrst consideration is related to parallelism. The system
should synchronize multiple branches that run in parallel and try to remove some
branches based on both syntactic and semantic information. Thus SYN should be
designed in such a way that management of parallel branches is relatively easy to imple-
ment. The second consideration is the early attachment strategy. A partial NP is created
until the head noun is found. Then this partial noun phrase is treated as if it is the com-
pleted noun phrase. Thus the partial noun phrase can be attached to other structures as
soon as its head noun is known. The NP’s post modiﬁers after the head noun can then be
attached to the NP incrementally. The third consideration is to delay some syntactic
decision until the context is collected enough to make the correct decision. Based on

these considerations, the rule-based syntactic analysis component has been developed.

4.1 Syntactic Working Memory

The syntactic working memory(SWM) is the data structure which contains the syn-
tactic representations that are being built during the parsing of the input sentence. Two
major components of the SWM are the two lists called CLIST and CASH. CLIST and
CASH contain the syntactic trees. CASH is storage which contains syntactic trees that
have been pushed out of CLIST temporarily as illustrated in Figure 4.1. When a word is
input to the system, it is pushed onto the right side of CLIST after going through mor-
phological processing (let us just distinguish the two ends of a list as the left and right

ends). The actual entity that is pushed is a node representing the word. A node in the

50

51

CLIST

trel tre2 tre3
A <—_: 000

input string

 

 

(‘ASI—I

AA

Figure 4.1 Syntactic Working Memory

 

trees is called an Snode which stands for a "syntactic node". The content of an Snode is

shown in Figure 4.2. It shows a node for the word "eat" and "rocks".

 

 

 

 

 

 

 

 

.Snode ,Snode
type=V =(Noun or V)
parent=nil parent=nil
son=nil son=nil
word=eat word=rocks
Figure 4.2 A Syntactic Node

Note that an Snode is itself a tree with one node. Syntactic trees are made with the

Snodes. The structure of a syntactic tree is illustrated in Figure 4.3.

As explained above, each word that is input by MON is pushed into the right side
of CLIST after it is converted into an Snode by the morphological routine. The right-
most tree of CLIST can be popped out and pushed into CASH by the request of an
action of a rule. This request is made if the rule used the rightmost tree just to determine
what analysis should be done for the second rightmost tree in CLIST (i.e. the rightmost
tree is used only for providing context for making decisions related to the second right—

most tree). After the execution of this rule, that rightmost tree need not be in CLIST

52

 

name= S1

WPC= S
parent=( )

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Figure 4.3 Tree in SWM

because the purpose of its being in CLIST has been fulﬁlled and the second rightmost
node needs to become the rightmost tree of CLIST for further processing related to it.
This fact will become clear when the format of rules is explained in the next section. So
one of the actions of a rule can be the request for the transfer of a tree from CLIST to
CASH. When the system needs to read a new input into CLIST, then the tree in CASH
is pushed into the right side of CLIST if CASH is not empty. If CASH is empty, the

input should come from the input string.

A new tree can be created and put into CLIST (usually a tree of one node is
created). For example, the nodes of type "S", "NP" and "VP" are the nonterminal nodes
which are created by the system. One important operation related to trees is attaching a
tree in CLIST (usually the rightmost tree) to another u'ee in CLIST so that a bigger tree

can be constructed.

53

The name CLIST reminds us of the data structure called "clist" used in the concep-
tual analyzers. But there is only a weak relation. "Clist" in conceptual analyzers is a list
of semantic objects that are being built. CLIST of SYNSEM is a list of purely syntactic
trees that correspond to the phrase structure trees of the sentence. CLIST is similar to the
combination of the C-stack and the lookahead buffer in Parsifal (Marcus, 1980). The
stack and lookahead buffer is merged into one list in SYN SEM. The right side of
CLIST corresponds to the buffer cells and the left side corresponds to C—stack in Parsi-
fal. However there is no preset boundary in CLIST that divides it into two parts. The
use of the SWM will become clear from the examples that will follow later.

4.2 Structure of Rules

As mentioned before, SYNSEM is a rule-based analyzer that is similar to Parsifal.
In this section, the mles used in SYNSEM will be explained in detail. Figure 4.4 shows
the format of a rule in SYNSEM.

 

 

 

 

 

 

 

C Prioﬁty J
Pattern-part /
node - node . node - node I, tric
<chain Icstrrc > tghain restrrc] [chain restnc ] [chain es
base < >
I‘— pattern —’ raw patterns I
|‘—"—_ additional restrictions _—’ I
Action-part
[ a1 :1 E a2 1 [ a3

 

 

II

I < actions

 

Figure 4.4 Format of Rules

54

A rule is composed of the priority speciﬁcation, the pattern part and the action part. The
priority speciﬁcation part speciﬁes the priority of the rule relative to other rules. The
priority is speciﬁed by enumerating the rules which has higher priority than this rule and
the rules which has lower priority than this rule. Note that numerical number is not used
for the priority speciﬁcation in SYN SEM. It is either an empty list or a list of the form,
((1 rulel rule2 ...) (h rule-a rule-b ...)). The ﬁrst case is is used when there is no rule
which has relative priority relationship with this rule (most of the rules are in this case).
In the second case, the sublist starting with "1" contains the rules which should have
lower priority than this rule and the sublist starting with "h" contains the rules which
should have higher priority than this rule. The priority is used when more than one rule
is found during the pattern matching. From the set of the rules from the pattern matcher,

any rule with lower priority than a rule in the set need not be considered and removed.

The pattern part of a rule speciﬁes the condition that should be satisﬁed by CLIST
in order for the rule to be selected during the pattern matching. We also say that the rule
matches CLIST if the pattern part is satisﬁed by CLIST. The action part of a rule consists
of actions that need to be executed when the rule matches CLIST. The pattern part con-
sists of two parts: the patterns and the additional restrictions. The patterns are matched
against the trees in CLIST. There are two kinds of patterns. One is the base pattern and
the other is the raw pattern. One pattern is matched against one tree. A raw pattern is a
pattern which needs to be speciﬁed by only stating a property of the root node of the
corresponding tree. This speciﬁcation can only be done by either naming the type of the
node or specifying a feature of the node. Note that the speciﬁcation is only about the

root node of the tree.

There can be zero or one base pattern in a rule. In the base pattern, other nodes in
addition to the root node of the corresponding tree can be involved in the speciﬁcation.
This means that the inside of the tree is examined using the base pattern. The base pat-

tern is usually used for checking the state of parsing. The state of parsing is implicitly

55

represented by the trees in the SWM. By examining the syntactic trees, the state of pars-
ing can be gleaned. In the ATN parser, the state of parsing is represented by the state
node of the ATN to which the parsing has reached. Parsifal used the active packets to
represent the state of parsing. In SYNSEM, the rule base is nor partitioned into packets
as it is in Parsifal; the reason will be explained later. The root node of the tree that is
matched against the base pattern is usually of type S(clause node). The state information

can be checked by specifying the right side of the S tree as shown in Figure 4.5.

(‘I .IST

  
 

speciﬁed in P
base pattern

 

 

Figure 4.5 Speciﬁcation of the Base Pattern

The base pattern can specify only a part of the tree that corresponds to this pattern. Note
that it is not necessary to specify the shape of the full tree. This leads to the fact that
matching the base pattern to a tree won’t be computationally too costly. Let’s take an
example using Figure 4.5. For the rule that attaches a PP to a VP, the only portion of the
S node that needs to be speciﬁed in the base pattern is "VP-S" as shown in the ﬁgure. It
is not necessary to specify the NP or V node in the tree. The PP can be attached to a VP
without worrying about other parts of the S tree because the existence of the VP guaran-
tees that there is an NP under 8 (in the subject position) and a V under the VP. The base
pattern should always be the leftmost pattern if it exists in a rule. The number of raw
patterns can be from one to three and they are put consecutively on the right of the base

pattern.

56

A pattern consists of two parts: the node chain and the node restriction as shown in
Figure 4.4. The node chain speciﬁes the skeleton of the portion of the tree (actually the
right portion of the tree) that corresponds to the pattern. The node restriction speciﬁes
the restrictions that the nodes on the tree should satisfy, such as some speciﬁc features.
The node chain for the raw pattern should consist of only the node type of the root node
of the tree or the basic property of the root node. The node chain of the base pattern con-
sists of a set of chains of nodes such as "(VP-l S-l)"(the number sufﬁx will be explained
shortly).

The additional restrictions of the pattern part of a rule consists of a set of restric-
tions that the corresponding trees should satisfy. Each additional restriction speciﬁes the
constraint that is related to two nodes that are in the different patterns of the rule. As an

example the main-verb2 rule is shown in Figure 4.6.

(main-verb2 ()
[ ( < (np-l s-1) 1 ((link subj)) > ,‘basepatternﬁndicatedbyn
[ (v—1) 0 (main (not aux)) ] ) ;rawpattern(indicatedby0)
,' additional restrictions

( [equal (feature-of num np-l) (feature-of num v-l)]
) l
[ (create vp-l) ,'acti0n1
(attach v-l to vp-l) ,‘actionZ
(transfer-features v-l to vp-l)
(attach vp-l to s-l)
(sem np—l v-1 subj)
(if (have-feature intransitive in v-l)
(add-feature struc-ok to s-l))
(connect_main_sec2 v-l) ]

Figure 4.6 Main-verb2 Rule

This rule matches CLIST when the subject NP of a clause is followed by a verb which is
in the form of a main verb. This NP should already be attached to the S node and this is
checked in the base pattern of the rule (see the node chain "(np—l s-l)"). The ﬁrst raw
pattern checks to see if the node of the rightmost tree is a verb which is not an auxiliary

verb. The additional restriction checks the agreement of the number feature of the NP

57

and the verb. This rule actually interprets the NP as the subject of the sentence whose
main verb is the verb corresponding to v-l. In the action part of the rule, a node of type

VP is created on CLIST and the v-l node is attached to the VP node.

The action part of a rule consists of the list of actions. An action is a command to
the rule interpreter about manipulating the content of the SWM, setting or resetting the
global variables, sending a request to the semantic processing component, etc. Some

actions are shown in Table 4.1.

 

Table 4.1. Some Actions

(create vp—l) .' create andput a new node in Clist.

(attach v-1 to vp-l) :attachatree toanode.

(transfer—features v-l to vp-l) ,' copy thefeatures.

(if (have-feature passive in aux-1) . . .) .‘test afeature in a node.
(add-feature dummy to np-Z) ; addafeature to anode.

(sem connect np-l v-l (dative-prep v-1) 2) ,' semantic suggestion to SEM.
(push-to-cash-lastnode) ; move the rightmost node in Clist to Cash.
(change-link np-l to iobj) :changeanameofa link.
(remove-feature modifiable from np-l) ,' delete a feature from anode.
(set-whcomp np—2) .' set a wh-phrase pointer to an NP.

(make-pointer np-3 to np-l) ,' makeanodepoint to another node.

 

Actions use the node names such as N P1 or VPl identiﬁed during the pattern matching
process of the pattern part of the rule. Because more than one node of the same category
can appear in the pattern part, a number sufﬁx is used for identiﬁcation (example, NP—l,
NP-2, etc.). For example, the pointer corresponding to NP-l found during the pattern
matching is used during the execution of actions when NP-l is speciﬁed in an action of
the same rule. Some complexity arises because more than one rule can match during the
pattern matching and the actions of those rules can change the SWM in a different way.
Therefore the data structures of the branch are copied into each new branch created for
each rule found during the pattern matching. Then the actions of a rule are applied on

' the SWM of the rule’s branch.

58

Now, it is necessary to explain how each rule’s pattern part is matched against the
trees in CLIST. Note that one pattern is matched against one tree. The mapping of the

patterns to the trees is illustrated in Figure 4.7.

noon

3 5 5

CLIST

 

 

 

 

 

 

 

 

 

 

 

 

 

 

[p4] [p3] [p2] [p1]
Pattern-part
I p3] [P2] [P1]
Pattern-part '
[p1]
Pattern-part

Figure 4.7 Mapping Patterns to Trees

The rightmost pattern should be compared with the rightmost tree and the next rightmost
pattern with the next rightmost tree, and so forth. This mapping implies that a rule
always specifies the righthand part of CLIST. The rightmost tree is the starting point of
the matching process. This pattern matching su'ategy is different from that of Parsifal. In

Parsifal, the ﬁrst buffer cell is the starting point as illusuated in Figure 4.8.

[l [l

[l [ g]
stack [ [l ] [I]

l ill

top celll 2 3
Figure 4.8 Parsifal’s Pattern Matching

 

 

 

 

 

 

 

 

 

59

4.3 Delay Used for Increasing Syntactic Context

Because of the pattern matching scheme explained in the above section, it is possi-
ble for SYNSEM to delay the construction of structures. The lookahead capability in
Parsifal can be achieved in SYNSEM by delaying the syntactic decision and receiving
further material into CLIST until a sufﬁcient amount of context is collected in CLIST.
Figure 4.9 is for illustrating how delaying can be used for the construction of a PP. Con-

sider sentence fragment (4-1) along with the ﬁgure.

an“): A
\ 7’

Rule Build-PP: [Prep] [N P] => ([create new PP]
[attach Prep to PP]
[attach NP to PP])

 

Rule PP-create in Packet Cpool
[Prep][NP]=>([....]

Figure 4.9 Delay in the build-PP Rule

(4-1) 1 in2 the3 basket4

In the case of Parsifal, the PP for "in the basket" is constructed while the processing
point remains at position 1 by looking ahead to the preposition and the NP "the basket".
To prepare the NP "the basket" which the normal processing point has not yet reached,
the attention shifting mechanism is used. While the normal processing point remains at
position 1, Parsifal moves its attention to position 2 and it processes the input starting
from this position. After completing the construction of the NP, the parser returns to
position 1. At this time, the input looks like "in [NP]". Now the PP-creation rule matches
this input. The working of Parsifal can be understood in a way that the parsing process

forks a child process and the child process analyzes and prepares an NP that lies ahead

(this is the attention shifting mechanism).

But SYNSEM does not use the complex processing mechanism such as the atten-
tion shifting mechanism. When the parser’s processing point is at 2 in (4-1), the parser
gets to the blocked situation because no rule matches CLIST in which the preposition
"in" is the righunost tree and no part of the NP has yet appeared. This is not the dead end

situation but the read-again situation. This situation is shown in Figure 4.10(a).

 

 

 

pr p Prep Det Prep Det N Prep )K /’K
D t p NP
in in the in the basket in the basket
(20 (b) (c) (d) (6)
Figure 4.10 Steps of Building a PP

 

 

 

Thus SYNSEM reads the next word "the" to provide more context, and the result is
shown in (b) in the ﬁgure. But CLIST in (b) still does not have enough context to ﬁnd a
rule that matches it. Thus, the next word, "basket", is input and CLIST becomes (c).
Note that the processing point is at position 4 instead of staying at position 1. Against
CLIST in (c), the NP construction rule matches and it builds the NP as shown in Figure
4.10(d). (This account is a simplied version. In the actual grammar of SYN SEM, it is
more complicated than this. Actually several rules are used to build the NP.) Still the
processing point (i.e. the rightmost node) is at position 4. Now the rule build-pp is found
during the pattern matching and creates the PP as shown in (e). Note that the PP has
been built using only the uniform processing mechanism. SYNSEM did not use some
special mechanism such as the attention shifting mechanism. Instead of looking ahead
and shifting the attention, SYNSEM delayed the decision of creating the PP (i.e. apply-
ing the build-pp rule) until CLIST gets the sufﬁcient amount of context. "Delaying the
decision" is implicitly implemented in "reading the next word if no rule matches

‘ CLIST".

61

A complex NP such as that shown in (4-2) gives more of a problem to Parsifal.
(4-2) r in 2 the 3 basket 4 which 5 the man 6 brought 7 from a the 9 store...

In Parsifal, while the main processing point remains at position 1, the sub-processing
begins at 2 to prepare the NP headed by "the basket". But the relative clause is a part of
this NP and it contains several NP’s such as "the man" and "the store". Thus, inside of
the sub-processing (i.e. the attention shifting), another sub-processing should be embed-
ded. This results in the recursive embedding of the processing levels as shown in Figure

4.11 because another attention shifting should be done inside an attention shifting.

 

 

main prncrssing
Moccssing-l
suhpmcessingL
in the basket which the man

 

 

 

 

 

 

Figure 4.11 Recursive Embedding of Attention Shifting

This results in undesirable complexity in the processing. It seems intuitive that this is not
the processing strategy that is used in the human language processing system. To avoid
this problem, the "node reactivation" mechanism is used in the grammar of Parsifal. As
soon as "the basket" is analyzed, the parser returns to the main processing point. Thus
when the relative clause is being analyzed in (4-2), there is no embedded attention shift-
ing. As soon as "which" is seen, Parsifal uses the node reactivation rule(sirnilar to the
attention shifting rule) to resume the processing of the NP "the basket". Note that the NP
may have already been attached to a previous tree at this point. This technique solves the
problem of deep recursion of the attention shifting, but it is directly against the "limited
looking ahead" theory in Parsifal. The "lookahead" theory of parsing can achieve its full

power only when the whole NP can be prepared by one attention shifting. (43) can be

62

used to illustrate this point:
(4—3) It is natural for the man who stole the apple to be punished.

After "the man" is analyzed as an NP, the buffer would be like:
[Prep;for] [NP; the man] [who]

The "PP-create" rule in Parsifal will match against the buffer cells shown above. But the
"For-np-to" rule in Parsifal is the rule that actually needs to match instead of the PP-

create rule.

SYNSEM parses sentence (4-2) using the uniform processing strategy. As soon as
"the basket" is analyzed as an NP, the rule build-pp matches CLIST. As soon as the
head noun of the NP is found, the partial NP can be regarded as a complete NP. In other
words, the partial NP corresponding to "the basket" is treated as a legitimate NP and can
be used by other rules. This is an early attachment strategy. This idea is similar to the
node reactivation rule in Parsifal but the account in SYNSEM is more easily conceived
and implemented. When SYNSEM reads "which", the rule which initiates a relative
clause is found and it attaches the secondary clause node "S" to the NP, "the basket"
(here the NP has already been attached to the PP node). Let’s call this rule wh-relative.

This rule is shown in Figure 4.12.

 

relplo

liP which
Rule Wh-relatiiC:
<low-right-NP> [which;relpro] =>
Figure 4.12 Matching Lowest and Righunost NP
In the rule wh-relative, the base pattern is <low-right-np> (low-right-np is a special case
of node-chain), which indicates that the lowest and rightmost node of the corresponding

u'ee (i.e. the second rightmost tree in CLIST) should be an NP. After the execution of

63

this rule, the snapshot of CLIST will be Figure 4.13.

x/9 [$60]

, , relpro

 

which

 

Figure 4.13 Implicit Attachment of Secondary Clause

The secondary clause node, "S", is implicitly attached to the NP which is attached to the
PP node. When the NP for "the man" is built, it will be attached to the S node as a sub-
ject. As the following input is received, the secondary clause node, S, is built incre-
mentally as if it is the main clause. The secondary 8 node is implicitly attached to be
seen as the root of a tree (indicated by the dotted line) so that the rules for the main
clause processing can be used for the processing of the secondary clause. When the
secondary clause S is complete, this tree will just be popped out (but it will still be
linked to the NP node). As will be shown in the examples later, the PP can be attached

to the previous tree before the processing of its relative clause begins.

But SYNSEM’s processing strategy is not optimal. For the sentence (4-3), SYN-
SEM can not use a rule whose pattern part is of the form "[Prep;for] [NP] [Prep;to]" (let
us call this the For-np-to rule) because of the following reason: as soon as the NP for
"the man" is constructed the build-pp rule will match the SWM and create a PP. One
solution is to add some restriction to the build-pp rule so that it can not match if the
preposition is "for". Another solution is just to add another rule whose pattern part is
"[PP;prep=for][Prep;to]" that corresponds to the for-NP-to rule. Neither solution is the
best.

4.4 Rules and State of Parsing

In Parsifal, the grammar rules are partitioned into packets. Each rule is a member
of only one packet. SYN SEM has no packets. It has only one pool of all rules. This
design strategy requires some explanation. The reason that Parsifal use packets is two-
fold. The ﬁrst is efﬁciency and the second is conciseness of the grammar. Each node in
the stack of Parsifal is associated with a list of packets. The node at the top of the stack
(the most recently pushed node) is called the current active node. The packets associated
with the current active node are the active packets. Adding(deleting) a packet to(from)

the set of active packets is requested in the action part of a rule.

When a node is pushed into the top of the stack, the packets related to the previous
active node are not active any more. If a node is popped from the top of the stack, the
second node from the top of the stack becomes the current active node and the packets
that are associated with this node become active. The whole idea of having the active
packets for the active node is that only a small subset of the rule base is related to the
parsing at each point (represented by the current active node). The other rules are
irrelevant to the parsing and need not be accessed. For example, a small number of rules
are related to the parsing of the subject of a clause. These rules are contained in the
packet called PARSE-SUBJ. The efﬁciency of parsing is obtained by looking at only the
rules in the packet PARSE-SUBJ during the time when the subject is analyzed. Now it is
likely that the set of active packets represents the state of parsing (see (Marcus, 1980)
for more detailed account). The rules in a packet will be formulated under the assump-
tion that they will be applied only during the state of parsing represented by the packet.
Thus the rules need not test the state of parsing in their pattern part, which makes the

grammar rules concise.

In SYNSEM, the advantage of using packets can not be used because of its parsing
strategy that "any rules that match CLIST should be found. " Consider the next sentence

to see the reason:

65

(4-4) The granite rocks near the shore.

Note that "rocks" can be both a noun and a verb. After "rocks" is input, two rules should
be found during the pattern matching: a rule that analyze "rocks" as the head noun of the
NP and another rule that analyze "rocks" as the main verb (actually there is one more
rule). But these two rules can not be put in the same packet because they are related to
different parsing stages, which makes SYNSEM unable to use packets. One of the
disadvantages of using packets is that there should be some rules(or actions) that add or
delete packets to and from the set of active packets. This causes some difﬁculty to the
grammar writer. The management of the active packets gives difﬁculty to both the
designer and the reader of the grammar. Charniak(1983b) attempted to remove this
difﬁculty without much success. This difﬁculty can be totally eliminated by removing
the packets. Our experience is that it is hard to keep track of the active packets during

the reading of Parsifal’s grammar rules.

Instead of using the active packets to get the state information of parsing, SYN-
SEM uses rules with the base pattern to check the state of parsing. It has been found
that the base pattern can be speciﬁed in a simple way. There are many rules that do not
even need the base pattern. It is argued that rules are still concise even with the base pat-
tern. We need to consider the possible inefﬁciency caused by not adopting the packets.
During the pattern matching, all rules in the rule base should be checked to ﬁnd the rules
that match CLIST. But this computational cost can be minimized by using a special rule
indexing technique. It has been veriﬁed in the experiment that the time taken for the
pattern matching is a small fraction of the total parsing time. The bottleneck of the

speed of SYNSEM is in the semantic processing instead of syntactic processing.

The raw patterns in the rules of SYNSEM are analogous to the patterns for check-
in g the buffer cells in Parsifal. Marcus argued that the lookahead of three buffer cells is
enough for deterministic parsing. He used this lookahead of only up to three buffer cells

to explain why the garden-path effect occurs in sentence (4-5). But this argument has

66

been challenged in several ways. Refer to (Kurtzman, 1985).
(4-5)The horse raced past the barn fell.

Following Parsifal, SYNSEM uses up to three raw patterns to make a rule. It has been

found that three raw patterns are enough to write the grammar.

Marcus explained that some linguistic generalizations such as Ross’s complex NP
constraint can be obtained automatically by using Parsifal’s grammatical constraint that
a rule can not access the inside of the nodes in the active node stack other than the
current active node and the ﬁrst S node. In SYNSEM, such grammatical constraint does
not exist. The base pattern should be able to access the inside of the corresponding tree.
Hirst(l984:211) provides the following counter example to Marcus’ claim. In the
prepositional attachment problem, the verb of the VP should be examined to determine
the attachment point (either the VP or the NP appearing after the main verb). The NP is
the current active node (i.e. the node at the bottom of the stack) and thus it can be exam-
ined during the processing. But the VP is the node which is the node just above the NP

in the stack and thus it can not be looked at according to Parsifal’s constraint.

Another counter example can be found in the parsing of conjunctions. Consider

the following sentences with conjunctions in them:
(4-6) The principal ordered the pupil to be punished and scolded.

(4-7)The teacher ordered the student to be punished and taught the other students

not to commit a similar misdemeanor.

To determine the counterpart of the conjunction of "scolded" and "taught" in the above
sentences, the verb "ordered" should be considered. But according to the constraint of
Parsifal, the verb "ordered" should not be available for consideration because it is nei-

ther the current active node nor the ﬁrst 8 node above the current active node.

,67

4.5 Rules and Ambiguity

In SYNSEM, the existence of a local syntactic ambiguity at a certain point is exhi-
bited by the fact that multiple rules are found during the pattern matching. One of the
important guidelines to writing the grammar rules in SYN SEM is to prepare a rule for
each possible analysis of the ambiguity. Multiple rules found by the pattern matcher
invoke the same number of parsing branches. As the parsing goes on, some branches
will die out because of either syntactic or semantic reasons. Later there will remain only

one branch, which means that the ambiguity has been resolved. In sentence (4-8),
(4-8) The boy ate a cake with a fork.

there is an ambiguity about the attachment of the PP "with a for " as shown in Figure

/\ /5\
ﬁ /vl\ & /K
the boy /v ‘2’); /PP\ the boy IV A

4.14.

 

a cake
ate 174’ 5 ate
2K )"\
With a fork .
a cake Wlth
a fork

Figure 4.14 PP Attachment Ambiguity

The snapshot of CLIST just after the analysis of the PP "with a fork" is shown in Figure
4.15. There are two rules(np-pp and vp-pp) that match this CLIST (these two rules are
also shown in Figure 4.15):

Note that the NP "a cake" (possibly a partial NP) has been already attached to the VP
because the head noun has been already analyzed and "a cake" is considered to be a legi-
timate NP. As explained before, the base pattern <low-right-np> of the rule np-pp

matches the NP "a cake" because this NP is the lowest and rightmost node of the second

68

(np-pp ((l )(h passive-by))
[ ( [ (np-l low—right) 1 (modifiable) l; 1 indicates base pattern.
[ (pp-1) 0 l) ] ) ; 0 indicates raw pattern.
() ] ; no additional restrictions.
[ (attach pp-l to np-l)
(if (have-feature trace in pp-l)
(sem connect (np pp-l) np-l (prep pp-l) 2)
(sem connect np-l (np pp-l) (prep pp-1)) )
(remove-feature modifiable from np-l) ]

(vp-pp ((1 )(h wh-dative-np-pp))
[ ( [ (vp-l 3-1) 1 () l ; base pattern.
[ (PP‘I) 0 () l ) ; raw pattern.
0 l
[ (attach pp-l to vp-l)
(if (have-feature trace in pp-l)
(sem connect (np pp-l) (v vp-l) (prep pp-l) 2)
(sem connect (v vp—l) (np pp-l) (prep pp-1)) )

 

CLIST
AK /m\2i
P / p
the% Y & :iTth a fork

ate a cake

 

Figure 4.15 PP Attachment and the Corresponding Rules

rightmost tree in CLIST. The base pattern of the rule vp-pp checks if the second right-
most tree of CLIST has the root node of type S and its rightmost child node is of type
VP (i.e. the verb of the clause has been analyzed). This is to check the state of parsing if
there is a VP which can be modiﬁed by the PP that is the rightmost tree in CLIST. For
each of np-pp and vp-pp, a branch is created. The actions in the rules are executed and
some of them may invoke semantic processing. The result of semantic processing of the
rules is used for comparing the branches and the better one is chosen (which will be vp-

pp in this case).

69

The sentence (4-9) shows another kind of ambiguity. After the word "raced" has
been input, the SWM becomes that shown in Figure 4.16. Two rules should be found

during the pattern matching.

(4-9) The horse raced past the barn fell.

 

 

(‘I .IST
A V
the horse raced

 

Figure 4.16 Verb Form Ambiguity
The rule main-verb2 analyzes "raced" as the past main verb of the sentence and the rule
np-vpp considers "raced" as the past participle leading a reduced relative clause.
Ambiguity also occurs when a word is categorically ambiguous as shown in (4-10).
(4-10)The granite rocks near the shore. (= (4-4))

The word, "rocks", can be both a noun and a verb. The snapshot of CLIST and the rules
that match this CLIST are shown in Figure 4.17:

 

 

(‘1 .IST
/"’\ I
the granite rocks

 

Figure 4.17 CLIST and Categorical Ambiguity

Three rules are found in this example: the noun-parseS rule which interprets "rocks" as
a main verb, the noun-parse4 rule which considers "rocks" as the head noun of the NP
and the noun-parseS rule which closes the NP "granite" and considers "rocks" as a
noun that begins a new NP. The disambiguation will be done later by either semantic
information or syntactic reasons. This example will be explained in more detail in
Chapter 6. These examples show that SYNSEM prepares one rule for each alternative

of an ambiguity.

70

4.5.1 Delay and Parallelism

The fact that a rule can have up to three raw patterns indicates that the parser can
delay the decision up to three constituents (i.e. at least three words) to collect the
sufﬁcient context. But it seems that the human sentence processing system does not
always utilize the delay as fully as possible. Let’s consider rules for parsing (4-11) and
(4-12).

(4-11)I know the man.

(4-12)I know the man is a doctor.

After "the man" is analyzed as the NP, the rule object in (4-13) is used to analyze this

NP as the object of the verb "know". This rule is shown below:

(4-13) Rule Object:
<S-VP—V> [NP] => [attach NP to VP]

(4—14) Rule Reduced-that-comp
<S-VP-V; that-comp> [NP] [V] =>...

Note that V is speciﬁed in the node chain of the base pattern to make sure that there is
nothing attached to the VP except the verb. The problem of the object rule is that it will
analyze "the man" as the object of "know" in (4-12), too. This is the incorrect analysis.
In (4-12), the reduced-that-comp rule in (4-14) should be used to analyze the comple-
ment clause that is led by the NP, "the man". However this rule can match CLIST after a
verb(or auxiliary verb) is input to CLIST because of the pattern [v] in the rule. Before
this rule matches, the object rule will match CLIST. Then the parser will reach the
dead-end and the parsing will fail.

The solution for the above problem is to change the object rule so that it can not be
applied for the verbs that can have a "that" complement clause and to add another object
rule (called special-object shown below) that will wait and see the word after the NP.

The updated rules for this purpose are shown in (4-15) and (4-16):

71

(4—15) Rule Object:
<S—VP—V; not(that-comp)> [NP]
=> [attach NP to VP]

(4-16) Rule Special-object:
<S-VP-V; that-comp> [NP] [#V]
=> [attach NP to VP]

The object rule in (4-15) can not be used for (4-11) or (4-12) because the V, "know",
has the feature "that-comp" indicating that it can have a complement clause with com-
plementizer "that" (note that the "that" complementizer in English can be omitted). For
(4-11), the special-object rule in (4-16) will be used after "." is input, because "." will
match the pattern [¢V]. Thus both (4-11) and (4-12) can be successfully analyzed using

these two rules.
But these rules are not perfect. The sentence (4-17) causes a problem.
(4-17)I know the man who eats an apple is a doctor.

Just after "who" is input into CLIST, the special-object rule matches the CLIST and
"the man" is analyzed as the object of the verb "know". But this is a wrong analysis
because "the man" should be analyzed as the subject of the complement clause. This
decision should be delayed until the correct evidence occurs. To ﬁx this problem, the

special-object rule can be modiﬁed as follows:

Rule special-object:
<S-VP-V; that-comp> [NP] [.1 => ....

This solution works for the sentence (4-17), but there is a counter example such as (4-
18):
(4-18)I have known the man since 1930.

The special-object rule can not be used to analyze this sentence because "since 1930"

exists between "the man" and ".".

72

From the discussion done so far, we can notice that even delaying the decision for
three words is not enough. It is argued in Frazier and Raynor(1982) that "the man" in
(4-12) is ﬁrst analyzed as the direct object of the verb and then this analysis is thrown
away if it proves to be wrong later and the sentence is reanalyzed. They presented data
from eye movement experiments as evidence that (412) shows some garden path effect.

The position taken in SYNSEM is to use the following two rules:

(4-19) Rule Object:
<S—VP-V> [NP] => [attach NP to VP]

(4-20) Rule reduced-that-comp:

<S-VP-V; that-comp> [NP] => ..open sec clause..

The above two rules match CLIST (just after "the man" is analyzed as the NP) and the
corresponding two branches run in parallel. For the sentence (4-11), the branch created
by the rule reduced-that-comp will be removed after the input of "." because the
expected complement clause does not follow. The sentence is successfully analyzed by
another branch created by the rule object. In the case of (4-12), the input "is" will make
the parser remove the branch corresponding to the object rule. The branch issued by the
rule reduced-that-comp will analyze this sentence successfully. For the sentence (4-
17), the branch issued by the reduced-that-comp will carry out the parsing but the
branch for object will die in the middle of parsing. This parsing strategy indicates that
the partial parallelism concept introduced in Chapter 3 overcomes this problem. But

note that this is just an alternative to the reanalysis strategy.

4.6 Syntactic Processing in Operation

To understand the operation of the syntactic analysis clearly, a detailed trace for
the analysis of (4-21) will be given in this section (See Appendix D for the system out-

put and see Appendix F. for the rules mentioned in the thesis).

73

(4-21)The teachers taught by the pool passed the test.

The parsing of a sentence starts with the empty CLIST and CASH. Just before reading
the ﬁrst word, a dummy node of type "SS" is inserted into the empty CLIST (the
snapshot of the SWM is shown in Figure 4.18). "SS" node indicates

 

 

 

 

CLIST CLIST
ss 33 (£3
CASH— ’ 9115.151—
(a) (b)

Figure 4.18 Snapshots of SWM

that it is the starting point of a sentence. MON reads the word "the" and calls the mor-
phological routine. This routine looks in the dictionary and ﬁnds the part of speech of
the word. If the word is an open class word, an instance concept for it is created in the
knowledge base and the MP passes a marker starting from this instance concept. Marker
passing will be explained in the next chapter. This word is inserted into CLIST of the
branches that are active (only one branch here). The SWM becomes as in Figure 4.18(b).
For this CLIST, the pattern matcher ﬁnds only one rule parse-det (Refer to Appendix A
for the NP structure used in SYNSEM.) After executing this rule, The SWM becomes
(c). No rules are found for this SWM, which means that the blocked situation has
occurred. However it is determined that this is not the dead end (Refer to Appendix B.)
Thus MON reads the next word "teachers" and inserts it into CLIST as shown in (d).
The pattern matcher ﬁnds only one rule, noun-parsel , for this SWM and its execution
results in (e). For (e), the pattern matcher ﬁnds the rule noun-parse33 and the SWM
becomes (f). This rule analyzes the noun "teachers" as the head noun of the NP that is
being built, because the noun is in plural form. Refer to Milne(1980) for the use of
singular/plural form of a noun to decide the boundary of the NP.

74

 

 

 

 

 

 

(‘1 181‘ CLIST CLIST
88 N S 83 NS N 38 bltc N
l I N5 |
Det Det teachers I teachers
(the) liner net
(C) (d) (6)

Figure 4.18 Snapshots of SWM(Continued)

Only one rule, subject, matches (f) which changes CLIST to (g).

 

 

 

SS ’NP\ /S\ A [main,en]
NC N SS P
SS NP
NA ' A 4;
NS teachers taught

——-Det

 

 

 

(f) (g) (h)
Figure 4.18 Snapshots of SWM(continued)
The NP has been analyzed as the subject of the sentence. No rule matches (g) and the
next word "taught" is read in by MON. Because this is an open word, the MP is called to
pass markers from the instance concept for this word. The Snode for this word has
features "main" and "en" because "taught" can be a main verb and a past participle. For
CLIST in (h), the pattern matcher ﬁnds three rules: main-verb2, np-vpp and np—vppl.
The main-verbZ rule analyzes "taught" as the main verb of the sentence. The other two
rules are for the reduced relative clause analysis of "taught". Note that this verb is a
dative verb as shown in "She taught the boy mathematics." Thus "the teachers" in (4-
21) can be either a direct or indirect object of "taught". For example, "the theory" in (4-
22) is the direct object but "the person" in (4-23) is the indirect object of the verb
"taught".
(4—22)The theory taught by the professor is really hard to understand.

' (4—23)The person taught by the professor is a farmer.

75

The np-vpp rule analyzes the NP, "the teachers", as the direct object and the np-
vppl rule analyzes the NP as the indirect object of the verb "taught". The pattern parts
of the two rules are almost the same. The difference comes from the action part. The
action part of np-vpp has an action which suggests that the NP and the V should be
related via the pseudo-preposition "obj". This indicates that the NP is the direct object of
the V. But an action in the action part of the rule np-vpp1 suggests that the NP and the
V should be related via the preposition "to" (which is the dative preposition of "teach").

This means that the NP is the indirect object of the V.

Each of the above three rules create a branch as shown in Figure 4.19.

 

 

 

 

 

 

 

 

 

 

 

 

 

main-verb2 > (I)
nP-vpp , 0)
np-vppl : (k)

 

 

 

Figure 4.19 Creation of Multiple Branches

The SWM’s in (i), (j) and (k) shown in Figure 4.18 are the results of the execution of the

rules of the three branches.

 

 

/I\ S/‘\ I~7’\ 7\ 17K
NPSS/VP latdumm mmy] fad SS P [dummy] 3%

V V NP P
)(taught) / taughtv) )1
V
(i) (k)
Figure 4.18 Snapshots of SWM(Continued)

 

In (j) and (k), the secondary clause 82 is attached to the NP. But 82 is a root node of the
rightmost tree because this is an implicit attachment as indicated with the dotted line.

The reason that 82 should be a root node is that this secondary clause is treated as if it is

76

the main clause so that the rules for building the main clause can be used to build the
secondary clause. When 82 is complete, it will just be popped out from CLIST (but it
will still be attached to the NP). An attachment like this is called an implicit attachment
because it is done in the background. 82 actually hides 81 until its construction is

ﬁnished and it is popped out.

For each branch, no rule is found during the pattern matching and a new word is
requested to be input. As explained in Chapter 3, the semantic comparison stage comes
at this point. The branch for the rule main-verb2 is selected after the comparison of the
semantic processing results (The reason for this selection will be explained in Chapter 6,
but it has to do with the knowledge that teachers do teach). The other two branches are
pushed into the backup stack. Out of these two branches, the one with the better seman-
tic result is pushed later so that it can be used ﬁrst when backtracking is needed later.
The output of this semantic comparison stage is one branch. Now MON inputs the new

word, "by", and the SWM for the selected branch becomes (1).

 

)l\ p 1 6P t 1 P NC T! 1 Prep NC N
38 NP ”(A
)'P pool
V (W) (by) NS passed
(the) (by) D“

(1) (m) (n) (0)
Figure 4.18 Snapshots of SWM(Continued)

 

No rule matches (1) and MON reads the next word "the" as shown in (m). After several
NP processing rules are used and the word "pool" is input, the SWM becomes (n) (this
change is same as the change from (b) to (e) before). At this point, it can not be deter-
mined that "pool" is the head noun of the NP without looking at the next words. For
example, "pool" is the head noun of the NP in "the pool of the house", but it is not the
head noun in the nominal compound, "the pool repair". No rule matches the SWM in

(n). Thus MON reads the next word "passed" and the SWM becomes (0). The rule

77

noun-parSGS matches (0). Seeing the V following the noun "pool", this rule decides
that "pool" is the head noun of the NP being built. Note the last action of this rule is
"(push-to—cash-last-node)" which pops the last node, V(passed), and puts it into CASH.
The reason for performing this action is that the V has been used by the rule only for
getting enough context to make the decision if "pool" is the head noun or not. After
completing this purpose, the V should be retracted from CLIST so that the NP just built
can be the rightmost node of CLIST and all further processing can be done for this situa-
tion.

After the execution of noun-parseS, the SWM becomes (p). For this new SWM,
the pattern matcher ﬁnds the rule build-pp which builds a PP as shown in (q).

 

ﬂPreP NP R e ik R IV R178

 

 

 

NC N NPYPPrcp Np NP ﬂ NP ﬁpassw SS PNP )1
v V ill V PL V P

CASH— EASE— NP NP

v IV Pm.“ Prey

ed
(ML ML CASVH— CASH=nil C A SH=nil
(passed)
(P) (q) (r) (S) (t)

Figure 4.18 Snapshots of SWM(Continued)

The vp-pp rule matches the SWM in (q) and this rule attaches the PP to the VP as
shown in (r). (The semantic component returns the signal of accepting this attachment
because a "teaching" action can happen at the location close to the pool.) No rule is
found for (r) and a new input is required for CLIST. Because CASH is not empty, MON
pushes the node, V(passed), in CASH into CLIST instead of reading a new word from
the input suing. The new SWM is shown in (s). The V of "passed" can be a past main
verb or a past participle. The main verb analysis rule does not match the SWM because
the main verb is already in the tree. The only rule that matches is np-vpp which

analyzes the V(passed) as the past participle that begins the reduced relative clause.

78

Let’s assume that "pass" has two meanings: (1) to hand something to a nearby person,
(2) to have enough qualiﬁcation in a test. The execution of the np-vpp rule gets the
response from the semantic component that the result is semantically bad. The reason
for this response is that a pool can not be passed because it is not a movable object and a
pool is not a kind of test. The branch whose execution of the corresponding rule results
in a bad semantic effect is just removed. This is another case of disambiguation due to

the semantic reason.

Note that there has been only one rule that is active and the execution of this rule
has been rejected by the semantic component. Thus, the parsing is in a dead end because
there is no active branch left. Therefore it is necessary to back up. The parser pops out
an image of a branch that was stored in the backup stack and makes a new branch which
is shown in (t). Note that the SWM of this branch is actually the SWM in Figure 4.18(k)
that was pushed into the backup stack. This branch is the branch that was created
corresponding to the rule np-vpp1. The next word to be read for the branch correspond-
ing to the SWM in (k) is "by".

No rule matches (t) and MON reads the next word "by". The new SWM is shown
in (u). The course of parsing from (u) to (v) is not explained because it is related to the

PP analysis which was explained from (1) to (q). Against the SWM in (v),

CLIST CLIST CLIST CLIST

”77‘? f"? 33

 

SS NP1???by

 

ss NP1 NP2 P29\)( PP ed
ass
P2v NP3by v NP3 VNP3 p
CASH=nil CASH=(V) CASH=(V) CASH=nil
(ll) (V) (w) (x)

Figure 4.18 Snapshots of SWM(Continued)

the pattern matcher ﬁnds two rules: the vp-pp rule and the passive-by rule. passive-by
analyzes the NP, "the pool", of the PP as the actual subject of the verb "taught" (of the

passive clause) and the vp-pp rule tries to attach the PP, "by the pool" to "taught". Thus,

79

two branches are created. The execution of the rule passive-by gets the failure signal
from the semantic component because a pool can not be the agent of the action "teach".
Thus the branch for this rule is removed. The execution of the rule vp-pp is accepted
because a satisfactory signal is returned from the semantic component for the execution
of the rule. Thus only one branch remains as active. The new SWM for this branch is
shown in (w). No rule matches (w) and the node V in CASH is pushed into CLIST as

shown in (x).

Only one rule(np-pp) matches (x) which attempts to analyze the V(passed) as the
past participle that begins the reduced relative clause that is attached to the NP, "the
pool". But this analysis is rejected by the semantic component as explained before.
Therefore the parsing is blocked and needs to back up. But before doing back up, the
parser always checks CLIST to see if there is an S of the secondary clause and if it is in
a complete form. If so, this 8 is popped out (this is the strategy of closing the secondary
clause in SYNSEM). Even if 82 is popped out, it is still attached to NP1. The SWM for
this branch becomes as shown in (y). Against (y), the main-verbZ rule matches. This

rule

 

as, I 2x A ”r a 2%
)K passed NP\ /vr> 2': TP the Xi TP 2:;

 

 

N” /‘[K i)" V
V NP3)K I CASH=nil CASH=(-)
Prep NP

 

passed
CASH=nil CASH=nil
(y) (z) (21) (22)

Figure 4.18 Snapshots of SWM(Continued)

analyzes the NP for "the teachers taught by the pool" as the subject of the verb "passed".

The result of the execution of this rule is shown in (2). No rule matches the SWM in (z)

80

and MON inputs the next word "the" as shown in (21). After the NP for "the test" has
been constructed, the SWM becomes (22). Note that CASH contains "." which was used

to provide the context for analyzing the NP.

The object rule matches (z2) and its execution results in the SWM in (23). No rule
matches (23) and "." in CASH is pushed into CLIST as shown in (24). The s-close-aff

rule matches and this rule

CLIST CLIST CLIST

pr>vK NP/s\ OI N

/V“’\ I.
lﬁ (2:; A A

 

 

 

passed the passed
test test
CASH=(.) CASH=nil CASH=nil
(23) (z4) (25)

Figure 4.18 Snapshots of SWM(Continued)

attaches the punctuation mark to the tree as shown in (25). No input is left and the ﬁnal
punctuation mark is attached to the tree. Therefore the parse of the sentence (4-14) is

ﬁnished.

4.7 Conclusion

The syntactic analysis component of the SYNSEM parser has been explained in
this chapter. The data structure employed in the parser is a list containing trees. The
grammar rules have the form of fairly straightforward production rules. The main design
objective of this component is to allow the parser to be able to delay the syntactic deci-
sion until a sufﬁcient amount of context becomes available. This "delaying" is realized
by reading the next word when no rule is found during the pattern matching and the
situation is not the dead-end. By reading the next input word, the SWM will contain

more context and rules that can match this added context are sought again.

81

The method of delaying the decision achieves the same power as the lookahead
used in Parsifal. The use of delaying enables the SYNSEM parser not to use the atten-
tion shifting mechanism which psychologically seems to require processing which is too
complex. SYNSEM can use a simple and uniform processing technique even for the

structures that require attention shifting in Parsifal.

Another feature of SYN is the base pattern that is used to ﬁnd the state of parsing in
each rule if necessary. Instead of employing states in ATN or packets in Parsifal, each
rule can test the state of the parsing by specifying the base pattern. This might make the
rules look a little complex but actually it makes the rules more readable and makes it

easier to follow the grammar.

The rules that have corresponding semantic processing contain actions which issue
a suggestion to the semantic processing component to ﬁnd a path in the knowledge base.
The path found for the suggestions are used as the measure of the goodness of the

rule(branch).

One important guideline to prepare rules in SYNSEM is to prepare one rule for
each alternative analysis of a structural ambiguity. If there are n possible ways of
analysis at some processing point, then It rules should be proposed by the pattern
matcher. Each alternative will become a branch, which means that multiple branches of
parsing can go on in parallel as explained in the previous chapter. Some branches will
drop out as the parsing is going on because of either a syntactic reason or a semantic rea-

SOIL

CHAPTER 5

KNOWLEDGE BASE AND MARKER PASSING

The major motivation of the SYNSEM parser’s control strategy is to use the
semantic information as much as possible as a resource for disambiguation. We use a
knowledge base to provide the parser with this semantic information. Because we are
building a general natural language understanding system, the knowledge representation
language should be one that can represent general world knowledge instead of one that
is suited for a speciﬁc domain. The enumeration of the possible candidates is ﬁrst order
logic, KL-One, KRL, Frail, Conceptual Dependency, KODIAK, and so forth. It seems
that these languages are the same as far as representational power is concerned, even
though each has its own representational characteristics. It is known that the ﬁrst order
logic is equivalent to the frame-based representation languages in the representational
power. First order logic has a straightforward inference engine which is the so-called
"resolution-based" theorem proving. The problem is that the amount of computation can

grow intractably as the size of the knowledge base grows.

The frame-based representation languages add some features to have more control
on intractability but end up with reduced capability for making inferences. In SYN SEM,
KODIAK is used as the knowledge representation language to build the knowledge base.
KODIAK was developed by Wilensky(l987). It has been used in some A.I. systems
such as FAUSTUS (the text inference system) (Norvig, 1986) and UC (natural language

processing system) (W ilensky & etc, 1986).

One of the motivations behind choosing KODIAK in SYN SEM is that it has a type
of concept called relation that relates the concepts of objects and actions. Another

motivation is that it is a semantic network in which it is appropriate to use the marker
82

83

passing method to facilitate the semantic processing. Charniak( 1983a) ﬁrst proposed the
use of marker passing in natural language processing systems. He pointed out that
marker passing-based systems can provide advantages for problems such as word sense
disambiguation, prepositional phrase attachment, and so forth. Norvig used the marker

passing mechanism to get plausible inferences in the text understanding system.

In this chapter, KODIAK and marker passing used in SYNSEM will be inuoduced,
which provides the reader the necessary background for the discussions in the following

chapters.

5.1 KODIAK

KODIAK will be introduced brieﬂy in this section. (The notations for KODIAK
elements used in this dissertation mostly follows the description of KODIAK in (Nor-
vig,l986).) KODIAK has three types of nodes: absolutes, relations, and aspectuals. The
nodes are connected via links. There are eight different types of links. Absolutes
represent the objects or entities in the world. Entities here include some concepts that
have a meaning by themselves. For example, objects in the world such as "boo ",

"train", "person", "dream , plan", etc. are represented by absolutes in KODIAK. Non-

object entities such as "action , eat , run", etc. are also represented by absolute nodes.
(The main intention of KODIAK is to derive the meaning of absolutes from that of rela-
tions and aspectuals.) A relation in KODIAK connects two absolutes. For each relation,
there are two corresponding aspectuals. The aspectuals for a relation can be considered
to be the formal parameters of the relation. The absolute which speciﬁes the class whose
elements can ﬁll the parameter is connected to the aspectual by a link. Let’s consider
Figure 5.1 as an example. The double circles are for relations, the single circles for
aspectuals, and rectangles for absolutes. The relation eater-eat relates the absolute

animal and the absolute eat. The absolute animal is connected to eater-eat

via the aspectual eater . eater-eat. (The names of the concepts can become long

84

 

 

 

 

 

 

 

 

 

 

 

 

eat food

   

eater.eater-eat

eateater-eat CﬂtﬁﬂtCC-Cﬂt eatee.eatee-eat

Figure 5.1 Elements of KODIAK

because each concept has a unique name so that a concept can be accessed by its name.
We tried to use self-explanatory names but the name itself does not have any meaning
for the Operation of the system.) We can consider the aspectuals, eater . eater—eat
and eat . eater-eat, as the parameters of the relation eater-eat. The absolute
animal constrains the argument which ﬁlls the parameter eater . eater-eat and
the argument of the parameter eat . eater-eat is constrained by the absolute eat.
As shown in Figure 5.1, the absolute animal is pointed by a C(constraint) link from
the aspectual eater . eater-eat. The "A" link between an aspectual and a relation
can be considered to imply argument. Anything that is a kind of animal has a relation
named eater-eat with any concept which is a kind of eat and can be connected to
(or involved in the relation with) the relation eater-eat which is again connected to
the concept eat. The fragment of knowledge represented in Figure 5.1 can be

represented in a frame-based language as shown in Figure 5.2.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

name eat
agent //\ animal
slots ,
patient -"‘—\
food

 

 

 

 

 

Figure 5.2 Frame Representation

85

Comparing Figure 5.1 with Figure 5.2, it can be noticed that slots in a frame-based
language correspond to the relation nodes in KODIAK. The agent slot of the frame
eat corresponds to the relation eater-e at in KODIAK. Relations are considered to
be the most important elements in KODIAK. Thus KODIAK is considered to be a
relation-based knowledge representation language. The name "aspectual" is derived
from the fact that the meaning of two aspectuals are derived from the relation that holds

between them.

KODIAK has eight types of links which are listed in Table 5.1 adapted from Nor-
vig(1986). The Prototype link is the newly added link which points to the concept which
can be the prototype ﬁller of an aspectual.

 

Table 5.1 Links in KODIAK
Dominate(D) - a concept is a subclass of another class
Instance(I) - a concept is an instance of some class
View(V) - a concept can be seen as another class
Constrain(C) - ﬁllers of an aspectual must be of some class
Prototype(P) - prototype ﬁller of an aspectual
Argument(A) - associates aspectuals with a relation
Fill(F) - an aspectual refers to some absolute
Equate(E) — two concepts are co-referential

Differ(Df) - two concepts are not co-referential

 

The D(dominate) link indicates the subclass relationship between two concepts. If a
node, say nodel, is a subclass of another node, say node2, then there is a D link which
starts from nodel and terminates at node2. D links are also used to specify the subclass
relationship between two relations. D links in KODIAK correspond to the IS-A or AKO

links in other representation languages. One important implication of D links is that the

86

properties of concepts are inherited by other concepts via D links. If nodel dominates
node2 (i.e. a D link points from node2 to nodel), all the relations and aspectuals that
nodel is involved with are inherited by node2.

The I(instance) link is used to represent the relationship between a member and the
class to which this member belongs. For example, to represent that a speciﬁc person
called Bill is a person, the I link is used between the absolute billO 1 and person
as shown Figure 5.3. Note that "01" in bil 10 1 is used to indicate the speciﬁc person,
our Bill and not, say, Bill Cosby:

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

person person
A
person-with-
_name-“ ill
BillOl
BillOl
(a) (b)

Figure 5.3 Instance Links

Because there can be many persons whose name is Bill, we can introduce a new concept
called persons-with-name-Bill as shown in (b) of the ﬁgure. The A and C links
have already been explained. For each relation, there are two A links that associate two
aspectuals to the relation. These two aspectuals correspond to the two arguments of the
relation. The C link is used to specify what kind of absolute can be connected to an
aspectual. This absolute can be considered to be the constraint of a concept which can
ﬁll the argument that the aspectual speciﬁes. An E(equate) link is used to connect two
absolutes to specify that the absolutes are co-referential. If an E link connects two aspec-
tuals, this means that the aspectuals are ﬁlled with the same absolute. Df(different) links

areiopposites of E links. They are used to assert that two absolutes are explicitly not co-

87

referential, or that two aspectuals can not be ﬁlled with the same absolute.

The V(view) link is used to regard an absolute as another absolute. For example,
"person" can be viewed as a "physical-object". In this case, "person" is used actually to
refer to the person’s body. The representation involving the V link is complex and will

not be explained in detail here.

5.1.1 Modeling Knowledge

To have a better feeling of KODIAK, it is useful to look at a piece of a knowledge
base written in KODIAK. Figure 5.4 models the knowledge involved with "giving" and
"having". Two absolutes named giver-has-given and givee-has-given are
two concepts which are kinds of the concept have. The absolute giver—has-
given represents the idea that one should have something to give to somebody. Note
that these ideas are represented by full-ﬂedged concept nodes in KODIAK. This shows
that KODIAK encourages the proliferation of concepts. giver-has-given is a
precondition of giving while givee-has-given is the result of giving, as
shown in the ﬁgure. Note that several E links are used to express the important relation-
ship among give, giver-has-given, and givee-has-given. The person
who gives something to somebody is the same person who has it. The thing that is given
by somebody is the same thing that the person has and the same thing that the recipient

will have. These are represented by the E links (indicated by "=") in the ﬁgure.

The D link between actor-act and giver-give indicates that the
giver-give relation is a kind of the actor-act relation. The concept have is
considered to be a kind of the concept experience which is again a kind of the con-
cept st at ive. The relation haver-have is subsumed by the relation
experiencer-experience. If giver-give had not been speciﬁed in the ﬁgure,
actor-act relation would be inherited by the absolute give because of the D link

between act and give. But giver-given was speciﬁed in the ﬁgure because

88

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

   

PantiCigant-
81 I . I [1 thing
. situatio ll event @
C O {f . C
/
person C
‘ person
stative
l
xpe 'encer- D
x ence

 

 

experietjce

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

D
have
D
mt
:4 ‘ I c

 

haver-of- .
grvee-has-grven

Figure 5.4 Example of a KB

89

giver has more speciﬁc meaning than actor such as the fact that "the giver of giv-

ing" is the same person who originally had the object (this is represented by the E link).

To be a more complete piece of knowledge, some nodes and links should be added
to the ﬁgure, which speciﬁes the time relation between the concepts. For example, the
period of time that giver-has—giving occurs is followed (disjointly) by the period

of the state givee-has-given.

5.2 Marker Passing in SYNSEM

Marker passing is one of the computational methods of using the knowledge base
encoded in a network-based model. A marker starts from an origin and spreads through
links and nodes. Some information is obtained by examining the collisions of markers
from different origins. The idea of marker passing is rooted in the research related to

spreading activation models in psychology.

The spreading activation model was ﬁrst proposed by Quillian(1968, 1969). After
that, much research has been done on this model such as Collins and Loftus(1975),
Anderson(1983), and Lorch(1982). The main idea in Quillian’s language understanding
system, TLC, was that there should be a connection in the memory between two con-
cepts if they are related semantically. The method to ﬁnd the connection(or path) is to
activate the two concepts and let the activation spread through the memory and then ﬁnd
the collisions of the activations from two origins. He called this memory "a semantic
network" which represents the factual assertions about the world. One type of node in
his semantic network is called a "unit" which represents the concept of some object,
event, idea, assertion, etc. Another type of node in his network is called "property"
which encodes any sort of predication (that can be stated by a verb phrase, adjectival,
etc.). A unit or property may contain pointers to other units or properties. In Figure 5.5,
a piece of knowledge about "client" is shown. These pointers are links of the memory

model. Based on this organization of memory, the program can trace to all the units that

client
W * *1
[person] ( a:

 
    
   

(13*

[employ]
[by]

professional]

Figure 5.5 Memory of TLC

can be reached from the starting unit by following the pointers, units, and properties.
The breadth-ﬁrst u'acing method was used to propagate the tags(i.e. markers). The inter-

sections of tags from different origins were used to interpret the input string.

Following Quillian’s idea, Collins and Loftus extended the spreading activation
model to provide psychological validity. One of the major extensions is that the activa-
tions at the nodes have suengths that decay over time and distance. The intersection can

occur only when the total activation at the node should be above some threshold.

The spreading activation model was ﬁrst introduced into A.I. by Fahlman(1979).
Each node of the network memory in his NETL system was implemented using a
hardware device called a node unit. The nodes are connected by links which are imple—
mented by a hardware unit called the link unit. He advocated that the massive parallel-
ism can be achieved with the independent propagations of markers by the node units via
the link units. In addition to these hardware units, NETL has a controller which is called
the network controller. The network controller can query all nodes to ﬁnd how they have
been marked. He tried to deal with some problems of the knowledge base such as type
hierarchies and property inheritance. Posed with certain kinds of questions such as
"what color is Clyde?", markers are passed starting from the nodes, "Clyde" and "color".

The collisions of the markers from different origins are used to answer the question.

91

Some special features such as cancellation links are used in NETL to make nodes use

more speciﬁc information (if it exists) rather than the inherited information.

Charniak(1983a) advocated the use of marker passing for language comprehension.
He suggested that the use of context for language understanding can be easily done by
using the marker passing paradigm. For example, the effort to solve the word sense
disambiguation problem beneﬁts much from the marker passing paradigm. The idea is
that the senses that do not form a reasonable path(or collision) are deleted from the list
of candidate senses of the ambiguous word. This strategy can even be applied in the case
of more than one sentence. Charniak(1986) provided the theoretical basis for the marker
massing paradigm. He showed that the Operation of ﬁnding the paths using marker pass—
ing is equivalent to theorem proving in logic. He suggested that anaphora resolution,
word sense disambiguation and the prepositional phrase attachment problems can be

attacked using the marker passing paradigm.

Granger and his colleagues used marker passing for their text understanding system
called ATLAST (Eiselt, 1985; Granger & etc. 1986). ATLAST consists of the Capsul-
izer, the Proposer, and the Filter. The Capsulizer does the intra-phrase syntactic analysis
whose result is passed to the Filter in the form of capsules. While the input word is input
into Capsulizer, spreading activation starts from the senses of the word. The activation
spreads through the memory which consists of MOPs(memory organization packets)
(Schank, 1982a, 1892b). When the intersection of two activations from different origins
occurs, the Proposer has found some plausible relationships between word senses. The
Proposer passes these possible inference paths to the Filter for evaluation. The Filter per-
forms inter-phrasal syntactic analysis to determine the actor, action, and the object slots.
Another major function of the Filter is to evaluate the inference paths and select one

path among the competing ones.

Hendler(1986) used marker passing for problem solving. His main idea is to reduce

the back tracking in planning by having the marker passing mechanism make

92

appropriate suggestions as to what to try ﬁrst.

Most recently, Norvig(l986) used marker passing for getting the plausible infer-
ences in a text understanding system. The knowledge base is written in KODIAK. From
the collisions of markers whose origins are input words, six types of plausible inferences
are produced. He argued that these inferences are the inferences that should be made to
understand the text properly. I-Iis major claim is that his approach of using the declara-
tive knowledge base and producing the inferences based on this knowledge base is a
generalization unifying the SCRIPT-based and Plan/Goal-based story understanding

approaches.

Another recent approach of using spreading activation can be found in the so-called
connectionist network model (Feldman and Ballard, 1982). In this memory model, all
the work related to syntax and semantics is presumed to be done by spreading the activa-
tions through the network. One of the features they use is inhibition as well as activa—
tion. (Waltz and Pollack, 1984) and (Small, Cottrell and Shastri, 1982) are examples of

attempts to build natural language parsers using the connectionist model.

5.2.1 Links and Flow of Markers

As each word is input into the system, MON creates a new instance node for the
word if it is an Open word. An instance node is a kind of absolute node. Then MON calls
the MP to pass markers starting from the instance node. (Marker passing done by the
MP is called primary marker passing to distinguish it from another marker passing done
by the semantic analysis component.) Figure 5.6 shows that an instance node Z is used
as the origin of a session of marker passing. Z is connected to three absolutes(W, X, and

Y) because the corresponding word has three senses.

In SYN SEM, the links and nodes through which markers can ﬂow are resuicted so
that the markers can not propagate too widely. In each row of Figure 5.7, the node in

the left column can send a marker via the link in the middle column to the node on the

 

 

 

 

 

 

 

 

 

 

 

 

 

Figure 5.6 Origin of Marker Passing

 

 

 

 

 

 

From Via T0
0 I
D >J

 

 

 

 

 

 

 

 

 

 

 

 

A
n
"U

 

 

 

 

 

 

 

 

O I >0

Figure 5.7 Links and Nodes Passing Markers

(ellipse: instance; box: absolute; circle: aspectual; double circle: relation)
right column.

Each marker carries a strength which diminishes as it travels. As a marker passes
through each link (except D and 1 links), its strength is reduced by one. When a
market’s strength becomes zero, the marker can not ﬂow any more. Thus, marker pass-
ing activates only a subgraph of limited diameter in the overall network. The marker

su'ength at the origin is set to be 6. This value was decided after considering several

94

factors. If it is too small, the set of collisions produced during marker passing may not
include some important paths. But, if it is too big, the initial marker strength may result
in too many spurious collisions, which can slow down the system. D and I links do not
reduce the marker strength. These links specify the ancestor relationship and it seems

that all the ancestors of an activated node should be activated too.

Note that the markers can not ﬂow in the reverse direction of D and I links. The
reason is to prevent the markers from propagating too widely in the network. For exam-
ple, if a high node such as thing is marked, then all the nodes under this node (i.e. all
objects in the network) will be marked without this resuiction. This restriction may
result in the reduction of some inferencing power of the system. It has not been found

that any important inference is lost because of this restriction.

SO the computational cost is the major factor for deciding the initial marker
strength and the resuiction of links. It is necessary to keep in mind that a marker can not
ﬂow back along any link through which it ﬂowed. Without this restriction, markers

would get into inﬁnite loops.

5.2.2 Basic Considerations of Marker Passing

The marker passing algorithm used by the MP is strictly a breadth-ﬁrst algorithm.
The whole marker passing operation related to a word is called a session of marker pass-
ing. A marker passing session starts from an instance node for an open word from the
input string. This instance node is the origin of markers in this session. During the
marker passing session, the MP knows the origin and thus the origin need not be
included in the information that a marker should carry. A marker carries the following
information: (1) the next nodes to visit, (2) the previous node it has just come from, (3)

the link through which it has just come, and (4) the strength of the marker.

When a marker visits a node, its trace is put on the node. The trace is a kind of

stamp that is used later to ﬁnd the collisions on the node. A marker trace has the

95

following information: (1) the origin of the marker, (2) the suength of the marker, (3)
the immediate previous node, and (4) the link through which the marker has come to this
node from the immediate previous node. A node may be marked by markers from dif-
ferent origins during the parsing of a sentence. Thus a node may have more than one
trace. During a session of marker passing, a marker may collide with other markers from
other origins. This happens when a marker reaches a node which has a trace of a marker
from any previous marker passing session. The collisions which occur during the marker

passing session are collected and reported to the caller of the MP after the session.

If a marker reaches a node which does not have the marker trace of the same ses-
sion, this means that this node is marked for the ﬁrst time in the session. There is no
problem in this case. A marker trace for this marker is put on the node and the marker is

passed to the neighboring nodes. This case is illustrated in Figure 5.8.

 

A 11 E

Figure 5.8 Split of Markers

Consider a case that a marker has arrived at node B from node A, and E has never been
visited during the session. A marker trace is put on E which is a list: (origin strength A
11). If the strength is not 0, the marker should be propagated to the neighboring nodes A,
B, and C assuming that 11, 12, and 13 are legal for the ﬂow of markers. Because this
marker has just come from A via 11, it can not ﬂow through 11 again. But 12 and 13 can
be used for passing the marker, and thus B and C will get marked. It can be said that the
marker at E is split into two markers. One of them will be a marker that ﬂows through 12

and mark B. The other one will flow through 13 and mark C. The mark at B will mark D

96

and F later. If a marker at a node, say node-x, ﬂows to the next node, say node-y, via a

link, then the marker at node-y is called a descendant of the marker at node-x.

Complication arises if a marker reaches a node which already has a marker trace of
the same marker passing session (i.e. this node was already visited during this marker
passing session). Two possibilities can be considered in this case: (1) the marker is the
descendant of the marker that already marked the node, (2) the marker is not the descen-

dant. Case (1) happens when a loop has been formed as shown in Figure 5.9.

 

 

 

 

 

Figure 5.9 Loop in Marker Passing

The marker that marked the node B propagates to B again (in the guise of a descendant)
via the path that is a loop in this ﬁgure. In this case, B is not marked again. This marker
which returned via a loop is just thrown away; Otherwise, marker passing will go on

inﬁnitely along the loop. Case (2) is illustrated in Figure 5.10. Node B was marked by a

,Q

 

{I V4

   

Figure 5.10 Merge of Markers

marker that reached B via A from the origin. Later, a marker which followed a different

path reaches B via C. This marker is not the descendant of the marker which put a

97

marker trace on B earlier. In this case, the marker trace at B is modiﬁed in such a way
that the previous node ﬁeld of the marker trace contains both A and C. We call this
operation "merging of two markers". But additional processing is required for this opera-
tion. Note the following fact. When a collision is reported, the paths to the origins are
computed and stored as part of the data about the collision. Because of this, the path
information of the collisions whose path goes through B should be updated so that the
new added branch can be a part of the paths in the path information of the collisions.
The easiest way to explain this processing is to use an example. In Figure 5.10, let’s
assume that a marker propagated from origin 01 to E via A, B, and D. At E, a collision
occurred with a marker from origin 02 (of a previous marker passing session). Accord-
ing to this collision, the path (01 A B D E F O2) is computed and stored. Later,
"merging of two markers" occurs at node B. Then the path information of the above col-
lision should be updated so that another path corresponding to the new branch can be

added. Thus the following path should be added to the collision: (Ol .. C B D E F .. 02).

5.2.3 Follow-on Collisions

As explained before, a collision is reported if a marker reaches a node which was
marked from another origin in a previous marker passing session. As soon as a collision
is detected, the MP ﬁnds all the paths corresponding to the collision by tracing the mark-
ers to the two origins. There can be more than one path found because of the merge of
two markers as explained in the previous section. For each collision, the following infor-
mation is gathered and stored: the node at which the collision occurred, paths
corresponding to the collision, and the two origins involved. At the node of a collision,

the fact that a collision between two origins occurred is logged for later use.

Normally collisions are expected to provide useful information. But, in reality,
many useless collisions called follow-on collisions (Hendler,l986) are reported. Let’s

consider Figure 5.11 and assume that the nodes, A, B, and C were marked by markers

98

 

 

Figure 5.11 Follow-on Collisions

whose origin was 01 during the previous marker passing session. Assume that a marker
reaches C during the marker passing session corresponding to the origin 02. This colli-
sion is a normal collision and is called a regular collision. The marker will reach B
from C. A collision is also reported at B because B was marked by a marker from O].
The marker will reach A from B and a collision will also be reported at A. But these col-
lisions at A and B are the dummy collisions corresponding to the regular collision at C,
because they provide no additional information. All these collisions correspond to the
same path, (01 .. A B C D .. OZ). From this example, we can see that more than one col-
lision can be reported for the same path. Out of these collisions, only one can represent
the path and others can be considered to be dummy collisions. Thus we call the ﬁrst col-
lision a regular collision and the Others the follow-on collisions. But note that there can

be more than one regular collision between two origins as shown in Figure 5.12.

 

    

X: regular collisions
O: follow-on collisions

Figure 5.12 Regular and Follow—on Collisions

Detection of the follow-on collision is done as follows:

99

When a marker (whose origin is O2) reaches a node, say A, which was visited by a
marker of the previous session (related to the origin, say 01); this means that a col—
lision occurred at A; Identify the previous node (say B) of this marker and check if
either a follow-on or regular collision between markers from 01 and OZ occurred
at that previous node (and the marker from 01 at node A was from node B); If it
did, the collision at node A is a follow-on collision; Otherwise, the collision is a

regular collision.

According to this detection method, it is clear that the collision at node C in Figure
5.12 is the regular collision and the collisions at node B and A are follow-on collisions.
The follow-on collisions are also logged on the node at which they occur so that the

information can be used later.

5.2.4 Overlapping Collisions

It is necessary to consider another kind of useless collision in addition to the

follow-on collisions. Let’s consider Figure 5.13 and assume that the marker

 

Figure 5.13 Overlapping Collisions

whose origin is 01 ﬂows through the nodes in the following order: A, C, D, E, etc.
Assume also that the marker from origin 02 ﬂows through the nodes: B, C, D, E, etc.
The collisions between the two markers occur at C, D, and E. The collision at C is the

regular collision. The fact that the collisions at D and E are false collisions is clear when

100

the paths for the collisions are compared as shown below:

Collision C1 at C: (01 A C B ..... 02)
Collision C2 at D: (01 A C D C B ....02)
Collision C3 at E: (01 A C D E D C B ....O2)

The paths C2 and C3 have no more information than C1, because the segment "CDC"
and "CDEDC" does not add any information. Thus the collisions, C2 and C3, are the
dummy collisions of C1. Note that these dummy collisions can not be detected by the
method to detect the follow-on collisions. We call the collisions such as those at D and E

the overlapping collisions because the ﬂows of the two markers overlap.

The scheme to detect an overlapping collision is simple. At a node at which a colli-
sion occurs, say D, compute the paths to the two origins. If the two paths share some
nodes and links, then the collision at the node D is an overlapping collision. In Figure
5.13, a collision occurred at D. The paths to the two origins are (D C A 01) and (D C
B ....02). Because (D C) is shared between the two paths, the collision at D is an over-
lapping collision. The collision at E is also an overlapping collision because the path
from E to O], i.e. (E D C A ...Ol), and the path from E to 02, i.e. (E D C B ...O2), share

the portion (E D C). Overlapping collisions are just thrown away.

5.2.5 Marker Passing Algorithm

Now the overall ﬂow of the marker passing algorithm used in SYNSEM can be
introduced. It is shown in Figure 5.14.

The initial strength of the marker is set to some ﬁxed number(6 in the current
implementation) and the strength is reduced by one as it passes a link(except D or I
link). Therefore the number(say W) of nodes being marked during a marker passing ses-

sion does not increase exponentially according to the size of the KB(say N) (

W=(kN)6

101

because the branch factor can be assumed to be kN where k is a constant and N is the

number of nodes in the KB). It follows that the number of collisions and the number of

possible paths for collisions do not increase exponentially according to the size of the

KB.

 

step 0:
step 1:

step 2:

insert a marker (origin nil nil 6) into an empty queue, Mqueue.
if Mqueue is empty, goto stop.
Marker <— a marker popped out from the front of Mqueue.
Set currentnode, previousnode, fromlink and strength from Marker.
if currentnode already has a trace of this session
then if the situation is infinite loop
then goto step 1
else begin
update the trace so that previousnode is added;
update related collisions to change the path
information;
goto step 1
end
else begin
find and prepare the markers for
all neighboring nodes to which the marker can flow;
(if strength is 0, A or C links can not be used.)
insert the markers into the rear of Mqueue;
if there are traces for previous sessions
then
for each trace do
if overlapping collision
then do nothing
else if follow-on collision
then report and log follow—on collision
else report and log regular collision;
goto stepl:
end
stop;
Figure 5.14 Marker Passing Algorithm

 

102

5.3 Conclusion

A knowledge base written in KODIAK is used as the source of semantic informa-
tion for the SYNSEM parser. KODIAK has three types of nodes: absolutes representing
the objects, states, and actions in the world; relations representing the possible relation-
ships that can exist between two absolutes; and aspectuals which are formal parameters
of relations. The important feature of KODIAK is that it is a relation-based knowledge
representation language. The concepts hidden behind the slots in the frame-based
languages become full-ﬂedged concepts in KODIAK. In KODIAK, the proliferation of
concepts is encouraged in contrast to languages such as Conceptual Dependency that
provide a set of primitives. The meaning of a concept in KODIAK is generated by its
connections to other concepts. KODIAK provides nine primitive links that can connect

the concepts.

The marker passing mechanism is used in SYNSEM as a simple way of providing
semantics to the parser. The detailed considerations on the design of the marker passer
have been explained. To reduce the space of the KB that gets marked and thus minimize
the computational cost, some constraints have been put on the marker passing: resuict-
ing the possible link and node combinations that markers can ﬂow through and restrict-
ing the distance that a marker can travel. The critical design problems such as follow-on
collisions, merging of markers, detecting the loop, and overlapping collisions have been
explained. Use of marker passing for providing semantics to parsing and for resolving

word sense ambiguity is discussed in the next two chapters.

CHAPTER 6

PROVIDING SEMANTICS TO PARSING

It was explained that the parsing is split into several branches when more than one
rule is found by the syntactic pattern matcher. Each branch is a legitimate way of pars-
ing at that point. The parser removes the branches whose semantic processing result is
worse than others by pushing these into the backup stack (i.e. the parser pursues only the
top-rated parses). The number of branches will be reduced (usually down to one) as soon
as syntactic or semantic information allows the parser to remove some branches. There-
fore the important problem in this parsing framework is how the semantic processing
result is computed and how the semantic processing results of the multiple branches are

compared.

It was explained in the previous chapter that the source of semantic information is
the KODIAK knowledge base and the major technique to utilize it is the marker passing
mechanism. A decision situation that requires semantic information which can not be
represented in KODIAK can not be handled in SYNSEM. Thus, we can only claim that
SYNSEM can handle the ambiguities which can be resolved by the kind of knowledge
that can be encoded using KODIAK.

This chapter will start with the explanation of the interface between the syntactic
analysis and semantic processing component. The most important concept is the sugges-
tion which the syntactic analysis component sends to the semantic processing com-
ponent. Any rule that invokes some semantic processing has an action which issues a
suggestion. Each suggestion is interpreted by the semantic processing component and
the result is stored as a part of the semantic interpretation of the branch. The semantic

processing result of each suggestion is realized as a path (or a set of multiple paths if
103

104

there is ambiguity) in the KODIAK knowledge base. Marker passing is useful for
ﬁnding a path for the suggestion. One problem with marker passing is that its computa—
tional cost is big. One reason for this big computational cost is that there are too many
spurious collisions(or paths) reported by the marker passing routine. In this chapter, an
efﬁcient way to remove false collisions is introduced. The problem of comparing the
branches according to the semantic processing result becomes a matter of comparing the

paths. This problem is also investigated in this chapter.

6.1 Interfacing Syntax and Semantics

The main method of interfacing the syntactic processing component(SYN) and the
semantic processing component(SEM) is to use the suggestions that SYN sends to SEM.
A rule whose execution should invoke some semantic processing contains actions that

issue a suggestion. For example, the rule main-verb2 has an action:
(sug NP1 V1 subj)

This suggestion indicates that the syntactic constituent NP1 and V1 are related by the
syntactic relationship "subj". It says that NP1 is the syntactic subject of the clause whose
main verb is V1. The semantic component should consider this relationship when it is

computing the semantic relationship between them.

 

 

 

 

 

 

suggestions _\

SYN ,) SEM
(C1 C2 Prep)
(C1 C2)

 

 

Figure 6.1 Flow of Information from Syntax to Semantics

What the rule main-verb2 tells SEM is that SEM should ﬁnd the appropriate semantic
connection between the semantic object corresponding to the syntactic constituent NP1
and the semantic Object corresponding to V1. But SYN includes a hint "subj" which
should be utilized by SEM. When SYN has analyzed a syntactic object of a clause, it

105

will send a suggestion to SEM saying that the verb of the clause and this object should
be connected via the semantic relation that can be implied by the syntactic relationship

"object". Thus the rule object (which does this analysis) has an action of the form:
(sug V1 NP1 obj)

The suggestion sent by SYN according to the action has the form of "(C1 C2 Prep)". C1
and C2 are syntactic constituents and Prep is a preposition (including pseudo-
prepositions) indicating the syntactic relationship between C1 and C2. Prep is used as a
hint to SEM for connecting the semantic objects corresponding to Cl and C2. Pseudo-
prepositions are the dummy prepositions that actually do not exist but can be assumed to
exist. "Subj" and "obj" are examples of pseudo-prepositions. "Indirect-object" can also
be a pseudo-preposition. But an indirect object can eventually be replaced by the dative
preposition of the dative verb. For example, "to" is the dative preposition of the dative
verb "give", and "for" is the dative preposition of the dative verb "buy". Thus, sugges—

tion (V 1 NP1 indirect-object) can be replaced with (V1 NP1 to) if V1 is the verb " give".
Let’s consider (6-1) and (6-2) to provide examples for suggestions.
(6-1)The man ate a cake with a fork.
(6—2) The man ate a cake with frosting.

In (6—1), "eat" and "a for " are related via the PP which is headed by the preposition
"with" (refer to Figure 6.2(a)). The preposition "with" provides some information about
how "eat" and "a fork" are related. Using the fact that the V, "eat", and the PP, "with a
fo ", compose the VP syntactically, SYN knows that the V and the NP of the PP should
be connected in some way and this relationship needs to be implied by the meaning of
the preposition "with". The preposition "with" might have many meanings. It is up to the
semantic component to use the correct meaning of "with" to connect "eat" and "a fork".
So, for (6-1), the syntactic rule which recognizes this syntactic relationship (i.e. the rule
vp-pp) has an action which issues the suggestion (V 1=eat NP1=a fork prep=with) and

sends it to SEM. For this suggestion, SEM ﬁnds the KB path which connects "eat" and

106

"a for ". The preposition "with" is used to ﬁnd the correct path (i.e. the path that con-

veys one of the meanings of "with").

2””; K A/{\
1 P P1 P1 /pp,1\
“:1 A a cake Preﬁvl X

a fork with frosting
(a) (b)

Figure 6.2 PP Attachment Depending on Semantics

eat

In the case of (6-2), the NP "a cake" is modiﬁed by the PP "with frosting" (consider
Figure 6.2(b)). This analysis is done by the rule np-pp. This rule knows that NP1 and
NP2 would be related in some way semantically and this relation is implied by one of
the meanings of "with". Thus, this rule has an action which issues the suggestion
(NP1=a cake NP2=frosting prep=with).

According to the framework described so far, it is clear that a syntactic rule to
which some semantic processing is related has some actions which issue suggestions to
SEM. The decision about which rules should have what kind of suggestions should be
determined by the grammar writer. Note that a suggestion action is of the form (C1 C2
Prep) where C1 and C2 are syntactic constituents. Thus the speciﬁcation of a suggestion

action is in purely syntactic form which is domain independent.

So far, we explained how the syntactic analysis component provides SEM with a
message which is used for semantic processing. This is the way syntax requests seman-
tics to do some semantic processing. The next question that arises is how semantic pro-
cessing inﬂuences the processing of syntax. This is achieved by the removal of some
branches based on the semantic processing result. Consider Figure 6.3. Each branch is
created for each syntactic rule whose patterns match CLIST. For each branch, there are

suggestions which are sent to SEM. The KB path for each suggestion is found by SEM

107

br ch-l br ch- branch-k
provide
$5331?“ 0 o o
sem. proc. i 3

component . ' .
semantrc comparison stage

(ﬁlter branphes)

%

 
    
 
   
   
 

 

 

 

 

continuation
of branch-2

Figure 6.3 Inﬂuence of Semantics in Parsing

and it is returned to the branch to be stored. At the semantic comparison stage, the
branches are ﬁltered according to the semantic processing result (actually the goodness
of the KB path). Thus the branches having inferior semantic effect are suppressed by
pushing them into the backup stack and the branches (i.e. syntactic analysis alternatives)
with best semantic effect survive. This is one way that semantics can inﬂuence the pro-

cessing of syntax.

Another way that semantics can inﬂuence syntactic processing is to notify SYN
that the suggestion is so bad that a reasonable path can not be found. In this case, the
branch that issued the suggestion is just thrown away (without being backed up in the

stack).

Interfacing syntax and semantics explained in this section are the major ways of
how syntax and semantics communicate in SYNSEM. But it should be noted that it
might be necessary to have other ways of making syntax and semantics exchange infor-
mation. Natural language is too complex to classify the communication between the syn-
tactic processing level and the semantic processing level in only a couple of ways. But

we claim that the methods of interfacing SYN and SEM explained in this section are

108

major ways of communication.

6.2 Finding a Path for a Suggestion

As explained in the previous section, SEM should ﬁnd a path in the KB
corresponding to the suggestion. The path should be the best path for the suggestion.
Normally only one path is the best one, but there can be more than one best path if there
is word sense ambiguity. The suggestion that is issued by the suggestion action of a syn-
tactic rule is speciﬁed using the syntactic constituents. Then this suggestion is converted
into the suggestion speciﬁed using the nodes in the knowledge base. Let’s consider sen-
tence (6—1) as an example. The result of syntactic analysis after "ate" has been read is

shown in Figure 6.4(a). This analysis is done by the rule main-verb2 which contains a

/\

NP01 VP02 (NP01 V03 subj) (man.l ate.2 subj) (man.1 ate.2)

/

V|03

ate

(a) (b) (C) (d)

suggestion

Figure 6.4 Conversion of Suggestions

action (sug NP1 V1 subj) as explained before. During the pattern matching of the rule,
NP-l is mapped to NP01 and V-l is mapped to V03. According to this action, the
suggestion in (b) is produced by SYN. Note that NP01 and V03 are actually pointers to
the syntactic constituents in the tree in CLIST (shown in (a)). The suggestion in (b) is
converted into the suggestion in (c) by replacing the pointers with the corresponding
concept nodes in the KB. Refer to the KB shown in Figure 6.5 to follow the example in
this section. The suggestion in (c) is of the form (I1 12 Preposition) where
11 and IZ are the KB nodes and Preposition is the preposition given by the rule

("subj" in this example). NP01 in (b) is replaced with man . 1 which is the instance

 

thing Ob' event
A
O O O O o D O O
D
thy_0bj

 

 

n

, “i , “‘ .O.

food namma nimat “

I j)
cake [Erna] .
1 [MI

® - D

C ‘ t__t001 h h_ani [I .
D

D
. .
m (C

Figure 6.5 Another Example of the KB

 

 

 

absolute created by MON. MON created it when it input the word "man". In a similar
way, V03 in (b) is replaced with the instance node ate . 2 in the KB. Thus the sugges-

tion in (c) is obtained. The conversion of the form is done by the interface routine.

After getting the suggestion in (c), all SEM should do is to ﬁnd the best path
between the node man . 1 and ate . 2 via some relation node that has one of the mean-
ings that "subj" can indicate. Note that the MP passed markers from the instance nodes
when the nodes were created. The collisions of markers whose origins are man . l and
ate . 2 are available. A collision corresponds to one or more paths between the origins
of the markers that collide. One of the biggest problem that marker passing-based sys-

tems suffer from is that there are too many spurious(false) collisions (Charniak, 1983a,

110

1986). Charniak said that the ratio of good collisions to the spurious collisions is 1 to 10.
Without solving this problem, the marker passing-based system can be slow, which

prevents it from being used as a practical system.

6.2.1 Secondary Marker Passing: Motivation and Basic Concept

We have developed a technique which can efﬁciently remove most of the spurious
collisions. The basic idea is to use the information that the preposition in the suggestion
can provide. Let’s assume that the suggestion to be processed by SEM is (I1 12
Preposition). Preposition provides some meaning which should be reﬂected
on the path between II and I2. The collisions of the markers from I1 and I2
reported by the MP are called two-way collisions because each collision consists of two
markers. The set of two-way collisions between I 1 and I 2 contains not only the
correct paths (reﬂecting the meaning of Prepos it ion) but also many other incorrect
paths. From now on, the collision and the path will be used interchangeably, except for
cases in which some confusion may occur. Note that a collision can correspond to more
than one path because of the "merging" of markers explained in Chapter 5. Without the
information Preposit ion provides, it is difﬁcult to remove the spurious collisions.
(Even if it can be done, it takes too much time.) It is necessary to develop a technique by

which the spurious collisions are removed efﬁciently.

The technique developed for this purpose is called secondary marker passing. Most
of the spurious collisions can be efﬁciently removed using this technique. This technique
removes all the collisions whose path does not reﬂect one of the meanings of Prepo-
sition. Let’s consider sentence (6-1). The fact that "the man" has the syntactic rela-
tion of "subject" with "eat" indicates that they are related by a semantic relation that can
be implied by the pseudo-preposition "subj". The input string "ate" and "a for " are syn-
tactically related via the preposition "with". This indicates that the concept for "ate" and

the concept for "a fork" should be connected in the KB via a semantic relation that can

111

be implied by one of the meanings of the preposition "with". From this consideration,
we can say that the best path between I 1 and I 2 should pass through a concept node
whose meaning can be implied by Preposit ion in the suggestion (11 12

Preposition).

The concept nodes which can represent the semantic relations that are implied by
the meanings of the prepositions are the nodes of the type "relation" in KODIAK. The
relation nodes such as actor-action, eater-eat, recipient-receive,
etc. are some of the nodes that can be implied by the pseudo-preposition "subj". The
relation nodes such as instrument-action, instrument-eat, has-part,
etc., are implied by the preposition "with". We can think of the hierarchy of the relation
nodes for each preposition in the KB. The root of the hierarchy can be assumed to be
the preposition itself. We can recognize these hierarchies in the KODIAK knowledge
base. The hierarchy in Figure 6.6(a) is for the pseudo-preposition "subj", and

 

 

  
 

instru

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

(a) m (c)

 

 

 

\'7

Figure 6.6 Hierarchies of Relations

that in (b) is for the pseudo-preposition "obj", and that in (c) is for the preposition

112

"with". Note in the ﬁgure that breakee-break belongs to the two hierarchies. For
example, "the rock" in (6-3) is the syntactic object of "broke" and the relation
breakee-break is implied by "obj" but the rock in (6-4) is a syntactic subject of

"broke".
(6—3)The man broke the rock with a hammer.
(64) The rock broke.

Thus the relation node breakee-break is implied by "subj". The root node of each
hierarchy in Figure 6.6 is a dummy node representing the preposition. The link between
the relation nodes in the hierarchy is implemented using the D link in KODIAK. If the
relation node A is a super-class of another relation node B, then there is a D link coming

from B to A.

For the suggestion (Instancel Instance2 Preposition) , the best path
between instancel and instance2 should pass through a relation node that is
implied by Preposition. This means that the path should pass through a relation
node that is in the hierarchy whose root is Preposition. The paths found by the MP
(i.e. two-way collisions) that do not pass through a relation node in the corresponding

hierarchy need not be considered.

This idea can be implemented using another marker passing. A marker different
from the markers used by the MP are passed from the root node of the hierarchy for the
preposition in the suggestion. The markers used in this process are called the secondary
markers (note that the markers used by the MP are called the primary markers). All the
relation nodes in the hierarchy can be marked by the secondary markers if the markers
propagate via D links between relation nodes in the hierarchy. Any path which does not
contain a relation node that is marked by a secondary marker can be removed as a false
path because it does not pass through any relation node that has a meaning of the prepo-
sition. At this point, it is important to note that the relation node of a good path which is

marked by a secondary marker should also have markers from instancel and

113

instance2 as shown in Figure 6.7.

 

Figure 6.7 Three-way Collision

The reason is that this relation node, node A in the ﬁgure, represents the semantic rela-
tion between instancel and instance2 and thus the node is related to both
instancel and inst anceZ. Thus, the markers from the two origins should be able
to reach the node A. This means that three markers should collide at the relation node A
as shown in Figure 6.7. This collision is called a three-way collision because the mark-
ers from three different origins form the collision. From this consideration, the task of
removing the paths which do not contain a relation node marked by a secondary marker
can be achieved by collecting only the three-way collisions. Then the routine for ﬁnding
the best path needs to consider the three-way collisions instead of the two-way colli-
sions. This will save a lot of computation because the number of two-way collisions(or
the paths) should be much bigger than the number of three-way collisions (the data of
the experiment will be given later).

It has been explained that all the relation nodes in the hierarchy for Preposi-
t ion are marked during the secondary marker passing session. But this is not true. The
amount of computation can be reduced using an updated scheme. For a large knowledge
base, the number of relation nodes in a hierarchy for a preposition can still be big. Con-
sider, for example, the pseudo-preposition "subj". The concept nodes for almost all verbs
in the domain are linked to relation nodes that are implied by "subj", which means that
the number of relation nodes in the "subj" hierarchy is almost the same as the number of

verbs in the domain. To reduce the number of relation nodes that are marked during the

114

secondary market passing session, the following scheme is used. When a relation node
gets marked by a secondary marker, SEM can pass the secondary markers from this
node only if a three—way collision has occurred at this node. Thus further propagation of
the secondary markers stops at a relation node at which no three-way collision occurs.
According to this scheme, the system need not mark all the relation nodes in the hierar-
chy. Only part of the hierarchy will get marked. Consider the suggestion (eat . 1
spoon . 3 with) for an example. Only the portion of the "with" hierarchy
corresponding to the node inst rument-acti on will be marked as shown in Figure

6.8.

  

 

 

instancel in stance2

Figure 6.8 Cut-down of Marker Passing Space

The sub-hierarchies corresponding to has—part and participant-event will
not get marked during secondary marker passing for the suggestion because there is no

three-way collision at has-part or participant—event.

Now, the secondary marker passing session can be stated more precisely as fol-

lows:

l. A secondary marker passing(SMP) session occurs for each suggestion from
SYN. But the suggestion should contain a preposition. Secondary marker
passing is done by SEM before it attempts to ﬁnd the best path for the sugges-
tion.

2. The output of an SMP session is a set of three-way collisions. This output is
used by the routines for ﬁnding the best path.

115

3. The origin of secondary marker passing is the dummy root node of the hierar-
chy corresponding to the preposition.

4. A marker is passed only through the D links (in the reverse direction) between
the relation nodes.

5. A marker which reaches a relation node can not ﬂow any more if a three-way
collision does not occur at this relation node.

There is no limit to distance that a secondary marker can travel.

All secondary markers are removed after each SMP session.

6.2.2 Consideration on the Polarity of Paths

Related to secondary marker pas sing, is one problem that should be considered dur-

ing the SMP session. This problem can explained using (6-5) and (6-6).
(6-5)The man with moustache
(6-6)The moustache with the man

During the analysis of (6-5), SYN would send a suggestion (man.1 mous-
tache.2 with) to SEM and SEM will ﬁnd the path, (1) in the ﬁgure, between
man.1 and moustache.2 which contains a relation node has-moustache,

because a three-way collision occurs at has-moust ache (illusu'ated in Figure 6.9).

 

 

 

 

 

 

 

 

 

with

(j

kvj

oustache
/man
/ ‘
(1) has-moustache .
man.2 man.1 moustache.2 moutasche.1

Figure 6.9 Two Collisions with One to be Removed

In the case of (6-6), the suggestion (moustache . 1 man . 2 with) is sent to SEM.
Note that the same path, (2) in the ﬁgure, as the path for (6—5) is found in this case

according to the three-way collision during the secondary marker passing operation. But

116

(6-6) is semantically unacceptable even if a good path is found. Thus either this path
should not be found during secondary marker passing, or the path is found during secon-
dary marker passing but is rejected later by some ﬁltering routine. From this observa-
tion, it is realized that the order of two concepts in the suggestion is important.
(man . 1 moustache . 2 with) is semantically acceptable and a good path is found,
but (moustache . 1 man . 2 with) is semantically bad and thus no path should be

found by SEM.

This kind of false path can be removed during secondary marker passing. The idea
is to use the order of aspectuals for a relation. We put an order for two aspectuals of
each relation in such a way that the aspectual whose corresponding word appears before
the word corresponding to the other aspectual is given the ﬁrst place and the other aspec-
tual is given the second place. For example, the relation subj (the root node for the
hierarchy of pseudo-preposition "subj") has two aspectuals, pre . subj and
pos . subj. Pre . subj is the aspectual that corresponds to the input suing of the syn-
tactic subject. Pos . subj is the aspectual that corresponds to the verb of the clause.
Pre . subj is called the ﬁrst aspectual and pos . subj is called the second aspectual
of subj because the subject occurs before the main verb in a clause. Note that aspec-
tuals of two relations that are connected by a D link are paired by the role/play relation-
ship as illustrated in Figure 6.10. Actor-act dominates eater-eat (i.e. there is a
D link from the latter to the former node). The aspectual actor . actor-act is in the
role/play relationship with eater . eater-eat. Similarly act . actor-action is
in the role/play relationship with eat .eater-eat. The ﬁrst aspectual of subj,
pre . subj, is linked to actor. actor-act. Actor. actor-act is linked to
eater.eater-eat by the role/play relationship. The second aspectual of subj,
pos . subj, is linked to act . actor-act which is linked to eat . eater—eat by
role/play. The role/play relationship is used to link the aspectuals as shown in the ﬁgure.

Because the subject appears before the verb the aspectuals that denote the subject are

117

man ate
I
I

' I
' I

pre.subj “ ” pos.subj
1 t J R i

I I

a . . .

I

gimme *‘C’ “ A a A I '

I
. ’ l D
D = ;

._ of
person 0 AOAw

eaten-I eater-eat eat.-

    

 

     
 

 

 

 

 

 

act

 

act.-

 

 

 

 

 

 

 

 

 

 

 

 

 

Figure 6.10 Role and Play of Aspectuals of Relations

linked to the ﬁrst aspectual of subj, pre . subj.

Figure 6.11 shows the case for the preposition "with". Note that the ﬁrst aspectual
of with is linked to the aspectual whole . has-moust ache Of has—moustache

because the referent of whole.has-moust ache appears before the preposition

"with" followed by the referent of part . has —mousta che.

pre with .s.with
O A A

' R
. . .

‘ hy obj ct @h phy_obiect
«C,— A A . -

who C. 311:.- ll
D . D

\_ ‘ F A: h
person A A w mousta e

whole.- has-moustache part.-

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Figure 6.11 Example of Role and Play

Let’s explain how the path is rejected for (6—6). SYN always puts the constituent appear-

ing before the preposition (here "with") in the ﬁrst element of the suggestion and the

118

constituent after the preposition in the second element of the suggestion. Thus the
suggestion for (6-6) is (moustache . 1 man . 2 with) . For each three-way colli-
sion that has been found, SEM checks to see if the ﬁrst instance node in the suggestion
(moustache . 1 in this example) and the aspectual (of the relation of the three-way
collision) which is in the role/play relationship with the ﬁrst aspectual of the preposition
node, are on the same part when the path is divided into two halves by the three-way
collision node. In the example shown in Figure 6.11, the aspectual of has-
moustache which is linked to the ﬁrst aspectual of with is whole .has-

moustache. The two half paths are:
(man.2 person whole.has-moustache has-moustache)

(has-moustache part.has-moustache moustache mous-

tache.1)

We can see that whole . has-moust ache and moustache . 1 are not on the same

half path. Therefore this three-way collision is thrown away.
Related to the problem discussed so far, it is worth considering (6-7) and (6-8).
(6-7)The man ate the apple.
(6—8)The apple was eaten by the man.

During the analysis of (6-7), SYN will send the suggestion (man . 1 ate . 2 subj)
after "ate" is analyzed. During secondary marker passing for this suggestion, a three-way
collision occurs at the relation node eater-eat. This suggestion is issued by the
action (sug NP1 V1 subj) of the rule main-verbz. NP1 is put before V1 because NP1
appears before V1 in the input suing. What should be the suggestion action of the rule

passive-by which analyzes "the man" as the actual subject of "eaten" in sentence (6-8)?

119

This rule is shown below:

(passive-by ((l np—pp np—pp1)(h ))
[ ( [ (DP-1 8‘1) 1 () l
[ (aux-1 s-1) 1 (passive) ]
[ (VP-1 8'1) 1 () l
[ (PP—1) 0 () l )
( [equal-exact (symbol by) (prep—of-pp pp-1)] )
l
[ (sem connect (v vp-l) (np pp-l) subj 2)
(attach pp—l to vp-l)
]
)

In the suggestion of this rule, "(sem connect (v vp-l) (np pp-l) subj 2)", the ﬂag, 2, is
used to signal to SEM that the V appeared before the NP in the input string but it should
be assumed that it appeared after the NP. Then SEM uses this information to change the

order of the V and NP and the analysis is done correctly.

6.2.3 Heuristics for Finding the Best Path

After the secondary marker passing session for a suggestion, a set of three-way col-
lisions is reported. The number of three-way collisions in this set is small compared with
the number of two-way collisions, but there is no guarantee that only one three-way col-
lision is returned. Even if there is only one three-way collision reported, the best paths
should be found (note that one collision can represent more than one path). Thus it is
necessary to develop heuristics that can ﬁnd the best paths, given a set of three-way col-

lisions. The course of ﬁnding the best path is shown in Figure 6.12.

 

 

 

suggestion Secondary
(11 12 prep) Marker :> Heuristic-1 > Heuristic-n >
Passing 3W3Y‘ best paths

collisiofls

 

 

 

 

 

 

 

 

 

Figure 6.12 Stages of Finding Best Paths

120

Note that the scheme in this ﬁgure applies only to the suggestions containing a preposi-
tion. Each heuristic in Figure 6.12 is actually for reducing the number of paths(or colli-

sions) which are the candidates for the best path.

6.2.3.1 Hierarchy: H-heuristic

The ﬁrst heuristic that is in the order is called the H-heuristic (H stands for hierar-
chy). The input to this heuristic should be the set of three-way collisions which is the
output from the secondary marker passing session. The idea behind the H—heuristic can

be illustrated using Figure 6.13.

 

Figure 6.13 Removal of a Collision Using Ancestor Information

Three-way collisions occur at node A and B. B is the ancestor of A. This means that A
can be reached from B via D links (in the reverse direction). It is obvious that the path
for the three-way collision at A is better than the path for the collision at B because A is

a more speciﬁc concept than B. The H-heuristic is deﬁned precisely as follows:

H-heuristic: A three-way collision is removed if its collision node is an ancestor of

a node at which a three-way collision also occurs.

The result after applying the H-heuristic becomes the input to the next stage of heuris-

tics.

121

6.2.3.2 Length: L-heuristic

The heuristic called the L-heuristic is applied to the result after applying the H-
heuristic. Thus the input to this stage is a set of three-way collisions. The set is con-
verted into a set of the KB paths according to the following scheme: For each three-way

collision,

(i) Starting from the collision node, trace the marker back to the origin node.

This results in a set of paths.

(ii) Starting from the collision node, trace the marker back to the other origin

node and get the set of paths.

(iii) For each path from the set in (i) and for each path from the set in (ii), form a
full path by combining the two paths. The result of this step is the set of full
paths.

To the output of step (iii) above, the L—heuristic(L stands for length) is applied. This

heuristic is deﬁned as follows:
L-heuristic: Select the paths with the smallest number of relation nodes.

The idea behind this heuristic is that the shorter the path is the better the path is. The
reason why the number of relation nodes is used instead of the number of all nodes
should be noted. The number of absolutes and D links between the absolutes might not
be a reliable source Of information for ﬁltering paths. Consider an example in Figure
6.14 to illustrate this. Let’s assume that node A is an absolute which is as speciﬁc as the
absolute node B. But the knowledge engineer put more descendants under B than under
A because he happened to know the hierarchy under B was better than the hierarchy
under A. Therefore the number of nodes between A(or B) and the instance node should

not be used in computing the length that is used in the L-heuristic.

122

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

in
A1 path-2
D . 132
B3 D
I I
instance

Figure 6.14 Number of Nodes in the D-hierarchy

6.2.3.3 Speciﬁcity: S-heuristic

The heuristic called the S-heuristic is applied to the set of paths output from the
stage of the L-heuristic. The source of information for the S-heuristic is the speciﬁcity of
the absolutes of the path (S stands for the speciﬁcity). To compare two paths based on
the speciﬁcity, it is necessary to decide which absolutes(or relations) of the paths should
be used for comparison. Note that a path might have many absolutes and up to three
relations (because the initial strength of a marker is 6). There are three possible forms of

paths as shown in Figure 6.15.

@‘OQVP PM Is WP
[j .

i if. Ii E.

Figure 6.15 Possible Forms of Paths

A path with any of the three forms might be compared with another path with any
of three forms. The nodes that are used for the comparison should be chosen in a way
that does not lose any information. In the S-heuristic, two absolute nodes are used for

this purpose: the absolute node that is pointed to by the ﬁrst C(constraint) link and the

123

absolute node that is pointed to by the last C link (the dark nodes in the ﬁgure). These
two absolute nodes are called the end absolutes of a path. The comparison of any two
paths (say, path-a and path-b) is depicted in Figure 6.16. Note, in this ﬁgure, that
instancel of path-a is the same concept as instancel of path-b. Similarly,
instance2 of path-a is the same as instance2 of path-b. The reason for this is
that path-a and path-b are the paths for the same suggestion, (inst ancel

instance2 preposition).

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

   

instancel instance2

Figure 6.16 End Absolutes of two Paths to be Compared
The ﬁrst end absolute of path-a, a1, is compared with the ﬁrst end absolute of path-b, bl,

and the second end absolute of path-a, a2, is compared with the second end absolute of

path-b, b2. Now, the S-heuristic can be deﬁned as follows (using the above notation) :

S-heuristic: If a1 is an ancestor of bl and a2 is an ancestor of(or same with) b2,
then path-b is better than path-a and thus path-a is removed. Symmetrically, if a2
is an ancestor of b2 and a1 is an ancestor of(or same with) b1, then path-b is better

than path-a and thus path-a is removed.

Any path which is inferior to any other path based on this heuristic is removed.

6.2.4 Operating Example of Finding the Best Path

In this subsection, some examples for ﬁnding the best path will be given. Let’s
consider how the path ﬁnding task is done for the suggestions sent by SYN during the

analysis of sentence (6]). The syntactic rule main-verb2 sends the suggestion (man . 1

124

ate . 2 subj) to SEM after "ate" is analyzed. Let’s use the small KB shown in Figure

6.5 to make the explanation easy. There are four two-way collisions of primary markers

from the origins, man . 1 and ate . 2. The collisions correspond to the following six

paths (nodes at which a two-way collision occurred are underlined. Links and aspectuals

are omitted):

(6—9)(man.l man person high_animate animal animate

actor-act act trsact eat ate.2)

(6-10)(man.l man person high_animate animal animate

phy obj food eatee_eat eat ate.2)

«$1D(man.1 man person high_animate animal animate

phy obj inanimate eat_tool instr-eat eat ate.2)

(6—12)(man.1 man person high_animate animal animate

phy obj inanimate instr-act act trsact eat ate.2)

(6-13)(man.1 man person high_animate animal animate

phy_obj thing actee-trsact trsact eat ate.2)

(6—14)(man.1 man person high_animate animal eater—eat eat

ate.2)

SEM passes secondary markers starting from subj. Subj, actor—act,

and eater-eat get marked during this secondary marker passing session. Then two

three-way collisions at actor-act and eater-eat are reported. The paths

corresponding to these three-way collisions are (6-9) and (6-14) and they are shown

below again with the three-way collision node underlined:

(6-9)(man.1 man person high_animate animal animate

actor—act act trsact eat ate.2)

 

(6-14)(man.1 man person high_animate animal eater-eat eat

 

ate.2)

125

Note that the collision node has been changed from act to actor—act in (6-9). The
reason is that a follow-on collision collected during primary marker passing can also be
used to form a three-way collision. Note that SEM needs to consider only two paths
after secondary marker passing. Otherwise the six paths from (6-9) to (614) should have
been considered. Then the H-heuristic is applied to the paths (6-9) and (6-14). Because
actor-act (the three-way collision node for (6-9)) is an ancestor of eater-eat
(the three-way collision node of (6-14)), the path (69) is removed. So only one path,
(6-14), is left after the stage of the H-heuristic, which means that the path ﬁnding opera-
tion need not go through the stages of the L-heuristic and the S-heuristic. The best path

for the suggestion (man . 1 ate . 2 subj) has been determined to be the path (614).

After "the cake" has been analyzed in sentence (6-1), the object rule sends the
suggestion (ate . 2 cake . 3 obj) to SEM. The two-way collisions (actually the

corresponding paths) between ate . 2 and cake . 3 are as follows:
(cake.3 cake food eatee-eat eat ate.2)
(cake.3 cake has-part1 frosting food eatee-eat eat ate.2
(cake.3 cake food asymbjggt animate animal eater-eat eat ate.2)
(cake.3 cake food psy object animate actor-act act trsact eat ate.2)
(cake cake food psy object inanimate eat_tool instr-eat eat ate.2)
(cake.3 cake food psy_o_bjec_t inanimate instr-act act trsact eat ate.2)
(cake.3 cake has-part1 frosting food phy object animate animal eater-eat eat ate.2)
(cake.3 cake has-part1 frosting food phy_objgt animate actor-act act trsact eat ate.2)
(cake.3 cake has-part1 frosting food W inanimate eat_tool instt-eat eat ate.2)
(cake.3 cake has-part1 frosting food phy_object inanimate instr-act act trsact eat ate.2)
(cake.3 cake food phy_object Mtg actee-trsact trsact eat ate.2)
(cake.3 cake has-partl frosting food phy_object _thi_ng actee-trsact trsact eat ate.2)

Instead of considering these 12 two-way collision paths, secondary marker passing is
done starting from the node obj. Actee-trsact and eatee-eat get secondary
markers and three-way collisions occur at these two relation nodes. The three—way colli-
sions and the corresponding paths are as follows (the three-way collision nodes are

underlined):

126

(6-15)(cake.3 cake food psy_object thing actee-trsact

 

trsact eat ate.2)

(6—16)(cake.3 cake has-part1 frosting food phy_object

thing actee-trsact trsact eat ate.2)

(6-17)(cake.3 cake food eatee-eat eat ate.2)

 

The H-heuristic is applied to the two three-way collisions. Because actee-trsact
is ancestor of eatee-eat, the paths for the three-way collision at actee-act, i.e.
(6-15) and (6-16), are removed. The output of the stage of the H-heuristic is only one
three-way collision and this collision corresponds to only one path, (6-17). Thus the best
path is determined to be (6-17). The following stages of the heuristics need not be
applied because only one path is left.

Let’s consider the other situation during the analysis of the sentence (61). After
the PP, "with a fork", has been analyzed, two rules(vp-pp and np-pp) are found by the
pattern matcher. Each of these two rules creates a branch of parsing. The branch for vp-
pp issues the suggestion (ate . 2 fork . 4 with) and the branch for np-pp issues
the suggestion (cake . 3 fork . 3 with) . The best paths for these two suggestions
are found in a similar way. For these two collisions, the data will be provided using the

bigger knowledge base below.

So far we used the KB shown in Figure 6.5. But this KB is very small. The follow-
ing data was obtained using a bigger KB which has about 100 concepts. The rules and

their suggestions are from the analyses of the same sentence (6-1).

0 Suggestion (man . 2 ate . 2 subj) for the main-verb2 rule :
Number of paths corresponding to two-way collisions : 295
Number of three-way collisions after secondary marker passing : 2
Number of three-way collisions after applying H-heuristic : 1
Best path found:

(618) (<abso ate.2> I <abso eat> C <asp eat.eater-eat> A <rel eater-eat> A <asp

127

eater.eater-eat> C <abso animal> D <abso high_animate> D <abso person> D

<abso man> I <abso man. 1>)

Note in the above data that the number of paths for the two-way collisions has been

increased from 6 to 295 according to the change of the KB. But the result of secondary

marker passing is still 2. This shows the good performance of the proposed mechanism.

This good performance is also observed in the case of the other suggestions shown

below. The L-heuristic and the S—heuristic need not even be applied because only one

path is found after applying the H-heuristic. (According to the experiments, it has been

found that the L-heuristic and the S-heuristic are usually used when there is word sense

ambiguity.)

Suggestion (ate . 2 cake . 2 ob j) for the object rule :

Number of paths corresponding to two-way collisions : 220

Number of three-way collisions after secondary marker passing : 2

Number of three-way collisions after applying H-heuristic : 1

Best path found:

(6—19) (<abso cake.3> I <abso cake> D <abso food> C <asp eatee.eatee-eat> A

<rel eatee-eat> A <asp eat.eatee-eat> C <abso eat> I <abso ate.2>)

Suggestion (ate . 2 fork . 4 with) for the rule vp-pp:

Number of paths corresponding to two-way collisions : 112

Number of three-way collisions after secondary marker passing : 2

Number of three-way collisions after applying H-heuristic : 1

Best path found:

(6-20) (<abso fork.4> I <abso fork> D <abso eat_tool> C <asp instr.instr-eat> A

<rel instr-eat> A <asp eat.insu'-eat> C <abso eat> I <abso ate.2>)

128

° Suggestion (cake . 3 fork . 4 with) for the rule np-pp :

Number of paths corresponding to two-way collisions : 808

Number of three-way collisions after secondary marker passing : 1

Best path found:

(621) (<abso fork.4> I <abso fork> D <abso eat_tool> D <abso inanimate> D
<abso phy_object> C <asp part.has-part> A <re1 has-part> A <asp whole.has-part>

C <abso phy_object> D <abso food> D <abso cake> I <abso cake.3>)

In the above data, it is shown how effective the secondary marker passing mechanism is
for the path ﬁnding problem. This can be contrasted with the case where secondary
marker passing can not be used. When "the granite rocks" is analyzed as a nominal com-
pound, a suggestion (granite . 1 rocks . 2) is sent to SEM. Note that this sugges-
tion has no preposition in the suggestion because there is no syntactic cue about the rela-
tion between the nouns in the nominal compound. The semantic relation between nouns
should be inferred by the semantic processing component (Finin, 1980). Data from the
path ﬁnding operation for the suggestion (granite . 1 rocks . 2) of the noun-

parSGS rule is:

Number of paths corresponding to two-way collisions = 309
Number of paths after applying S-heuristic : 29
Best path found:

(6-22) (<abso rocks.2> I <abso stone> D <abso granite> I <abso granite. 1>)

Secondary marker passing could not be used because there is no preposition available.
After the stage of the S—heuristic, the number of paths is 29 which is still a large number
of paths. The time taken for the stage of the S-heuristic in this case is about 50 times
longer than the time taken for the whole path ﬁnding operation for the suggestions with
prepositions. This data shows how efﬁcient the secondary marker passing mechanism is

for the path ﬁnding problem.

129

6.3 Semantic Comparison of Branches

It has been explained in this chapter that the branches running in parallel are
ﬁltered at the semantic comparison stage(SCS). The comparison is made based upon the
goodness of semantic processing results. The semantic processing results are represented
by the KB paths returned by SEM for the suggestions from SYN. The KB paths are
stored in the corresponding branches so that they are used in the semantic comparison
stage. Thus the problem of comparing the branches based on the semantic processing

results can be reduced to the problem of comparing the paths of the branches.

bran t - s
continue>
j Semantic

Comparison
N backup

tack

 

 

 

    
  

Stage r -

 

 

 

worst

 

Figure 6.17 Rank-ordering the Branches

For the moment, let’s assume that any two paths can be compared and the decision can
be made about which path is the better path (or equally good path). Then the branches
that are input into the semantic comparison stage can be rank-ordered. The branches
with the highest rank are kept as active (see Figure 6.17) and the other branches are
pushed into the backup stack, the worse one ﬁrst so that the branch with better rank can
be used ﬁrst in case of backtracking. Only one branch will normally be given the highest
rank. But there can be more than one branch which receives the highest rank because
two branches may have the same paths as a result of semantic processing. The branches
with no semantic processing result are always inserted in the set of branches having the
. highest rank for the following reason: these branches can not be compared to the

branches with semantic processing results and should be kept as active branches. Having

130

no semantic processing result means that the corresponding rule has no action issuing a

suggestion. Examples of these cases will be shown later.

One simple example can be given using the analysis of sentence (6-1). It was
explained that the np-pp rule and the vp-pp rule were found during the same pattern
matching operation. Two branches are created for the two rules. For the branch of the
np-pp rule, the semantic processing result is the path (6-21) that is found for the sugges—
tion (cake . 3 fork . 4 with) . The path (6-20) found for the suggestion (ate . 2
fork . 4 with) is the semantic processing result for the branch of the vp-pp rule. Fig-

ure 6.18 illustrates this case.

 

 

path=(6-20)

 

 

NP-PP :
path=(6-21)

 

 

 

Figure 6.18 Comparison of Two Rules(vp-pp & np-pp)

It has been stated that rank-ordering the branches during the semantic comparison
stage can be done if any two paths can be compared. We have seen in section 6.2.3 that
the process of ﬁnding the best path for a suggestion also requires the capability of com-
paring two paths and the paths to be compared have the same origins as illustrated in
Figure 6.16. But it is necessary to consider one more possibility in the case of the path

comparison problem in this section. This is illustrated in Figure 6.19.

 
 

path-b

instancel instance2 instancel instance3 instance2
(a) (b)
Figure 6.19 Instances and Two Paths to be Compared

131

In Figure 6.19(b), path-a and path-b share one end (i.e. instance2) but not the other

end. The reason that two paths should share at least one end(origin) is that the newest

constituent built by the parser is always involved in the suggestions issued by the rules

that are found during the same pattern matching. This fact makes it easy to decide which

end of a path corresponds to which end of the other path.

The comparison of path-a and path-b can be classiﬁed into four possible cases (note

that al and a2 are the end absolutes of path-a, and b1 and b2 are the end absolutes of

path-b) as shown in the following heuristic.

SCS Heuristic:

(case-i):

(case-ii):

(case-iii):

The relation of end absolutes conforms to the condition of the S-
heuristic explained in section 6.2.3.3. Let’s state it here again. If a1 is
an ancestor of b1 and a2 is an ancestor of(or same with) b2, then path-b
is better than path-a. Symmetrically, if a2 is an ancestor of b2 and a1 is

an ancestor of(or same with) b1, then path-b is better than path-a.

If a1 is an ancestor of b1 but a2 and b2 have no ancestor relationship,
then consider the following three possibilities, (a), (b) and (c). Here we
need to consider what the term promiscuous concept is. A concept is
said to be promiscuous if it is high in the hierarchy of the KB and has

many descendants(Charniak, 1983a). For example, "action , thing",

"physical-object", etc. are promiscuous concepts.
(a) If a2 is promiscuous, then path-b is better than path-a.

(b) If neither a2 nor b2 is promiscuous, then path-b is better than
path-a.
(c) If b2 is promiscuous but not a2, then it is undecidable.

If a1 is the same as b1 and both a2 and b2 are promiscuous concepts(or

symmetrically, a2 is the same with b2 and both a1 and bl are

promiscuous), path-a and path-b ties in goodness. In this case, any path
with a smaller number of relation nodes is better than the other. If the
number of relation nodes can not even break the tie, other information

such as the frequency of usage of relation nodes in the path is used to

132

decide which path is better.

(case-iv): If a1 is not b1 and they do not have an ancestor relationship, and a2 and

b2 do not have an ancestor relationship, then any path having a smaller

number of promiscuous end absolutes is better than the other.

In Figure 6.18, the path (6-20) for the vp-pp rule was chosen over the path (6-21),
which is determined using (case-ii) of the SCS heuristic. The end absolutes of (6—20) are
eat and eat_tool, and those of (621) are phy_object and phy_object.
Consider Figure 6.20 for this problem. Because phy_object is an ancestor of
eat_tool but phy_object and eat have no ancestor relationship, (case-ii) can

be applied and the decision that (6-20) is better than (6-21) can be made. Thus the

branch for the vp-pp rule is chosen in the SCS.

 

 

phy_object

 

 

,4
/D*

/p

 

 

 

cake

 

 

(6—21)

 

 

eat

 

 

l

 

I

(6—20)

Figure 6.20 Two Paths in "eat" Example

6.4 Semantic Processing Component in Action

In this section, two examples of semantic processing will be given to provide a

"clearer picture of the actual working of the SYN SEM parser. Especially, in the second

 

[phy_object

l

 

 

 

 

 

rk.4

eat_tool

133

subsection, it will be shown in detail how the parallel running of branches is inﬂuenced

by semantic processing.

6.4.1 Example 1

Let’s ﬁrst consider the example given in Section 4.6 which used the example sen-

tence (4-21) shown below again:
The teachers taught by the pool passed the test.

It was explained that three rules are selected during the pattern matching operation as
shown in Figure 4.19: main-verb2, np-vpp, and np-vppl. For each of these rules, a
branch is created. During the execution of the actions of each branch, a suggestion is
issued as shown in Table 6.1. The data related to ﬁnding the best path for each sugges-

tion is shown in the table, too.

After executing the rules of the branches, the semantic comparison stage comes
because each branch needs a new input word. The semantic processing result of the
branch corresponding to the rule main-verb2 is (6—23). Similarly, the path (624) is the
semantic processing result of np-vpp, and (6—25) is for the branch of np-vpp1. Thus
these three paths are compared at the semantic comparison stage. Note that the end abso-
lutes for (6-23) are teach and teach_professional, those for (6-24) are
trsact and thing, and those for (6-25) are teach and person. Each pair of

paths is compared as follows:

(i) Comparison of (6-23) and (6—24):
The ﬁrst EA of (6-24) is an ancestor of the ﬁrst EA of (6-23) and the second
EA of (6-24) is also an ancestor of the second EA of (6-23). Therefore,
according to (case-i) of the SCS heuristic, (6-23) is better than (6-24).

(ii) Comparison of (6-23) and (6—25):
The ﬁrst EA of (6-23) is the same as the ﬁrst EA of (6-25) and the second EA

134

Table 6.1 Data for Finding Paths

 

 

 

 

 

 

 

Branch I main-verb2 "2'22 “E'QEI I
Suggestion (teachersJ taught2 subj) (teachers.l taughtz obj) (teachers.l taught.2 to)
# of 2-way collisions 398 398 398

# of 3-way collisions 2 l l
H-heuristic 1 1 1
S-heuristic l l 1
L-heuristic 1 1 1

path (623) (6-24) (6-25)

 

 

 

 

 

 

(6-23)(<abso taught.2> I <abso teach> C <asp teach.teacher-teach> A <rel teacher-
teach> A <asp teacher.teacher-teach> P <abso teach_professional> D <abso
teacher> I <abso teachers. 1>)

(6-24)(<abso taught.2> I <abso teach> D <abso trsact> C <asp trsactactee-trsact) A
<rel acme-trsact> A <asp actee.actee-trsact> C <abso thing> D <abso phy_object>
D <abso animate> D <abso animal> D <abso high_animate> D <abso person> D
<abso teach_professional> D <abso teacher> I <abso teachers. 1>)

(6-25)(<abso taught.2> I <abso teach> C <asp teach.rec-teach> A <rel rec-teach> A
<asp rec.rec-teach> C <abso person> D <abso teach_professional> D <abso
teacher> I <abso teachers. 1>)

of (6—25) is an ancestor of the second EA of (6-23). Thus, according to (case-
i) of the SCS heuristic, (6-23) is better than (625).

(iii) Comparison of (6-24) and (6-25):
The ﬁrst EA of (6-24) is an ancestor of the ﬁrst EA of (6-25), and the second
EA of (6-24) is also an ancestor of the second EA of (625). Thus, according
to (case-i) of the SCS heuristic, (6-25) is better than (6-24).

From (i), (ii) and (iii), it is concluded that (623) is the best, (6-25) is the next best, and
(6—24) is the worst. Therefore the branch for main-verb2 is the one that is selected by
the semantic comparison stage. The other two branches are pushed into the backup
stack. The branch for np-vpp1 is pushed later than that for np-vpp so that the branch
with the better semantic result is used ﬁrst when backtrack is required later. This situa-

tion is shown in Figure 6.21.

135

 

 

 

main-verb2

 

v
S”
.0
S”

nP'VPP

 

 

ll

 

 

nP-VPPI
Figure 6.21 Competition of Three Branches

 

Consider the situation shown in Figure 4.18(w). It was explained that the pattern
matcher found the passive-by rule and the vp-pp rule that matched Figure 4.18(w).
Thus, for each of the two rules, a branch is created. These branches, the corresponding

suggestions, and the paths for the suggestions are shown in Figure 6.22.

 

 

 

 

 

 

VP-PP (taught.2 pool.3 by):
path=(6-26)

Figure 6.22 Removal of a Branch with Bad Semantic Processing

 

(6-26)(pool.3 I pool D place C loc.loc-teach A loc-teach A teach.loc-teach C teach
I taught.2)

Note that no path can be found for the suggestion for the passive-by rule. Therefore
the branch corresponding to this rule is just thrown away because the semantic effect is
too bad to ﬁnd a reasonable path. But for the suggestion for the vp-pp rule, the path (6-
26) is found. Thus the branch for the vp-pp rule is the only rule that is left and it is kept

active without any comparison as shown in Figure 6.22.

136

6.4.2 Example 2
Let’s consider how sentence (6-27) is processed in SYN SEM.
(6-27)The granite rocks near the shore.

The snapshot of the SWM after reading "rocks" is shown in Figure 6.23. The root of the

tree that corresponds

 

CLIST
)C IV V__N
)A | CASH=nil
)3 granite rocks
Det
the

 

Figure 6.23 Snapshot of CLIST

to the word "rocks" is of two types(V or N), because "rocks" can be either a verb or a
noun. Three rules match this SWM. They are noun-parseS, noun-parse4 and noun-
parseS. Nounéparse4 analyzes "rocks" as the head noun of the NP that corresponds to
the input suing "the granite rocks". This rule matches the SWM because "rocks" is a
plural noun. If it is not a plural noun, it is necessary to see the next word to make the
decision. The rule noun-parse?) regards "rocks" as a verb and analyzes "the granite" as
the complete NP. noun-parse5 also analyzes "the granite" as the complete NP but inter—
prets "rocks" as a noun that leads a relative clause (i.e. "rocks" is considered to be the

NP which is the subject of the relative clause that will follow).

Each of the three rules creates a branch of parsing. The system takes a note that
CASH was empty at this time. From this point to the next semantic comparison stage,
each of the branches will run independently. Note that the semantic comparison stage
comes when all branches want to read the next word (just after "rocks"). Let’s follow
each branch until the semantic comparison stage comes. First, the ﬂow of the branch for
‘ noun-parse3 is shown in Figure 6.24. After the execution of noun-parse3, the SWM

becomes (a) in the ﬁgure. But there is no suggestion issued by this rule. The V of

137

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

m rsr n r. T
S
. WK 1’
11311116. 3 15C N Rule: no rule
oun- arse '
Figure 623 p > NA S—“QPEL K . :
granite mput to
N3 the CLIST
Det '
(the) gram”
CASH Y CAf Y
rocks (a)
CLIST "’
V
§ | Rule: A
P marn-verb2 NP AUX )(P no rule ._
rocks suggestion read S.C.S
(granite.l rocks.2 subj Y next word
path=(6-28) rocks
CASH=ni1 CASH=nil
(C) ((1)

Figure 6.24 Flow of Branch for Noun-parse3

"roc " is pushed into CASH as shown in (a). The subject rule matches this SWM and
its execution results in the SWM in (b). No rule matches (b) and CLIST needs an input.
Because CASH is not empty, the V of "rocks" in CASH is inserted into CLIST as shown
in (c). The main-verbZ rule matches this SWM, and its execution updates (0) to (d).
This main-verb2 rule issues a suggestion (granite. 1 rocks.2 subj). The

best path for this suggestion is found by SEM and is shown in (6-28).
(6-28)(<abso rocks.2> I <abso shake> C <asp act.stuff-shake> A <rel stuff-shake>
A <asp stuff.stuff-shake> C <abso phy_object> D <abso inanimate> D <abso
stone> D <abso granite> I <abso granite.l>)
Against the SWM in ((1), no rule matches and a new input is required to be inserted into

CLIST. Because CASH is empty, 3 new word needs to be read from the input suing

4 which means that this branch has reached the semantic comparison stage.

138

Second, the ﬂow of the branch for the rule noun-parse4 is shown in Figure 6.25.
After the execution of noun-parse4, the SWM of this branch becomes (a) in the ﬁgure.
The rule noun-parse33 matches (a) whose execution results in the SWM in (b). This
rule notices that "rocks" is in plural form of the noun and thus it detemrines that "rocks"
is the head noun of the NP. Because this NP is a nominal compound, a suggestion

(granite . 1 rocks . 2) is sent to the SEM. Note that a preposition is missing in

this suggestion.

 

 

    

 

 

 

 

CLIST
5 NR I‘ rule:
)1 A If nou'n-parse33 ‘
o . . T991“ suggestion
suggestion granite (granite. 1 rocks. 2)

Det path: 29

_(the)

CASH=nil (a)

CLIST m IQT

 

 

S /NK]1\J rule: /S
“II-DEC subject > /NP\ I n9 rule ,

 

 

 

 

 

 

 

rocks no & N S.C.s.
su estion
granite gg rocks
gmncite
CASH=nil CASH=nil
(b) (C)

Figure 6.25 Flow of Branch for Noun-parse4
From the nominal compound, no information about the related preposition can be
obtained by SYN. SEM ﬁnds the path (6-29) for this suggestion.
(6-29)(<abso rocks.2> I <abso stone> D <abso granite> I <abso granite.l>)
The parsing of this branch continues and the subject rule matches the SWM in (b). The
execution of this rule results in the SWM shown in (c). No rule is found for (c). Because

CASH is empty, a new input word should be read from the input string. Therefore this

‘ branch has reached the semantic comparison stage.

139

Third, let’s follow the branch for the noun-parseS rule. The ﬂow of this branch is

shown in Figure 6.26. After the execution of

CLIST FLIL
S K / S

 

    
 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

NC N rule:
i 6— “ . subject ; no rule _
gure noun-parse5 VA granite no & 7
)1 S suggestion the
1112? granite
E N
(a) rnclrc (b)
- 131* - .rsr
rule:
/S I] noun-parseO & /S R
rocks noun-parse33 = P NC N no rule ;
no ' s.c.s.
the, suggestion the rocks
m 31611.19
CASH=ml CASH=nil
(c) ((1)

Figure 6.26 Flow of Branch for Noun-parseS

noun-parseS, the SWM of this branch becomes that in (a) of the ﬁgure. This rule
regards "rocks" as the plural noun which starts a new NP and thus "granite" is analyzed
as the head noun of the complete NP as shown in (a). Note that "rocks" is pushed into
CASH so that the NP, "the granite", can be used in any further processing while it is the
rightmost edge in CLIST. The subject rule matches (a) and its execution results in the
SWM in (b). No rule matches (b) and thus a node in CASH is input into CLIST as
shown in (c). The SWM in (c) becomes (d) by the noun-parseO rule and noun-
parse33 rule. These two rules determine that "rocks" is the NP by itself. No rule is
found that matches ((1) and CASH is empty. Thus a new word needs to be input from the
input string. This branch has reached the semantic comparison stage. Note that no rule
. issued a suggestion during the parse of this branch. This is an interesting case since there

is no semantic processing result for this branch.

140

At the semantic comparison stage, these three branches are compared based on the
semantic processing results. The semantic processing result of the branch for noun-
parse3 is the path (628), while the path (629) is the semantic processing result of the
branch for the rule noun-parse4. But the branch for the rule noun-parseS has no

semantic processing result. This situation is illustrated in Figure 6.27.

 

 

II

nmrn-parseB
path=(6-28)

nonn-partsc4 _ \
path=(6-29)

V
II

 

 

 

._uoun;narsr:5 > __________ >
no semantic result

I!

 

 

 

Figure 6.27 Comparison of Three Branches

The path (628) will be compared with the path (6-29) to compare the two branches.
Note that there is no relation node in the path (6-28). The paths that have no relation

generally have the form shown in Figure 6.28.

v

énstancel éinstanceZ

Figure 6.28 Path without a Relation Node

 

 

 

 

 

 

 

 

 

 

 

 

In this case, we set both the ﬁrst end absolute and the second end absolute to be the pla-
teau (the node A in the ﬁgure). Therefore the end absolutes of (6-29) are the concept
stone. The end absolutes of (6—28) are shake and phy_object. Because the ﬁrst
end absolutes of the two paths have no ancestor relationship but the second end absolute,
. phy_object, of (6-28) is an ancestor of the second end absolute, stone, of (6—29),

(6-29) is better than (628) according to (case ii-b) of the SCS heuristic. Because the

141

branch for noun-parseS has no semantic processing result collected until the semantic
comparison stage, it can not be compared with other branches and thus it is kept active
by the semantic comparison stage. As shown in Figure 6.27, the branch for noun-
parse4 and that for noun-parseS are selected in the semantic comparison stage. The

branch for noun-parse3 is pushed into the backup stack.

After the semantic comparison stage, two branches are active and run in parallel as
shown in Figure 6.29 (branch-1 is for noun-parse4 and branch-2 is for noun-parseS).
The new input word("near") is input by MON and is inserted into CLIST of every active
branch. The SWM at this point is shown in Figure 6.29(a). No rule is found by the pat-
tern matcher in either branch in (a). Thus, a new input word is required to be input in
each branch and the semantic comparison stage comes. But there is no semantic process-
ing result collected after the last semantic comparison stage. So both branches are kept
without comparison. The next input word "the" is read and pushed into CLIST of both
branches as shown in (b). After the node of type "N8" is made, both branches need to
read another new input word. The semantic comparison stage comes but both branches
are kept because neither has any semantic processing result collected since the last stage.
After reading "shore", the SWM’s of the branches are shown in (c). After the PP pro-
cessing is done, branches are shown in ((1). Against (d) in the ﬁgure, the np-pp rule
matches the SWM’s of both branches. The semantic suggestion is (rocks . 2

shore . 3 near) and the best path is found to be (6-30):

(6-30)(<abso shore.3> I <abso shore> D <abso place> C <asp pos.loc-place> A
<rel loo-place> A <asp pre.loc-place> C <abso phy_object> D <abso inani-
mate> D <abso stone> I <abso rocks.2>)

After the punctuation mark(".") in CASH is input into CLIST’s, the branches become
(f). Neither branch ﬁnds any rule that matches the SWM and it is found that both

. branches have reached the dead-end situation. Thus, both branches are thrown away for

the reason of syntactic anomaly. Backtracking is required because there is no active

142

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

PTfP DFt _ pr p NS
near the no 11.119; 4; f T I __,,
r "w l“
S the
nil shore
I ................. tC - - _. .CASHEnil .......
PTTP Bled S $ NP prep 3
P l
rocks near the > I, A near Det I»
[11 I shore
C , the
granite
CASH=nil — CASH=nil T CASH=nil
(a) (b) (C)
branch—1
E . punc
i ar 1:11;)? )L no rule 7 no mtifr
e ._ ow
the s ore (rocks.2 shore.3 C III away
I'OC S near) A0616 A
CASH=() ""1
.......................... .CASHEC).--“_---CASI-L=nil-_-.._---
branch 2
P P P P punc
f; A rule: /5 A f A no rule
NProcks np-pp : g I“ P m NP _,throw
near (rocks.2 shore.3 rocks A , away
the . 6 near) the . I
grarute S OTC granite
CASH=(.) CASH=(.) CASH=nj1
(d) (e) (0

Figure 6.29 Parallel Running of Two Branches

branch. A branch is popped out from the backup stack and it is made to be an active
branch. This is the branch that was shown in Figure 6.24(d). From this point, the parsing
goes on as shown in Figure 6.30. Against the SWM in (b) of the ﬁgure, the vp-pp rule
matches. This rule issues a suggestion (rocks.2 shore.3 near). The path
. found for this suggestion is (6-31). Note that this suggestion is the same as the sugges-

tion issued at Figure 6.29(d)-(e). But the path for the suggestion issued at Figure

143

NP)\ PTP§ A rule:
Z:§ 713nm“ shore ___> NP VP vp-pp K/K IV];
“113' near (rocks.2 shore3

 

 

 

 

 

 

 

 

 

 

 

 

 

build-pp the near) the
granite lks shore granite near
rock's the
CASH=(.) CASH=(.) Shore_
CASH=(.)
(a) (b) (C)

Figure 6.30 Final Branch for the Parse of the Sentence

6.29(d)-(e) is (6—30). The reason that the paths found for the same suggestions are dif-
ferent is that "rocks" is used as a verb in Figure 6.30 (thus with sense shake) while as
a noun in Figure 6.29 (thus with sense stone). This shows that SYNSEM uses the

correct sense that is appropriate for the part of speech used in the analysis.

(6-31)(<abso shore.3> I <abso shore> D <abso place> C <asp loc.loc-act> A <rel loc-

act> A <asp act.loc-act> C <abso act> D <abso shake> I <abso rocks.2>)

6.5 Conclusion

It has been explained in this chapter how the SYNSEM parser gets the semantic
information that is necessary to drive the parsing. Each syntactic rule that needs seman-
tic processing contains actions that issue a suggestion to ﬁnd a best path between two
concepts. A preposition or a pseudo-preposition is provided in the suggestion whenever
it is available. The semantic component should ﬁnd the best path for this suggestion.
This path is used to build semantic interpretation and is stored in the branch as part of its

semantic processing result.

The marker passing paradigm provides a conceptually simple method to ﬁnd a path
between two concepts. But this simple method is plagued by the reality that there are too
' many spurious collisions(paths) collected during marker passing. This results in a seri-

ous problem when attempting to use the marker passing paradigm in natural language

144

processing. A mechanism called secondary marker passing has been developed to solve
this problem. It has been found that the secondary marker passing method removes most
of the spurious collisions efﬁciently. The key idea behind the secondary marker passing
method is to utilize the information that can be provided by the preposition or pseudo-
preposition supplied in the suggestion. Along with this mechanism, three heuristics

have been developed to ﬁnd only the best paths for a suggestion.

Another important task for the semantic processing component is to compare and
ﬁlter the branches according to the semantic processing result at the semantic com-
parison stage. Each branch keeps the set of the knowledge base paths found for the
suggestions issued by the rules of the branch. The problem of comparing the branches is
reduced to comparing the paths. How to compare the two paths to determine the better
one has been studied in this chapter. This is one way to solve the problem of scoring the
alternative rules and choosing the one with the best score. Instead of producing the abso-
lute score for the rules, the branches are rank-ordered according to the comparison of the
paths and the branch with the highest rank is chosen. Using this scheme, the problem of

producing an absolute score can be avoided.

CHAPTER 7

WORD SENSE DISAMBIGUATION

It is recognized that word sense disambiguation is one of the several natural
language understanding problems that are hard to solve completely. Since any natural
language understanding system should have some mechanism to do word sense disam-
biguation to achieve a reasonable degree of understanding, researchers in natural
language understanding have devoted much research to this problem. It seems that we
are far from achieving a complete solution to the problem. However humans seem to
have no double in selecting the correct sense of a word in a sentence. Psycholinguists
have tried to ﬁnd the mechanism that humans use to achieve this high performance, but

they are also far from completely understanding this mechanism.

There are three kinds of ambiguity related to a word (so-called lexical ambiguity) :
homonymy, polysemy and categorical ambiguity. Homonymous words (for example,
ball/toy-ball, ball/dance-party) have multiple meanings(senses) that are not related.
Polysemous words (for example, "go") have several meanings that are related. The
senses of a polysemous word have a meaning in common. A polysemous word is also
called a vague word. A word is categorically ambiguous when it can be used in more

than one syntactic category (for example, rock/shake, rock/stone).

There were two recent research efforts in word sense disambiguation that used
knowledge bases built with frame-based knowledge representation. Hirst’s(1984, 1986)
research is largely related to the disambiguation of homonymous words. He used what
he called Polaroid Words(PW) to realize his idea about the gradual disambiguation of
ambiguous words. A PW is created and assigned to each word as it is input to the

parser. PWs can be considered to be communicating processes that cooperate together.
145

146

A PW ﬁrst starts with all possible senses of its word. The PW gradually eliminates the
senses that are inappropriate to the surrounding context by communicating with other
PWs. When a PW eliminates some of its senses, it announces this. Then other PWs use
this information to try to eliminate some of their senses. This elimination and announce-
ment process goes on until all words have been read and the exchanging of information
stabilizes. Two major sources of disambiguation information that are used by PWs are
selectional restriction and semantic association. The use of semantic association is
achieved by incorporating marker passing in his system.

Lytinen(1984, 1988) put his focus on the disambiguation of vague words. He used
the concept reﬁnement process for this purpose. A vague word is initially represented by
a concept that is high in his hierarchically organized knowledge base. As more words
are read in and more contextual information is available, a set of concept reﬁnement
rules are used to reﬁne the concept to a more speciﬁc concept in the knowledge base.
But it is not clear how the homonymous words (especially nouns) are disambiguated
using Lytinen’s strategy.

In this chapter, we will explain a word sense disambiguation method used in the
SYNSEM parser. Inﬂuenced by the work of Charniak(1983a, 1986), SYNSEM has been
built based on the marker passing mechanism. SYNSEM keeps track of the semantic
paths which reﬂect the process of semantic interpretation of an input sentence. If a word
has multiple senses and it has not been disambiguated yet, there can be false paths being
kept due to the inappropriate senses of the ambiguous word. If a word has a vague word
sense, then it will be represented by an abstract concept until it is disambiguated to a
more speciﬁc concept. As the parsing goes on, these false paths will be eliminated and
abstract concepts will be replaced by concrete concepts. All these disambiguation stra-
tegies adopted by SYNSEM are made possible by the use of the knowledge base that

contains the system’s knowledge about the world.

147

7.1 Use of Semantic Paths for Word Sense Elimination

To explain several aspects of the disambiguation method used in SYNSEM, the
knowledge base shown in Figure 7.1 will be used throughout this chapter.

© subj
1k

     

 

 
 
 
  
 
 
  

semi.

 

like. smart
I

C 0 Q o C
obj-c.function.pz.par@,ar10r_3

Figure 7.1 Knowledge Base for Examples

Let’s use the example sentence (7~1) (Hirst, 1983) to show how the disambiguation goes

on.
(7-1)The crook operated the pizza parlor.

In this section, the use of semantic paths for word sense disambiguation will be
described. The mechanism is based on the word sense elimination strategy proposed by
Hirst. But instead of using the complex interaction among Polaroid Words, we utilize
the semantic paths that are found during the analysis of the sentence. As we saw in the
previous chapter, the best path in the KB is found for each suggestion sent from SYN to

SEM. The best path(or paths) for the suggestion is stored in a set in the branch of

148

parsing. This set represents the semantic interpretation of the sentence (in the loose
sense). Each branch has its own set of paths that represent its semantic interpretation.
Let’s call this set of the best paths the set of semantic paths. When a branch, say branch
A, is forked into two branches, say B and C (because of two rules found by the pattern
matcher), A’s semantic paths are copied to B’s and C’s. Then the set of semantic paths
for branch B is the same as that of branch C. But B and C can add different semantic

paths according to their own processing as the parsing goes on.

If a word has multiple senses out of which only one is the correct one in the sen-
tence, there will be wrong paths found by SEM because of the wrong senses. As the
parsing goes on and more context is available, some paths will be proved to be wrong
and be removed. The word senses are also eliminated if they are in the paths that are
proved to be false later. The elimination of a word sense is propagated via semantic
paths, which results in the removal of other word senses. Using this idea, we are able to
characterize the word sense elimination process more clearly and implement it more
efﬁciently. How this process works will be explained in detail using the parsing of sen-

tence (7-1).

When "croo " is input, MON will create the instance node crook . 1. "croo " is a
homonymous word that has two senses: criminal and a staff used by the shepherds. The
node crook . 1 has I-links pointing to all senses of the word as shown in Figure 7.2.

Each sense is represented by an absolute node in the KB.

    

Disambiguation
Process

Figure 7.2. Representation of Multiple Word Senses

This means that SYNSEM supports the all-readings hypothesis for the word sense

access (Swinney, 1979). To each I-link, a status ﬂag is attached to indicate the status of

149

the link (the status of an I-link is written in the parentheses in Figure 7.2). A status ﬂag
can be one of the following values: "i", "u", and "r". "i" means that the link is in the ini-
tial state; In other words, it has not been processed yet since its creation. "u" indicates
that the I-link is currently being used in a semantic path and the corresponding sense is
still a candidate for the correct sense of the word. An I-link for the sense that has been
eliminated is tagged with "r"(removed). When a word has been disambiguated fully,
there should be only one I-link with the status ﬂag "u" as illusuated in the left side of

Figure 7.2.

Here, it is necessary to inu'oduce a complicated matter that SYNSEM has to han-
dle. As explained before, each branch of parsing has its own set of semantic paths.
Therefore the word sense disambiguation status of a word may be different among the
branches, which means that each branch should have its own representation of the
disambiguation status of the words. This problem becomes more complicated with
marker passing and semantic priming (that will be explained later). For example, con-
sider Figure 7.3. Note the possible connections between 2 and the instance for the new
word. If branch-i is selected instead of branch-ii, then the sense of the new word should
be the concept "V" in the ﬁgure. But if branch-ii is chosen instead of branch-i, then the

concept "U" should be the correct sense of the new word.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

. /——/I— \

branch-r

x Y

> )6
Z U / v

branch-ii N I I

x /Y W

new word
I (u)

 

 

 

Figure 7.3 Different Word Senses for Different Branches

150

For the input word "croo ", the MP passes markers starting from its instance node.

The next word "operat is input which is represented by the instance node
operated . 2. This word has two possible senses: to cause something to function and
to perform surgery. Marker passing is done for this word, too. After syntactic analysis,
SYN will send a suggestion (crook . 1 operated. 2 subj) to SEM according to
the rule main-verb2. SEM will ﬁnd all possible best paths as shown in Figure 7.4. There
should be only one best path for a suggestion, but if one of the words related to the

suggestion has multiple senses as in this example, more than one best path can be found

by SEM. It is important for SEM not to remove any legitimate good path.

 

 

 

 

 

o ‘I
instr-c. unction \
0 0 ' C ' I I
agent-c. unction Sm — cause functionL _S_1_,

I

. u) v’v”’
ag —. orm_surg w) ’ I(r)

Figure 7.4 Paths between Two Instance Nodes

 

Three paths have been found in this example as in Figure 7.4. These paths are stored in
the set of semantic paths. Note that shepherd_staf f has no path to
perform_surgery because the KB in ﬁgure 7.1 deﬁnes that the instrument of per-
forming surgery should be some kind of medical tool and the shepherd’s staff is not one
of them. At this point of paring, the two senses of "crook" are still used in the semantic
paths. Thus, it is not fully disambiguated yet. The word "operated" has not been disam-
biguated either. If we assume that the word "operated" has another sense 81 (the dot-
ted box in the ﬁgure) and it is not used in any path found, then the sense 81 will be
removed by attaching the ﬂag "r" to the corresponding I-link. For a suggestion, usually
one path is found. But three paths have been found because of the ambiguous words in

this case. As the parsing goes on, the two false paths will be removed.

151

Now the next input is the canned phrase "pizza parlor" which is represented by the
single concept pizza _parlor. Let’s assume a canned phrase can be analyzed.
, After marker passing is done for this word, SYN will send the suggestion
(operated.2 pizza_parlor.3 obj) to SEM by the object rule. For this

suggestion, there is only one path (d) found as shown in Figure 7.5.

, © , - .
Kim ‘ iris ©gss"
S v d. : . © _
W h “‘ ' ent—c.function(c m, "- ca ﬁp ction pizza 53—1) 10f

© 1 u)
agent-perform_surg o .- . -. .2 zza_p or.3

Figure 7.5 Effect of Reading a New Word

 

 

 

 

 

 

 

  

 

I( (u)

Note that the sense perform_surgery of "operated" is not used in this path. This
means that perform_surgery cannot be the sense of the word "operated". The

II I,"

sense perform _surgery is removed by attaching the ﬂagr to the corresponding
I-link. The removal of the sense perform_surgery makes SEM remove the path
(c)(in the ﬁgure) from the set of semantic paths, which results in the set of semantic

paths shown in Figure 7.6.

© .
instr-c.function , (' -c© unc on
s r'. -. . © _ $12me or
$5 _ Mal agent-c.function m. - ca ﬂp ction piz —P 101'

I (
(u) H. , ‘3 @or3

Figure 7.6 Paths after Removing a Sense and a Path

     
 

 

 

 

At this point of parsing, the word' 'operated" has only one sense with "u " ﬂag, which
means that it has been fully disambiguated. But there are still two senses with " u" ﬂag

for the word "crook". Thus, it has still word sense ambiguity.

152

If there were not path (b) between crook . 1 and operated . 2, then the remo-
val of the path (c) would result in the removal of the sense criminal of the word
"croo " because criminal would not be used in any path connecting crook . 1 and
operated . 2. As soon as a sense gets an "r" ﬂag, it will never be considered by SEM

during the path-ﬁnding process.

At this point, let us consider a general case which can be illustrated in Figure 7.7.
W.8, Y.l and X.4 are instance nodes for the words already processed by the parser. W.8
has two senses which are P and Q and these two senses are still candidate senses. A and
B are candidate senses for Y.1. D and E are candidate senses for X.4. At this point, the
new word 26 has just been input. The suggestion between Z6 and X.4 resulted in only
one path (e) as shown in Figure 7.7. The sense D of X.4 is not used in this path (e).
Thus, the sense D can not be the sense of X.4 and it is removed. The removal of the
sense D causes the path ((1) to be removed from the set of semantic paths. The removal
of ((1) causes the sense B to be removed from the possible senses of Y.1. The removal of
the sense B again causes the removal of the path (a) from the set of semantic paths. This
again causes the removal of the sense P from the possible senses of W.8. P is not con-
nected to any other paths and thus the removal operation stops here. From this process,
we can notice that the removal of a word sense recursively propagates through the KB
via the semantic paths. We call this operation the recursive word sense removal(RWSR)

operation.

 

 

 

 

 

   

Figure 7.7. RWSR Operation

But there is a case which requires more careful consideration. Consider the case

shown in Figure 7.8 which has one more path, (f), added to the case in Figure 7.7.

153

 

 

 

 

 

Figure 7.8 Consideration for the Set of Non-removable Senses

During the RWSR operation invoked by the removal of D of X.4, P will be considered
for removal after the removal of the path (a). Note that P is also used in the path (f).
Thus, P should not be removed. The sense P is an example of the so-called non-
removable senses. After the path ﬁnding process for a suggestion has been done, the
SEM computes the set of non-removable senses before the word sense disambiguation
process starts. SEM computes the set of non-removable senses as follows: Starting from
the word senses of the new word that are included in any semantic path found for the
suggestion, ﬁnd all word senses that can be reached via the semantic paths. In Figure
7.8, this set will be (F E A Q P). During the RWSR operation, any word sense that is in

the set of non-removable senses is not removed.

In this section, we have been considering the word sense disambiguation process
that is initiated by each suggestion sent by SYN. This operation starts from the set of
semantic paths that is found by SEM for each suggestion from SYN. Let’s assume that
S is the set of semantic paths found for a suggestion having old-instance and
new-instance as two origins. Then the word sense disambiguation is done as fol-

lows:

(1) Remove the senses of new-instance that are not in any of the paths in

S(for example, the node C in Figure 7.9).

(2) Compute the set of non-removable senses. (In Figure 7.9, all senses that are

reachable from A or B via semantic paths are non-removable senses.)

154

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

  

' tan
rns ce S=(path—a path—b)

Figure 7.9 Starting the Removal Operation

(3) For each of the senses of old-instance having "u" ﬂag, if the sense is not
used in any path in S, call the RWSR operation passing this sense and old—

instance as the arguments.

RWSR operation routine is shown in Figure 7.10.

 

Procedure RWSR_operation (sense instance)
begin
if sense is in non-removable-senses
then exit
else
begin
remove sense form the possible senses of instance;
find all (sense instance) pairs that can be reached from
(instance sense) pair via any path in the set of semantic
paths and has not been visited yet. Insert the pairs into
the set "reachable";
remove any path used in the step 1 from the set of semantic paths;
call RWSRgoperation recursively for each element in the set
"reachable" passing the element(which is pair) as argument;
end;

end;

Figure 7.10 RWSR Operation Routine

 

155

7.2 Use of Concretion

Another mechanism used by SYNSEM for word sense disambiguation is the con-
cretion operation. A sense of a word is changed to a more specialized sense if a path
with a special shape connecting the two instances of the suggestion are detected. Let’s
call this special shape the concretion shape. The concretion shape is shown in Figure

7.11.

 

 

 

 
  

 

 

 

 

 

 

 

, ’ ’ instance2

 

 

 

 

I
I
I

 

 

 

 

 

 

 

 

 

 

 

instancel

Figure 7.11 Shape of Concretion Path

In Figure 7.11, the path (instance2 A B D E F G H instancel) is of the
concretion shape. Note that there can be zero or more absolutes between G and
inst ancel and between A and B in the ﬁgure (as indicated with *). According to
this concretion path, the sense of instance2 can be specialized from A to B. The
reason for this is as follows: The conceptual association between instancel and
instance2 is via B and A; But B is a specialization of A; Thus, instance2 can
actually imply the meaning B (which is a specialization of A). The concretion path in
Figure 7.11 is the best path found for the suggestion (instancel instance2
Prep) -

The detection of the concretion path is done in a special way. Note that markers
from inst ancel and inst ance2 can not create a two-way collision at the relation
- node E because a marker from inst ance2 can not ﬂow in the reverse direction of D

links and thus can not reach the node B. But it is necessary to get a two-way collision at

156

B so that a three-way collision can occur at E: during secondary marker passing and thus
the path can be found as the best path for the suggestion. This problem is handled in the
following way. Note that the marker starting from instancel will reach node A via
the course of the concretion path (i.e. via H, G, F, E, D, and B). The two-way
collision occurs at the node A when a marker ﬁom inst ance2 reaches the node A.
For every two-way collision, the MP checks if the trace of one marker of the collision is
of the concretion shape and the trace for another marker of the collision is of the form
(collision-node I-link instance-node). Then a dummy two-way collision is put on the
relation node which is in the middle of the trace (the node E in Figure 7.11). We add
another case in which a three—way collision can occur in the case explained in Chapter 6.
The dummy collision node is also used to form a three-way collision when it collides
with a secondary marker. This is the way that the concretion path is found for a sugges-
tion. When a concretion path is found for a suggestion, the sense of an instance is spe-

cialized. In the current example, the sense of instance2 is changed from A to B.

Let’s look at the ongoing example that is using (7-1). In the previous section, it
was explained that the path ((1) in Figure 7.6 was found for the suggestion
(operated. 2 pizza_parlor . 3 obj). Actually this path is a path of the con-

cretion shape as shown in Figure 7.12.

obj cau ction

© 1‘1 ,1

ll , , ’operated.2

 

 

 

l
l
V D ,”I

77

cause funptionn
przza:par or

 

 

 

 

 

obj-c.function_pizza_parlor

 

pizza_parlor.3

Figure 7.12 Concretion for the Example

Noticing that (d) in Figure 7.6 is a concretion path, the sense of "operat " is changed

9 from cause_function to cause_function_pizza_parlor. This is called

157

the concretion operation. But this concretion operation requires the system to adjust the
semantic paths that goes through the old word sense (cause_funct ion in this exam-
ple). For example, the path (a) and (b) in Figure 7.6 goes through cause_function
to get to operated. 2 from crook . 1. Thus these paths should be adjusted so that
they can go through the concreted sense, cause_function_pizza__parlor,
instead of the old sense. This is called the path adjustment operation after concretion. It
is important to recognize that the path adjustment operation after concretion is also a
source of word sense disambiguation. To see this, let’s consider how the path adjust-

ment operation is done in the ongoing example.

The new adjusted path should satisfy the following condition: the relation node of
the new path is the specialization of(or same as) the relation node of the old path and
the end absolutes of the old path is an ancestor of(or same as) the end absolutes of the
new path. If a new path satisfying this condition can be found, the new path is added to
the set of semantic paths in place of the old path. If the new adjusted path can not be

found, the old path is simply thrown away.

The adjustment of the path (b) in Figure 7.6 is shown in Figure 7.13. The new
adjusted path

 
   
  

old path

 
   

116W

   

agent-cause_func.przza_parl cause_fu tion
pizza_par or

crook. 1

Figure 7.13 Success of Path Adjustment

shown in this ﬁgure satisﬁes the condition stated above. In the case of the adjustment

operation for the path (a) in Figure 7.6, the new adjusted path can not be found as

158

illustrated in Figure 7.14. The reason is twofold: shepherd_staf f is not an ances-

torof capital and there is noconnection between capital and crook.1.

 

 

 

 

 

 

 

 

. . instru ent_ .
rnanrmateW se function
old path \
D ’0‘
- x ’1

 

 

 

 

 

 

 

 

 

 

 

f N i

k) 0/ i f .
inStrument cause_functron_

cau se_func_pz_parlor pizza_parlor-

 

 

 

 

Figure 7.14 Failure of Path Adjustment

Therefore, the path (a) in Figure 7 .6 is removed from the set of semantic paths, which
results in the removal of shepherd_staf f from the sense of crook. 1. At this

point, the word "croo " has only one sense criminal and is fully disambiguated.

The method of concretion used in this section is similar to the concept reﬁnement
process (speciﬁcally the slot-ﬁller specialization rule) of Lytinen(1984). But the path
adjusunent operation after concretion was not used in Lytinen’s method and is a new
source of disambiguation information developed in SYNSEM. The disambiguation
information from the path adjustment operation may lead to the elimination technique
explained in section 7.1 to remove other senses, which means that the different sources

of disambiguation cooperate with each other.

7.3 Use of Semantic Association

Semantic association is one of the sources of information that is useful for word
sense disambiguation. It seems that Quillian(1968) is the ﬁrst researcher who used it.
Semantic association can be modeled naturally in a semantic network. Two concepts
have a semantic association if there is a path connecting them in the network. A concept

is said to be semantically primed by another concept if the access of the latter concept

2

159

facilitates the process(or access) of the former concept. In a marker passing system,
nodes X and Y have semantic association if markers originating from them collide at a
node. If X gets a marker whose origin is Y, it can be said that X is semantically primed
by Y.

SYNSEM uses these notions for the disambiguation. When the MP begins to do
marker passing for a new input word having multiple senses, it checks if there is a sense
of this word that has been primed by a previous word. Then the MP removes the other
senses of the new word by tagging the corresponding I-links with "prime-removed" (this
is an additional possible value of the status ﬂag). The MP does not pass markers from
the senses with the "prime-removed" tag, which implies that these senses are not
accessed. This is to follow the research in Seidenberg & etc.(1982) which found that
semantic priming allows only the primed sense to be used. An ambiguous word which
still has more than one sense remaining can be disambiguated by a new input word if
there is a strong semantic association (detected by the MP) between one sense of the
ambiguous word and one sense of the new word. Strong semantic association is deﬁned
as follows in SYNSEM: there should be a path that does not involve more than two
relation nodes and the end absolutes of the path should not be promiscuous concepts.
The MP removes the senses of the ambiguous words that are not involved in the strong
semantic association by tagging them with "prime-removed". This process can be

explained using the input sentence (Hirst, 1984):
(7-2) The slug operated the vending machine.

We assume that the possible senses of "slug" are: a gashopod without shell, a fake coin,
a bullet or a shot of liquor. After "operat " is processed, the following three paths are

found and registered :

(7-3)(slug.l fake_coin coin inanimate instr-c.function

cause_function operated . 2)

160

(7-4) (slug . l bullet inanimate instr-c . function

cause_function operated.2)

(7-5) (slug . 1 shot_liquor inanimate instr-c . function

cause_function operated.2)

Note that perform_surgery is removed from the possible candidates for the sense
of "opera " because it is not used in the paths. The sense
gastropod_without_shell is not used in any path and thus it is removed from
the possible senses of "slug". But the three senses(fake_coin, bullet, and
shot_liquor) are still possible candidates for the correct sense of "slug". After "the
vending machine" is input, the MP ﬁnds the following path indicating strong semantic
association between slug. 1 and vending_machine . 3 through the sense

fake_coin :

(7-6) (vending_machine.3 vending_machine obj-c.function_vendingmachine
cause_function_vendingmachine instr-c.function_vendingmachine coin

fake_coin slug.1)

Thus fake_coin is chosen as the sense of "slug", and other senses and their
corresponding paths((7-4) and (7-5)) are removed. Word sense disambiguation using
semantic association is graphically shown in Figure 7.15. Note that the removal of the

sense A in the ﬁgure invokes the RWSR operation explained in the above section.

 

 

 

 

NE] W“
I . .
(u) no 1(1) £50)
e e
forprevious for new
word worrL_

 

 

 

 

 

Figure 7.15 Use of Semantic Association

161

7.4 Use of Information about Syntactic Category

Another source of word sense disambiguation is syntactic categories of words
assigned during syntactic analysis. Let’s consider the sentence (7-7) to illustrate how

disambiguation is done using the categorical information.
(7-7)The granite rocks near the shore.

The word "rocks" has two possible senses: stone and shake. But stone is used as the
sense of "rocks" only when the word is interpreted as a noun and shake is used as the
sense only when the word has verb as its part of speech. As soon as the syntactic
analysis component determines that "rocks" is a verb, it can be determined that its word
sense is shake. If it is decided that "rocks" should have the syntactic category of noun,

its sense automatically becomes stone.

This disambiguation method is also used in SYNSEM. When an instance node is
created for the word "rocks", it attaches the part of speech to the I link of each sense as

shown in Figure 7.17. The MP ﬂows markers through all senses of

 

 

 

 

 

 

shake stone shake stone shake stone

I V1 IN] N [N]
1’ 1'
w as»

(b) (C)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Figure 7.17 Use of Categorical Information

the instance because it does not know yet which sense is the correct sense of the word.
Then three rules(noun-parsea, noun-parse4 and noun-parseS) are found by the pat-
tern matcher as shown in Section 6.4.2. Three branches are created for the three rules
and they run in parallel as described in that section. The branch for noun-parseS
analyzes "rocks" as a verb. (Note that each branch keeps its own copy of I links for
ambiguous words.) In this branch, the I link for the sense stone is removed so that it

can never be used during the process of ﬁnding the best path for the suggestion in the

162

branch as shown in Figure 7.17(b). Thus the paths to rocks . 2 can only go through

the sense shake.

The branch for noun-parse4 and the branch for noun-parseS set the I links for
rocks . 2 as shown in Figure 7.17(c) so that the link for the sense shake can never be

used for ﬁnding the paths for their suggestions (because they analyze "rocks" as a noun).

When one branch is ﬁnally chosen over the others, the correct sense of "rocks" is
also automatically chosen which conforms to the syntactic category of the word. In this
example, the branch for noun-parse3 is the one that the parser ﬁnally chooses and thus

the sense shake is automatically chosen as the correct sense of "rocks".

7.5 Conclusion

A new method for applying the word sense elimination mechanism has been pro-
posed. We have shown that a knowledge base encoded with the KODIAK knowledge
representation language can be used to facilitate word sense disambiguation in natural
language parsing. The basic strategy is to make the parser keep track of a set of seman-
tic paths corresponding to the semantic interpretation of the input sentence. These are
the paths that connect the instance concepts of input words. Senses of a word that are
not used in the semantic paths found for a suggestion from the syntactic analysis com-
ponent are removed. The removal of a sense at one end of a registered path results in
the removal of the sense at the other end of that path. The removal operation is applied

recursively.

A concretion mechanism similar to the concept reﬁnement process is also incor-
porated into the disambiguation method in the SYNSEM parser. The idea of the path
adjustment operation is newly employed to extend the concretion mechanism. The use
of marker passing enables the parser to use semantic association and semantic priming
as another source of disambiguation information. The syntactic analysis helps the parser

eliminate word senses that do not correspond to the correct part of speech.

163

A problem occurs if semantic priming leads the parser to select an incorrect word
sense. Further research is required to handle this problem. The fundamental problem
that waits for additional research is how more diverse kinds of knowledge such as goals,
intentions, discourse structures, and more global context (spanning to several sentences)

can be applied to the word sense disambiguation problem.

CHAPTER 8

CONCLUSION

The parsing model with improved interleaved semantic processing developed in the
thesis research will be summarized ﬁrst and then it will be argued that the proposed
approach of parsing is an integrated parsing approach. Then the achievements that have
been made in this thesis will be outlined. Finally the problems that need more research

related to the thesis will be enumerated and discussed.

8.1 Proposed Approach to Parsing

The objective of the research has been to develop a natural language parser that can
utilize both syntax and semantics as early and fully as possible. To achieve this objec—
tive, we have taken the approach of interleaved semantic processing which is based on
the syntax-oriented analyzer using semantics to get intermittent feedback. As suggested
in several parsing models in psycholinguistics which proposed the parallelism, all syn-
tactic alternatives of a structural ambiguity are considered in parallel. Each alternative
invokes a new branch of parsing. The branches run in parallel until all branches need to
read the next input word. At this time, the semantic comparison stage is invoked, during
which the semantic processing results of the branches are compared and only those with
best results are kept alive while the others are thrown away. This is a source of structural

disambiguation that uses semantic information.

The branches out of the semantic comparison stage run again in parallel until the
next semantic comparison stage comes. The semantic results collected during this inter-
val between the two semantic comparison stages are used again for ﬁltering the

branches. The branches run asynchronously in parallel, but they are synchronized

164

165

periodically by the reading of words. Some branches might be removed if they reach a
dead-end because the following input material does not agree with their analyses, which
is realized by the fact that no rules are found during the pattern matching step and the
situation is determined to be a dead-end. The removal of branches here is a source of
structural disambiguation that is based on syntactic information. Another type of ambi-
guity that is abundant in natural languages is word sense ambiguity. Much effort on the
word sense disambiguation problem has also been made in this research. The theoretical
support for the parsing approach taken in this thesis can be found in Kurtzman’s parsing
model that advocated for immediate parallel analysis(IPA) with strong parallelism and

conceptual selection hypothesis.

A parser that had a similar motivation as SYNSEM is Lytinen’s MOPTRANS
parser. But he approached the problem of building a parser with good use of syntax and
semantics from the integrated parsers such as the conceptual analyzers. MOPTRAN S
adopted the control in the sense of Crain and Steedman’s strong interaction between
syntax and semantics, which means that, in the MOPTRAN S parser, the selection of the
syntactic rules is inﬂuenced by semantics. Actually only one rule that corresponds to the
best semantic interpretation is chosen and other alternatives are not even considered.
Lytinen’s approach is directly against most of the parsing models proposed in psycho-
linguistics. In addition to the difference in general approach, SYN SEM and MOP-
TRANS are different in the inside detail of the parsing. It is argued in this thesis that
SYNSEM’s parsing approach constitutes a new distinct approach that utilizes syntax and

semantics as early and fully as possible.

8.2 Argument for the Proposed Approach

It was pointed out that the parsers with interleaved semantic processing are not
good enough in utilizing semantics. We will show that SYNSEM has improved on this

point over the previous parsers. (The exact paths in the KB corresponding to the

166

suggestions will not be provided in this discussion; We will also sometimes use the ter-
minologies such as "slot" of a frame representation language for the ease of explana-
tion.) Let us assume that SYNSEM parses sentence (8-1) and SYN has just determined

Mary to be a NP:
(8-1)John gave Mary a book.

At this point, two rules match the syntactic working memory(SWM): object and ind-
object. The ind-object rule matches the SWM because the verb gave is a dative verb.
SYN sends the suggestion (gave.2 Mary. 3 obj) according to the object rule.
Another suggestion (gave.2 Mary.3 to) is sent to SEM according to the ind-
object rule. Here gave . 2 is the instance node of the concept give, and Mary . 3 is
the instance node of the concept female. The path for the suggestion (gave.2
Mary. 3 to) is better than the path for the suggestion (gave.2 Mary. 3 obj)
because person ﬁts better into the RECIPIENT slot(implied by to) than the
PATIENT slot (implied by obj) of give (more precisely, because
high_animate, the constraint of the relation recipient-give, is a more speciﬁc
concept than phy_object, the consuaint of the relation pat ient-give). Thus
the branch for the ind-object rule is selected over that of the object rule. This shows

that SYNSEM correctly interprets "gave Mary" in (8-1) as an integrated parser does.
Another illustration can be provided by using the sentences from (8-2) to (8-4):
(8—2)The teachers taught by the Berlitz method passed the test.
(8-3) The students taught by the Berlitz method passed the test.
(8-4) The horse raced past the barn fell.

Lytinen(1984) argued that an integrated parser should be able to explain why (8-2) is a
garden path sentence while (8-3) is not, even though they have the same syntactic struc-
tures as (8-4) which is a well known garden path sentence. After SYN SEM inputs

"taught" in (82), three syntactic rules(main-verb2, np-vpp1, and np-vpp2) will match

167

the SWM because "taught" can be either a "past active" or a "past participle". The
main-verbZ rule analyzes the teachers as the subject of the past active "taught". The
np-vpp1 rule analyzes "the teachers" as the direct object, and the np-vpp2 rule analyzes
"the teachers" as the indirect object of the past participle "taught". Each of the three
rules creates a branch. These branches send three suggestions to SEM: (teacher . I
taught . 2 subj) for the main-verb2 rule, (teacher . 1 taught . 2 obj) for the
np-vpp1 rule, and (teacher.1 taught.2 to) for the np-vpp2 rule. For the
suggestion of each branch, the best path connecting the two instance concepts in the
suggestion will be found. Then at the semantic comparison stage, the three paths of the
three branches will be compared and the best one will be selected. The path for the
main-verb2 rule is best because the concept teacher can be a prototype for the
AGENT slot but not for the RECIPIENT or SUBJECT-MATTER slot of the concept
teach. (Note that the "P"(prototype) link is added in addition to the "C"(constraint)
link between an absolute and an aspectual in our version of KODIAK.) Thus, the branch
for the main-verb2 rule is selected which analyzes the teacher as the syntactic subject
and thus a garden path phenomenon occurs in (8-2). But in the case of (8-3), the path
for the suggestion of the np-vpp2 rule (that determines the students as the indirect
object) is best among the three paths for the three suggestions of the three rules because
the concept student is a prototype of the RECIPIENT slot of the concept teach.
Therefore a garden path phenomenon does not occur in sentence (8-3) even if it has the
same syntactic structure as sentence (84). So our parser behaves in the same way that

Lytinen expected for an integrated parser.

It can be shown that the integrated parsing approach of MOPTRANS which per-
forms semantic processing before syntactic processing in each step of parsing is more
inefﬁcient in some cases than the approach of SYNSEM. MOPTRAN S tries to connect
two semantic objects without using the information that syntactic knowledge can pro-

vide. Let’s consider (8-5) to illustrate this point:

168

(8-5)The man attacks a bear with a boy.

After inputing attacks, MOPTRANS will uy to ﬁnd all possible connections between the
semantic objects in memory: MAN and ATTACK. It will ﬁnd the three possible seman-
tic connections: MAN can be the AGENT of the action ATTACK; MAN can be the
PATIENT of the action ATTACK; and ﬁnally MAN can be the PARTICIPANT of the
event ATTACK. Then it will try to ﬁnd the best one among the three.

In the case of SYNSEM, only one syntactic rule, main-verbz, will match the
SWM. Therefore the relation node pat ient-of—attack and the relation node
participant-of-attack will never be considered because these two nodes will
never get marked during the secondary marker passing that starts from the node subj.
The three-way collision will occur only at the node agent-of—attack. SYNSEM
will have only one branch for one rule and never consider the semantic connections that
should be considered by MOPTRANS. In a large and complex knowledge base, there
can be many possible connections between semantic objects which may make the MOP-

TRANS’ parsing approach inefﬁcient.

From these examples, we argue that SYNSEM utilizes syntactic and semantic

knowledge as fully and efﬁciently as any existing integrated parser.

8.3 Observations on Implementation and Test

The SYNSEM parser has been implemented on a SUN 3/280 system using Com-
mon Lisp. Appendix C lists the test sentences used to test the parser developed to
demonstrate the working of the theory explained in this thesis. Appendix D shows the
total output of the parser to parse a sentence, which is useful to track the course of pars-
ing. (The comments are shown in italics. The sentences which the parser could not han-

dle satisfactorily are preceded by an asterisk.)

The size of the parser is reﬂected in the following data:

169

Number of absolutes : 172
Number of relations : 145
Number of aspectuals: 290
Number of syntactic rules : 80

Number of Lisp lines : 7400 (including comment lines)

As far as the speed is concerned, the prototype of the SYNSEM parser is not good. It is

slow, and needs some analysis to ﬁnd out ways to increase the speed. For sentences with

around 10 words, it takes about 3 rrrinutes for parsing. But it would be useful to report

what has been observed during the test:

(a)

(b)

(0)

Most of the parsing time is spent in the steps of doing marker passing for open
words (it has been called primary marker passing in this thesis). From three or
four open words in the input string, it sometimes takes more than a minute to
do primary marker passing. It should be mentioned that much portion of this
time is for doing garbage collections by the lisp interpreter. I conjecture that
the reason for the so many garbage collections is due to the use of a recursive
function for ﬁnding paths for each collision. It will be interesting to see what
will happen in the speed if the recursive function is replaced by the
corresponding non-recursive version. The time taken for marker passing is

about 90% of the total parsing time.

The syntactic analysis component is fast and is not affected much by the
increase of the number of words in the dictionary. The size of the rule base
did not change much after the initial writing of the rules. With the rule base
that has most of the rules that can do the clause level analysis, it has been
found that the time taken for syntactic analysis is only a fraction of the total

parsing time (around 4% of the total parsing time).

Time taken for ﬁnding the best paths for suggestions with a preposition is only

a fraction of the total parsing time and it did not vary much with the increase

170

of the knowledge base. This is due to the efﬁciency provided by the secondary

marker passing mechanism (around 6% of the total parsing time).

(d) For suggestions which do not contain a preposition, the time to ﬁnd the best
path is too long (even much longer than the time for marker passing.) Thus it

is really necessary to develop some method to speed up this process.

(e) Based on the above observations, it is concluded that it is necessary to
develop a mechanism using some parallelism to speed up the marker passing
process. (This is true regardless of the availability of the efﬁcient method of
path ﬁnding.)

8.4 What has been achieved

In this section, the items that have been achieved during the project will be listed

and summarized:

A new integrated parsing has been achieved by updating interleaved semantic pro-
cessing. The details have already been summarized and discussed in the previous

two sections in this chapter.

A rule-based syntactic analyzer with the following new features has been
developed. The base pattern is used to check the state of parsing which results in
the removal of packets or states from the parser. A method of "delaying the deci-
sion until enough context appears" has been created and implemented. This enables
the parser to have the same power as a parser using lookahead and to use simple

and uniform processing instead of the attention shifting mechanism.

A simple and clean interface between syntax and semantics has been developed.
This is achieved by using suggestions that are sent from the syntactic processing
component to the semantic processing component. But it seems that this method
can not cover everything. The use of the semantic comparison stage provides a way

of using semantic information to control the syntactic processing.

171

A mechanism called secondary marker passing has been developed to facilitate the
use of the marker passing paradigm in natural language parsers. This provides an
efﬁcient way of ﬁnding the best paths between two concepts in the knowledge
base. Without this mechanism, we should tolerate the high computational cost or
we should give up using the marker passing paradigm. As Charniak(1986) pointed

out, there is yet no better alternative to marker passing.

Along with the secondary marker passing mechanism, several heuristics have been
developed to ﬁnd the best paths between the concepts. It has been observed that
usually the secondary marker passing mechanism and the H-heuristic are powerful
enough to ﬁnd the best paths. The S-heuristic is necessary in a complex situation

such as the existence of word sense ambiguity.

The critical problem of scoring the alternative analyses has been circumvented by
the method of rank-ordering the branches by comparing the knowledge base paths.
A method of comparing the paths has been developed under the context of the
SYNSEM parser. The basic idea was to use the speciﬁcity of the concepts (these

were called the end absolutes) that constrain the ﬁrst and last relations in the path.

Attacking the disambiguation problem of structural ambiguity is not enough
because word sense ambiguity is another major source of ambiguity. Broad and
deep treaunent of this problem has been sought in this research. It has been sug-
gested that managing the set of knowledge base paths (called the semantic paths)
and propagating the elimination of word sense via these paths is an efﬁcient
method to remove the false senses. The path adjustment operation related to the
concretion of a word sense has been proposed as another source of word sense

disambiguation.

172

8.5 Future Research Item

In this section, limitations and the research items identiﬁed during the development

of the SYNSEM parser will be listed:

The ﬁrst and most important problem on the list seems to be the problem of
knowledge representation. Presently semantic information that can be used by
SYNSEM is limited to the kind of knowledge that can be encoded using KODIAK.
KODIAK is one of the most recent knowledge representation languages but it can
not satisfy the users fully. For example, how can we represent the fact that "eating
on the uain" is more plausible than "an apple on the train"? (for the sentence "The
man ate an apple on the train") How can we represent the fact that "discussing run»
ning as a sport" is more plausible than "discussing running as an act of engaging in
running"? Another problem related to knowledge representation is the use of other
diverse kinds of knowledge sources such as knowledge about discourse structure,
user modeling, goals and intentions, common sense knowledge such as time and
unexpected happenings, etc. Some of these might need a totally new kind of
knowledge representation language. To achieve the full use of semantics the sys-
tem should be able to utilize these various knowledge sources. It is clear that the
current version of KODIAK can not easily accommodate all of these knowledge
structures in a reasonably easy way (even if it can). It is not clear that the present
architecture of the parser can remain unchanged when these other knowledge

sources are eventually used.

The current grammar in SYNSEM needs to be extended so that more complex
syntactic structures can be analyzed, for example, topicalization, particles, etc. But

no major obstacles are anticipated in doing this.

Parsers which use semantics early have an advantage in handling ill-formed inputs.
It is worthwhile to extend the parser to be able to handle ill-formed inputs. The

basic strategy can be an attempt to ﬁnd the semantic connections between the

173

chunks that can not be linked by syntactic relations. Finding the paths based on
marker passing can be useful for this purpose.

Marker passing in SYNSEM requires a large amount of computation. Part of the
reason is that detection of the concretion paths and strong semantic association
paths related to word sense disambiguation should be done during marker passing.
It seems interesting to consider the use of massive parallelism in marker passing to

increase the speed.

The branches running in parallel build their own copies of syntactic representation.
It might be the case that the same structures are built in several branches. Intui-
tively, the human parsing system does not seem to have multiple copies of the
same structure. It is better to share the data structures among the branches as far as

possible.

SYNSEM has no efﬁcient way of ﬁnding the best path between two nouns that
appear in the nominal compound. It just uses the brute force method to compare
every pair of paths to ﬁnd the best path and thus it takes much time. The problem is
that there is no cue to facilitate the search (such as syntactic relation or preposi-
tion). It is necessary to develop a better scheme to ﬁnd the best semantic connec-

tion among the nouns in the nominal compound.

APPENDICES

Appendix A

Structure of Noun Phrase(NP)

Figure A—l is assumed to be the structure for an NP in SYNSEM. It is not claimed
that this su'ucture is the optimal one of the NP. But this NP structure is introduced here
for the ease of following the grammar rules and examples appearing in the thesis. Note
that the post NP modiﬁers such as the PP or the relative clause, etc. are not shown in the

ﬁgure. Instead, only the portion of the NP up to the head noun is shown in the ﬁgure.

 

Na Noun Noun Noun
/I\ [headnoun]
N s Adj Adj

Der Quant Ordinal Number

Figure A-l Structure of an NP

174

Appendix B

Determining Dead-end Situation

When the pattern matcher ﬁnds no rules that match the syntactic working
memory(i.e. CLIST), the parser should decide if the situation is the dead-end situation or
just the situation that requires a new input (called read-again situation). The parser uses
the routine called "justify-situation" for the decision. This routine is called by the parser
with no input argument(this actually use CLIST for its working). This routine is

explained below.

Let’s call a tree in CLIST Ci where "i" indicates the position of the tree from the
right side of CLIST. C1 is the rightmost tree, C2 is the second rightmost tree, etc. Let’s
call the raw patterns of rules RPi where i indicates the position of the raw pattern. Thus

RPl is the rightmost raw pattern, RP2 the second righurrost raw pattern(if it exists), etc.

Cl might lead to the construction of a basic nodes such as an NP or a PP. For
example, DET can be considered to be the leading edge of an NP. Let’s use
SC1(starting edge of C1) to specify the basic node of which C1 is a leading edge. SCl of
det, ns, na, or nc is NP. SCl of prep is PP. Except these, SCl of C1 is set to nil to indi-
cate that Cl does not lead to the basic node of NP or PP.

The basic scheme used in the routine "justify-situation" is that if the right side of
CLIST can match the pattern part of any rule by using SCl or the initial portion of the
pattern part of any rule(indicating that the rule can be matched in the near future after
more context is built in CLIST) with or without using SCl, then it is decided that the

situation is a read-again situation.

175

176

This routine is shown in more detail below:

Routine justify—situation:

{ The return with yes means that the situation is decided to be a read-again
situation and the return with no indicates the dead-end situation.
The length of a rule is the number of raw patterns in the rule. 1

(1)
(2)
(3)

(4)

if CLIST contains only one tree then return with yes;
if C1 is a conjunction node(and, or, but), then return with yes;
if C2 is of type S then
if SCl =/ nil
then begin assume CLIST whose C1 is replaced with 5C1;
call justify-situation with this new CLIST;
end
else begin if there is a rule of length 2 whose RP2 is C1
then return with yes;
if there is a rule of length 3 whose RP3 is Cl
then return with yes;
return with no;
end;
if SCl is non-nil and there is a rule of length 2
whose RP1=SC1 and RP2=C2, then return with yes;

{Sofar, checking with the rules of length I or 2 has been ﬁnished.
From now, try the rules with length 3. }

(5)
(6)
(7)

(8)

if there is a rule whose RP3=C2 and RP2=C1 then return with yes;
if there is a rule whose RP3=C2 and RP2=SC1 then return with yes;
if there is a rule whose RP3=C3, RP2=C2 and RP1=SC1,

then return with yes;
return with no;

Appendix C

Test Sentences

Diverse kinds of sentences have been used to test the features of the SYN SEM

parser. The test sentences are listed under the categories which indicates the feature of

ambiguity to be tested. The asterisk in front of a sentence indicates that SYNSEM can

not handle the sentence satisfactorily.

(1)

(2)

pp-attachment ambiguity

The man ate a cake with a fork.

The boy ate the pizza with the man.

The man ate the cake with frosting.

The man ate an apple from the orchard.

The employees of the company with the disease were examined by the doctor.
*The women discussed the dogs on the beach. ;Kurtzman(1985). SYNSEM does not have
enough conceptual expectation( described in Section 2.23) to handle this sentence.

I bought the book for the girl in the bookstore. ; the concept bookstore-buying has a slot of
location to which bookstore is the prototype. Thus "in the bookstore is attached to the VP.

The man ate in the restaurant with Tom.

The man ate the cake with Tom.

The man ate with Tom.

The spy saw the cop with the binoculars. ;Ferreira & Clifton(1986)

The spy saw the cop with the revolver. ;Ferreira & Clt'fton(1986)

Active past/past participle ambiguity

The teacher taught by the pool passed the test. ; Crain & Steedman(1985)
177

(3)

(4)

(5)

178

The present wrapped by Judy was given to Tom. ;Lytinen(1984)
The horse raced past the barn fell. ; Crain & Steedman(1985)

The students taught by the pool passed the test. ; Crain & Steedman(1985)

Case determination ambiguity

”The man ate the cracker with the dip. ;the fact that "with" can indicate the co-object of
some verbs should be encoded in the parser, which needs more consideration.

I told the story to John.

I told John the story.

I told John the story was the last thing I wanted to hear. ; Barton & Berwick(1985)

The person gave the cat to Mary.

Tom broke the window with a rock.

The rock broke the window.

The rock broke. Syntactic analysis is ﬁne. But it misanalyzes semantically("rock" is analyzed as

the instrument. Currently there is no way to detect it and reanalyze it later.

relative clause attachment ambiguity

The man bit the apple on the desk that the boy ate.

The man bit the apple on the desk that the boy painted.

*The man bit the apple on the desk that the boy bought. :1: is hard to determine where to
attach the relative clause. It is necessary to have some semantic information that it is likely for boys
to buy an apple but not a desk.

The girl saw the lion in the ﬁeld which Tom shot in the body.

missing "that" ambiguity Tom knew the man.

Tom knew the man ate the apple.

I told John the story was the last thing I wanted to hear. ;At ﬁrst "the story" is analyzed
both as the object of "tel " and as the subject of the complement clause. As soon as "was" is seen,

the object analysis is thrown away.

(6)

(7)

(8)

(9)

179

Conjunction

The boy and the girl ate an apple.

The man bit and ate the apple.

The man bit the apple and ate the pear.

The man ate an apple and the girl bit the pear.
The man ate the apple and the pear.

The boy ate and the girl saw the apple.

Wh-phrase sentences: Wh-question and relative clause

Which apple did the man eat ?

I ate the apple which the man bit.

I ate the apple that the man bit.

He ate the apple the man bit.

Who do you want to visit?

Who do you want to eat the apple? The equi-np deletion analysis and the Wh—analysis("who"
ﬁlls the gap between "want" and ”to") goes on in parallel until one is rejected.
Who do you want to give the apple to? ;Marcus(1980)

What did you give Tom the book for? ;Hirst(1984)

I gave the boy who you gave the apples to three books.

Equi-np deletion

The man wanted to eat the apple.

Tom wanted the man to eat an apple.

The man wanted the book.

Semantic anomaly
The boy drank the apple. ;the analysis fails because of semantic anomaly("apple" can not be
object of action ”drink".
”The boy ate the green apple. ;The semantic oddness of eating unripe apple can not be

detected in S YNSEM

180

(10) Word sense ambiguity
The slug operated the vending_machine. ;Hirst(1984)
The man operated the machine with the tool.
The granite rocks near the shore. ;Milne(1982)
The crook operated the pizza_parlor. ;Hirst(1984)
*The astronomer married a star. ; Charniak(1983). "star" is at ﬁrst disambiguated to be a
concept of "star in the sky" but later the parser gets into trouble because the object of "marry"
should be of type "person".
Tom ﬁxed the washer. ;Lytinen(1984)

Tom ﬁxed the game. ;Lytinen(1984)

(1 l) Metaphor: S YNSEM can not handle this type of sentence.

*The car drank the gasoline. ; Small(1983)

(12) Multiple part of speech ambiguity
The granite rocks near the shore. ;Milne(1982)
*The prime number few. ; Milne(1982). SYNSEM can not handle this one because of the lack

of necessary semantic information.

(13) Plain sentences
The apple is big.
The boy said that Tom had stolen the watch.
The girl sleeping in the room wanted to visit Tom.
(14) Ambiguity from present participle
*The man eating tiger growls. ; Small(1983). SYNSEM currently can not handle this sentence
because it does not have appropriate syntactic rules.
The man eating shrimp growls. ;Small(1983)
(15) Vague word sense ambiguity

I took a picture.

181

The man took aspirin.
The man got the goods. ; "got" is analyzed to be "buying".
(16) Semantic association between nouns in a nominal compound
The cotton clothing is made of grows in Mississippi. ;Marcus(1980). SYNSEM can not
model the garden pathing effect of this sentence shown by humans. Two possible analyses are

sought in parallel in S YNSEM .

Appendix D

Output of Parsing

The following system output of SYNSEM shows how the parser works during the

parsing of sentence "the teacher taught by the pool passed the test" which was explained

extensively in Chapter 4 and 6.

;;; Dribble ﬁle #P"lusr.MC68020/users/staff/ra/nlp/out-taught” started
T

> (mon)

RRRRRRRRRead next word : THE

A word read 111 read-word: word=@

Come into rule-matchin part...

rule succeeded. S-ST ART

Retum from rule-matching mtine. mlesz (S-START)
Only one rule selected == S-ST ART

        

 

 

 

mm: .(S) name: NIL range: (1 l) sansug .NIL
feature eaturezs ((S (MAIN))) trace .NIL

“ode.SSl SS 1 l :NIL
paEms: l2" 2((‘é‘é3‘f3ém NF ”mm“

Come mto rule-matching

pan...
Retum from rule-matching routine. rules: NIL
No rules are found.

------- justify ﬁnished.as success --------------

A word read in read-word: word=THE

Come into rule-matching part...

Retum from rule-matchin routine. rules: (PARSE DET)
Only one rule selected: ARSE- DET

 

 

 

 

          

ez‘Nlp‘or: .(S) namezNIL ran e: (1 l) semsug:NIL
parent features eatmes:((S (MAIN) )tracezNIL

node:SSl .(SS) narne:@ran e: (l l) sernsugzNIL

parent: 82 Puma: ((88)) trace: N'ﬁ.
nodezNS4 pos:(NS) name:g:gNILrane(22)semsu :NIL
perennNIL features: ((NS (NUM IS 28 38 1P 2P 3P))) trace: NIL

node:DET3pos:(DEI') namez'I‘HEg:rane(22)semsug :NIL
parentzNS4 tures: ((DET (NUM IS 28 BS 1P 2P 3P))) trace: NIL

Come mto rule-matchmg

part...
Retum from rule-matching routine. rules: NIL
No mles are found.

------ ﬁnished. as success -------------
cad next word: TEACHER

Busing markers. instance=<abso TEACHERJ>

ump into node carrying same ori mark at cnode=~<asp TEACHERTEACHER-TEAClb
Loop detected at cnode=<abso PERSON>
Loop detected at cnode=<abso PLACE>

182

183

A word read m read-word: word=TEACHER
Come mto rule—matching part...

Return from rule-matching routine. rules: (NOUN-PARSEI)
Only one rule selected = NOUN-PARSEI

3 Trees in CLIST
node:S2 :(S) name:NIL range:(l l) semsugNIL
parent: features:((8 (MAIN))) trace :NIL

node:SSl :(SS) name :ran@ e.(l l) semsugzNIL
parent:82 I”. eatuzres ((88)) trace: NIL:
node:NC6 pos:(NC) name:NIL ran g:ge(22)semsu :NIL
parent:NIL features: ((NC (NUM 18 28 3S 1P 2P 3P))) tracezNIL

node:NS4 pos:(N8) name:NIL range:(2 2) semsug:NIL
pumtzNCG features:((NS (NUM lS 28 38 1P 2P 3P))) trace:NIL

nodezDE'I‘3 pos:(DET) name:THE range:(2 2) semsug:NIL

parent:NS4 features:((DET (NUM lS 28 3S 1P 2P 3P») trace:NIL
node:NOUN5 pos:(NOUN) namezTEACHER ran e:(3 3) semsugzNIL
parentleL features:((NOUN (NUM 3S))) tracezNIf.

 

 

Come into rule-matching part.

Return from rule-matching routine. rules: NIL
No rules are found.

----- justify ﬁnished.as success “.-...-m.--
RRRRRRRRRead next word : TAUGHT

passing markers. instance:<abso TAUGHT. 2>

this 1s regular collision at <abso TEACH).

this rs regular collision at <abso TRSACIB.

Bump into node carrying same ori mark at cnode=<abso PERSON>
this is regular collision at <abso 'I'I'>.

Loop detected at cnode:<asp RECREC-TEACH)

A word read in read-word: word=TAUGHT

Come' mto rule-matching part...

Retum from rule-matching routine. rules: (NOUN-PARSE3)
Only one rule selected: N OUN -PAR8E3

 

        

 

 

 

mm18) name .NIL range: (1 l) semsug :NIL

s:((S (MAIN))) trace NIL
nod .88] SS 1 l NIL
punciuSZ I’mtu (res)((’S.Sn)‘)c£cem NE: ( )sernsug:

nodezNP8 pos:(NP) name:NIL range:(23) semsug :NIL
parent:NIL features:((NP (NUM 3S) (MODIFIABLE))) tracezNIL

nodezNCG pos:(NC) name:NIL range:(2 2) semsugzNIL
paratt:NP8 features:((NC (NUM IS 28 3S 1P 2P 3P))) trace:NIL

node :NS4 pos: (NS) name .NIL range: (22) semsug :NIL
parentzNCé features: ((NS (NUM 18 28 3S 1P 2P“‘8 3P))) trace .NIL

node:DET3 pos:(DET) namezTHE range:(2 2) sansugzNIL
parentzNS4 features:((DET (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:NOUN5 pos: (NOUN) name: TEACHER range: (3 3) semsug: NIL
parent: NP8 features: ((NOUN (NU M 38))) trace:NIL

Corne' mto rule-matching part...

Remm from rule-matching routine. rules: (SUBJECT -RULE)
Only one rule selected: SUBJECT -RULE

 

        

 

 

 

node:Sm(S) name:NIL ran e:(l 3)semsug:NIL
parent

eatures:((S (MAIN) )trace :NIL
”0d .881 SS 1 l :NIL
Sarge] 82 F” (res)((l8.8")l)etr.?cerzan NE: ( )semsug

.(NP) name :NIL range
parait: 82 mamas tures(:(NP (NUM 3S) (MODIFIABLE) (SUBD)) trace :NIL

nodezNCG pos:(NC) name:NIL range: (22) semsugSNIL
parent:NP8 features:((NC (NUM 18 28 38 IP 2P3 ))) trace: NIL

node:N84 pos:(N8) name:NIL range:(2 2) semsugzNIL

184

parent:NC6 features:((NS (NUM 18 28 38 1P 2P 3P») tracczNIL

;OdczDETB perisKDET) narne:THE ran e:(2)2 )scrnsug :NIL
‘ parentzNS4 tures ((DET(NUM 18 3S 1P2P3P)))tracc:NIL

nodezNOUNS p08: (NOUN) name. TEACHERran NIII. .(3 3) setnsugtNIL
mm features: ((NOUN (NUM as») trace

Come into rule-matching part...

Return from rule-matching routine. ntles= NIL

No rules are found.

~-—---- justify ﬁnishedas success --------------

anobjcctreadinfromcash:V7

Come into rule-matching part...

Return from rule-matching routine rules: (MAIN-VERBZ NP-VPP NP-VPPl)

come into go__on_scvera1__ntles. rules: (NP-VPPl NP-VPP MAIN-VERBZ)
start of branch in go_sevcral_rulcs = (NP-VPPl)

Suggestim received: dir=CONNECT ol =<abso TEACHERJ) 02=<abso TAUGHT.2> o3=TO whichasp: 2
two-way collisions betw <abso TAUGHT.2> and <abso TEACHER. l> = 31
# of paths for these two-way collisions : 1502

call into secondary marker passing

Number of original three-way collisions : 1

number of collisions before H-heuristics = 1

# of collisions after H-heuristic: l

Nmnberoffullpathsbefilrelhltering=l

number of full paths after num. relation ﬁltering : 1

Number of full paths after hierarchy (I end absos : l

Final paths selected in pickup_interp.
<<>> (<abso TEACH) <abso PERSON) (<abso TAUGHT.2> I <abso TEACH) C_INV <asp TEACHREC-TEACH) A_INV

<rel REC-TEACH) A_INV <asp RECREC-TEACH) C_INV <abso PERSON) D <abso TEACH_PROFESSION AL) D
<abso TEACHER) I <abso TEACHERJ>))

Trees in CLIST
node:SlO pos:(S) name:NIL range:(l 3) scmsugzNIL
parent:NIL fcatures:((S (MAIN))) tracczNIL

node:SSl] pos:)(88 name:@ range:(l l)scmsug:NIL
parent:810 eatures:((SS)) ):traceNIL

 

 

 

 

«SU BL-
ncdczNPl2 .(NP) name .NIL range: (2 NIL) scm .NIL
parent: SlO Ferd: tuzres ((NP (NUM 3S) (SUBJ))) tracc: N181.

nodezNCl3 pos: (NC) name: NIL range: (22) semsugNIL
parent .NP12 features: ((NC (NUM 18 828 38 1P 2P3 ))) trace: NIL

nodezNSl4 pos:(NS) name:NIL range: 2(2) semsugzNIL
parent:NCl3pm fcautres: ((NS (NUM 18828 38 IP 2P P))) trace: NIL

node: DETlS s: (DET) name: THE marge: .(2 2) sernsug:NIL
parent: N814 eatures: ((DE'I‘ (NUM IS 38 1P 2P 3P))) tracezNIL

;odeNOUNlé pos:(NOUN) name: TEACHER ruNﬁc: .(3 3) semsug NIL
parcnt:NP12 features: ((NOUN (NUM 38))) trace

node: 817 pos:(S) name: NIL range :(4 NIL) semsug: NIL

parfsuzNPIZ features: ((8 (SEC) (STRUC- 0K)» trace: NIL
nodezNP18 I)“: (NP) name :NIL range:NIL semsugzNIL
parent: 817 eatures: ((NP (DUMMY))) trace .NIL

node:AUXl9 pos:(AUX) name:NIL rangc:NIL semsugzNIL
parents 17 features:((AUX (PASSIVE))) tracetNIL

:NIL 4 N11. :NIL
parents 17 Gratin": ((3%)!)etrace range: ( )sernsug

node:V9 pos:(V) namezTAUGHT range:(4 4) sernsu :(<abso TEACHERJ) <abso TAUGHT.2> TO 2)

pIISEtImo features:((V (EN) (TENSE PAST) )(TRANSITIV E) (DATIVE) (DATIVE-PREP TO) ))

nodezNP2l pos: (NP) name:NIL range: NILscmsugNIL
parart:VP20 features. ((NP (TRACE)))trace:NP12
node. 817 .(S) name .NIL range: (4 NIL) semsug :NIL
parentzNPf20‘fca tures: ((8 (SEC) (STRUC-OK») tracc1NIL

nodczNP18:(NP) name:NIL ran ezNIL semsug:NIL
parurt:Sl7 eaturcs:((NP (DUMMY) )tracczNIL

. nodczAUXl9 pos:(AUX) name:NIL rangezNIL semsugzNIL
paratt:Sl7 features:((AUX (PASSIVE))) tracezNIL

185

;ode:VP20 for: .(VP) name. NIL range:(4 NIL) semsug1NIL
parent:817 eatures: ((VP)) trace:NIL

node: V9 pos: (V) name: TAUGHTran range: .(4 4) serum :(<abso TEACHER.1> <abso TAUGHT.2> TO 2)

Wm features: ((V (EN) (TENS PAST) (M (IRANSI'I'IVE) (DATIVE) (DATIVE-PREP TO) ))

nodezNP21 pos:(NP) name:NIL range:NIL semsugzNIL
parern:VP20 features:((NP (TRACE))) trace:NP12

beginning (1 NEXTJound.

Come into rule-matching part...

No rules matchcd.(just after NEXT_ROUND1)

--«--- justify ﬁnishedas success .-............

End of one rule-process for=(NP-VPP1) merit=(<abso TEACH) <abso PERSON) (<abso TAUGHT .2) I <abso TEACH) C__INV
<asp TEACH. REC-TEACH) A_ INV <rel REC-TEACH) A_INV <asp REC.REC-TEACH) C_INV <abso PERSON) D

<abso TEACH PROFESSION AI.) D <abso TEACHER) l <absoTE CHER.1>))

start of branch irt go_several_rulcs : (NP-VPP)

Suggestion received: dit=CONNECT 01 =<abso TEACHERJ) 02=<abso TAUGHT.2> o3=OBJ whichasp= 2

two-way collisiorts bctw <abso TAUGHT.2> and <abso TEACHER. l) = 31

4 of paths for these two-way collisions = 1502

call into secondary marker passin

Number of original three—way colﬁsions - 1

number of collisions before H -hcuristics = 1

# of collisions after H-heuristic: 1

Number of full paths bef it rel ﬁltering: 1

number of full paths after num. relation ﬁltering- - 1

Number of full paths after hierarchy of end absos: 1

Filter paths having too_ abstract _...conccpt

removed: (<abso TAUGHT. 2) I <abso TEACH) D <abso TRSACI‘) C_ INV <asp TRSACI‘. ACI'EE-TRSACI) A _INV <rcl ACI'EE~TRSACI‘)
A _INV ACI'EE. ACI‘EE- TRSACI‘) C _INV <abso THING) D <abso PHY_ OBJECI‘) D <abso ANIMATE) D <abso ANIMAL> D
<abso HIG _ANIMATE) D <abso PERSON) D <abso TEACH_ PROFESSIONAL) D <abso TEACHER) I <abso TEACHER. 1))
Final paths scleaed tn pickup_ interp: no paths found.

start of branch tn go_ several_ ntles— - (MAIN- VERBZ)

Suggestion received: dir=CONNECF ol=<abso TEACHER. l) 02=<abso TAUGHT. 2) o3-5UBJ whichasp: NIL
two-way collisions bctw <abso TAUGHT. 2) and <abso TEACHER. 1): 31
3 of paths for these two-way collisions: 1502

call into secondary market passing

Number of original three-way collisions : 3

nunber of collisions before H-heuristics : 3

drop collision at ﬁlter]: cnode=<rel ACTOR-AC1)

d collision at ﬁlterl: cnode:<rel AC'IOR-TRSACT)

# o collisions after H-heuristic: 1

Number of full paths bef it rel filtering- - 1

number of full paths after num. relation filtering- - 1

Number of full paths after hierarchy of end absos- - 1

Final paths selected tn pickup_ jitter-EH
<<>> (<abso TEACH) <absoTEA _PROFESSIONAL) (<abso TAUGHT. 2)I <abso TEACH) C_ INV <asp TEACH. TEACHER- TEACH)

A _INV (rel TEACHER-TEACH) A__ INV <asp TEACHER. TEACHER-TEACH) P_ INV <abso TEACH _PROFESSIONAL) D <abso
TEACHER) I <abso TEACHER. 1>)

_.— Trees tn CLIST
node: 836 pos: (8) name: NIL range: (1 4) semsug: NIL
parent:NIL features: ((8 (MAIN))) trace .NIL

0d 8837 SS 1 l :NIL
.....i"w " osoo F“ (mlté's'i‘fgﬂ ”if" ’ m“
nodeNP38 ):narne NIL range: (2 3) serum
paratt:S36 (mm cautres:((NP (NUM 3S) (MODIFIABL LE) (SUBD)) trace :NIL

node: NC39 pos: (NC) name: NIL range: (22) semsug)NIL
parent:NP38 features: ((NC (NUM 18828 38 IP 2P3 ))) trace:NIL

nodc:NS4O pos:(NS) name:NIL ran g:)e(22 )u:NILscmsg
parentzNC39 features. ((NS (NUM 18 28 38 1P 2P P))) trace :NIL

node: DET41 pos: (DET) name: THE c: (2 2) semsugzNIL
parattzNS4O featurcs:( ((DET (NUM 18 38 1P 2P 3P))) trace:NIL

;ooo- .NOUN42 poo (NOUN) name: TEACHER W' .(3 3) semsug: NIL
Mme features: ((NOUN (NUM 33)» trace

node: AUX43 pos. (AUX) nune :NIL range :NIL semsug: NIL
parent:S36 features: ((AUX)) trace:NIL

nodc:VP44 pos:(VP) name:NIL range:(4 4) semsugzNIL

 

 

 

186

parcnt:S36 features: ((VP (NUM 18 28 3S 1P 2P 3P) (DATIVE-PREP TO) (DATIVE) (TRANSITIVE) (MAIN) (T EVSE PAST) (EN )))

node: V35 pos:(V) namezTAUGHT ran ge:(4 4)sesn su:(<abso TEACHERJ) <abso TAUGHT.2> SUBJ NIL)
parent: VP44 features:((V (EN) (TENSE 8PAST) (TRANSITIVE) (DATIVE) (DATIVE-PREP TO) ))
Come mto rule-matchingpa
No rules matched. (just after NEXT _ROUNDI)
- ------ justify linishedas success ----------

End of one rule-process for=(MAIN-VERBZ) merit=(<abso TEACH) <abso TEACH_PROFESSIONAL> (<abso TAUGHT.2> I <abso TEACH)
C_INV <asp 'IEACH.TEACHER-'IEACH> A_INV <rel TEACHER-TEACH) A_INV <asp TEACHERTEACHER-TEACH) P_IN V
<abso 'IEACH_PROFESSIONAL> D <abso TEACHER) I <abso TEACHERJ)»

startof rmparallcltill cneis found.#of branchcs:2
#ofscmbranchcs=2 #ofnoscrnbranches:0

Backed up branches into I'backup-staclt“ : (NP-VPPl)
Branches to corttinue to run : (MAIN -VERB2)

FINAL One brutch chosen after parallel execution: (MAIN -VERB2)

Come into rule-matching part...
Retum from rule-matching routine. rules: NIL

------- justify ﬁniahed.as success ------------
RRRRRRRRRead next word : BY

A word read irt read-word: word=BY

Come into rule-matching part...

Return from rule—matching routine. rules: NIL
No rules are found.

-------- ' s ' ﬁnished.“ success ----«-—------
cad next word : THE

A word read in read-word: word=THE
Come into rule-matching part...
Retum from rule—matching routine. rules: (PARSE-DET)

Only one rule selected = PARSE-DET

Trees in CLIST
node:S36 pos:(S) name:NIL range:(l 4) semsugzNIL
parcntzNIL features:((S (MAIN))) trace:NIL

node:SS37 pos:(SS) namc:@ range:(l l) semsugzNIL

m-SU E11836 features:((88)) trace:NIL

node:NP38 pos:(NP) name:NIL range:(2 3) semsug:NIL

parent2836 features:((NP (NUM 3S) (MODIFIABLE) (SUBD)) trace:NIL

;odezNC39 pos:(NC) name:NIL range:(2 2) semsugzNIL
parentzNP38 features:((NC (NUM 18 28 38 1P 2P 3P)))tracc1NIL

node:NS40 pos:(NS) name:NIL range:(2 2) scmsug:NIL
parentzNC39 features:((NS (NUM 18 28 3S 1P 2P 3P))) trace:NIL

node:DET4l pos:(DET) namczTHE range:(2 2) semsugzNIL
parent:NS40 features:((DET (NUM 18 28 3S 1P 2P 3P))) trace:NIL

iiodeNOUN42 pos:(NOUN) namc:TEACHER range:(3 3) scmsug:NIL
paretn:NP38 features:((NOUN (NUM 3s») trace-.NIL

nodc:AUX43 pos:(AUX) name:NIL range:NIL semsugzNIL
parutt:836 features:((AUX)) trace:NIL

 

 

 

 

node:VP44 pos:(VP) name:NIL range:(4 4) semsugzNIL
parent:836 features:((VP (NUM 18 28 3S 1P 2P 3P) (DATIVE-PREP TO) (DATIVE) (TRANSITIV E) (MAIN) (T ENSE PAST) (EN)))

node:V35 pos:(V) name:TAUGHT range:(4 4) scmsug:(<abso TEACHERJ) <abso TAUGHT .2) SUB] NIL)
parentzVP44 features:((V (EN) (TENSE PAST) (MAIN) (TRANSITIVE) (DATIVE) (DATIVE-PREP TO) ))
nodczPREP45 pos:(PREP) namc:BY range:(S 5) semsugzNIL
parent:NIL features:((PREP)) trace:NIL
node: N847 pos: .(NS) name: NIL range: (6 6) sernsug .NIL
parent: NIL cat:ures ((NS (NUM 18 28 3S 1P 2P 3P))) trace:NIL

node.DET 46 per: .(DET) name: THE range: (6 6) semsug :NIL
parentzNS47 tures:((DET (NUM 18 28 3S 1P 2P 3P))) trace: NIL

Come tnto rule-matching part...
Return from rule-matching routine rules: NIL

187

No rules are found.

- ----- justify finishedas success --....m---”
RRRRRRRRRead next word : POOL

passing markers. instance=<abso POOL.3>

this is regular collision at <abso PLACE>.

this ts regular collision at <abso INANIMATE>.

this is regular collision at <asp ACT .ACI‘OR-ACb.

Loop detected at cnode=<asp ACT .INSTR-AC'b

this is regular collision at <asp ACLACCOMPANIER-AC'b.

this is regular collision at <abso TRSACIB.

Loop detected at mode-«asp POS.IDENTITY-REL>

Bump into node carrying same ori mark at cnode=<asp ACI'.PURPOSE-ACT>
this is regular collision at <rel AG-EXAMINE>.

A word read in read-word: p‘word=POOL
Come into rule-matching

Retum from rule-matching” routine. mles= (NOUN-PARSEI)
Only one rule selected = NOUN -PARSE1

 

Trees in CLIST
node: 836 pos: (8) name: NIL range: (1 4) semsugzNIL
parent: NIL features: ((8 (MAIN))) trace .NIL

node:SS37 s. (88) name: @ran e. (l l) semsugzNIL

parmt:836 p0 eatures:((SS)) trace: NIL:

-SUBJ--

node:NP38 s. (NP) name. NIL range. (2 3) sems

parent:836 Futures: ((NP (NUM 38) (MODIFIABLE) (SUBJ))) trace: NIL

node:NC39 pos:(NC) name:NIL range:(2 2) semsugzNIL
parentzNP38 features:((NC (NUM 18 28 38 1P 2P 3P))) trace:NIL

;ode.NS40 po:s(NS) name.NIL range: (22P'.NIL2)semsu§
parentzNC39p° features: ((NS (NUM rs 23 3s 1P2? P))) trace:NIL

node:DEI‘4l .(DE'I) name :THE e: (2 2) sansugzNIL
parentzNS40eatu your“ :((DET (NUM 18.2.8 38 1P 2P 3P)))trace1NIL

node: NOUN42 pos: (NOUN) name: TEACHER 7:31.243 3) semsug: NIL
parent :NP38 features: ((NOUN (NUM 3S))) trace

 

 

 

node: AUX43 pos:(AUX) name: NIL range: NIL semsug: NIL
pamt: S36 features: ((AUX)) trace:NIL

node: VP44 Foam .(VP) name:NILran ranze (4 4) semsug: NIL
parent: S36 :((VP (DA'I'IVE- REP 10) (DATIVE) (TRANSITIVE) (MAIN) (TENSE PAST) (EN)))

node: V35 pos: (V) name: TAUGHT range: (4 4) semsug: (<abso TEACHER. l> <abso TAUGHT. 2> SUBJ NIL)
parent: VP44 features: ((V (EN) (TENSE PAST) (MAIN) (TRANSITIVE) (DATIVE) (DAIIVE- PREP TO) ))
node:PREP45 pos: (PREP) name: BY range: (5 5) semsug: NIL
parent:NIL features: ((PREP» trace:NIL
node: NC49 pos: (NC) name: NIL rangze (6 6) semsugzNIL
parent:NIL features:((NC (N UM IS 28 38 1P 2P 3P))) trace:NIL

node:N847 pos:(NS) name:NIL range:(6 6) semsug:NIL
parent:NC49 features:((NS (NUM 18 28 38 1P 2P 3P))) trace:NIL

nodezDET46 pos:(DET) namezTHE range:(6 6) semsugzNIL

parentzNS47 features:((DET (NUM 18 28 38 1P 2P 3P))) trace :NIL
node: NOUN48 pos:(NOUN) namezPOOL range: (7 7) semsug:NIL
parent: NIL features: ((NOUN (NUM 38))) trace: NIL

Come into rule-matching part...

Retum from rule-matching routine. rules: NIL
No rules are found.

-------- justify ﬁnished.“ success --------------
RRRRRRRRRead next word : PASSED

passing markers. instancc=<abso PASSED.4>

this is regular collision at <abso PASS>.

this is regular collision at <abso TRSACIB.

this is regular collision at <abso PHY_OBIEC'I>.

Loop detected at atmqsp ACTOR.ACIOR-ACI>

Loop detected at anode=<asp ACCOMPANIERACCOMPANIER-ACI‘)
Bump into node carrying same ori mark at cnode=<abso IT>

this is regular collision at (rel PAT-WRAP>.

Loop detected at mode=<abso PLACE>

188

Come into syntax-semantics.

A word read in read-word: word=PASSED
Come into rule-matching part...

Retum from rule-matching routine. rules: (NOUN-PARSE3)
Only one rule selected = NOUN-PARSE3

 

 

Trees in CLIST
node:S36 pos:(S) name:NIL range:(l 4) semsug:NIL
parentzNIL features:((S (MAIN))) trace:NIL

node:SS37 pos:(SS) name:@ range:(l I) semsugzNIL

parcnt:836 fcatures:((88)) trace:NIL

«SUBL-

nodczNP38 pos:(NP) name:NIL range:(2 3) scmsug:NIL

paratt:836 features:((NP (NUM 3S) (MODIFIABLE) (SUBJ))) trace:NIL

node:NC39 pos:(NC) name:NIL ran g:c(‘2 2g)scmsu :NIL
parera:NP38 features. ((NC (NUM IS 28 38 1P 2P 3P))) trace:NIL

node:NS40 pos:(NS) name:NIL range:(2 2) semsugzNIL
parentzNC39 features:((NS (NUM IS 28 38 1P 2P 3P))) trace:NIL

node:DET41 pos:(DET) namezTHE range:(2 2) sansugzNIL
parentzN840 features:((DET (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:NOUN42 pos:(NOUN) namezTEACHER range:(3 3) semsugzNIL
parent:NP38 features:((NOUN (NU M 38))) trace:NIL

node:AUX43 pos:(AUX) name:NIL rangc:NIL semsugzNIL
parent:836 features:((AUX)) trace:NIL

 

 

nodezVP44 pos:(VP) name:NIL range:(4 4) semsug:NIL
parent'836 features ((VP (NUM 18 28 3S 1P 2P 3P) (DATIVE-PREP TO) (DATIVE) (TRANSITIVE) (MAIN) (T ENSE PAST) (EN)))

node V35 pos (V) name:TAUGHT range:(4 4) semsug:(<abso TEACHER.1> <abso TAUGHT.2> SUB] NIL)
parent:VP44 features:((V (EN) (TEN SE PAST) (MAIN) (TRANSITIVE) (DATIVE) (DATIVE-PREP TO) ))
nodc:PREP4S pos:(PREP) namc:BY range:(S 5) semsug'NIL
parentzNIL features: ((PREP)) trace: NIL
node: NPSI s. (NP) name: NIL rang c: (6 7) scmsu
parentzNIL r0 eatzures ((NP (N UM 38) (MODIFIABL LIE»)L trace: NIL

node:NC49 pos: (NC) name: NIL rang c: (6 6) semsug :NIL
parmtzNPSI features: ((NC (NUM 18 828 38 1P 2P 3P))) trace:NIL
node:N847 pos(N:8) nune: NIL range: (6 6) semsugzNIL
parentzNC49pos features: ((NS (NUM 18g 28 38 1P 2P P))) trace .NIL
node: DET46 g:(DET) amczTHE rang c:(6 6) semsugzNIL
parentzNS47 tures: ((DET (NUM IS 28 38 1P 2P 3P))) trace:NIL

node:NOUN48 pos:(NOUN) namczPOOL range:(7 7) semsugzNIL
parentzNPSI features:((NOUN (N UM 38))) trace:NIL

Come into rule-matching part...
Retum from rule-matching routine. rules: (build-PP)

Only one rule selected = build-PP

 

 

 

Trees in CLIST
node:S36 pos:(S) name:NIL range:(l 4) semsugzNIL
parentzNIL features:((S (MAIN))) trace:NIL

node:SS37 pas: .(SS) name: @ran NE: .(1 l) semsug: NIL

parent.836 tures: ((88)) trace:
«SUBL-

node:NP38 : (NP) name .NIL range: (2 3) semsu
parent:836 Shires: ((NP (NUM 3S) (MODIFIABL LE) (SUBD)) trace :NIL

node:NC39 pos:(NC) name:NIL range:(2 2) semsugzNIL
parent:NP38 features:((NC (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:NS40 po:s(N8) name:NIL range:(22) sern suzNILg
parent: NC39p0 features: ((NS (NUM IS 28 38 1P 2P P))) tracc .NIL

nodc:DET4l :(DET) namezTHEran
parentzNS40 eatures:((DET (N UM 18

;ochOUN42 pos:(NOUN) name: TEACHERran c: (3 3) semsug: NIL
parerrtzNP38 features: ((NOUN (NUM 35)» trace: NIL.

 

c:(2 2) semsugzNIL
38 1P 2P 3P))) trace:NIL

189

nodezAUX43 pos:(AUX) name:NIL rangczNIL semsugzNIL
parerrt:836 features:((AUX)) trace:NIL

node:VP44 Fos: (VP) )name :NIL range: (4 4) semsug :NIL
parcnt:836 eatures: ((VP (DATIVE-PREP T0) (DASTIVE) (TRANSITIVE) (MAIN) (TENSE PAST) (EN)))

node: V35 pos: (V) name. TAUGHT range: (4 4)scrn my (<abso TEACHERJ> <abso TAUGHT.2> SUBJ NIL)

parem: VP44 features: ((V (EN) (TENSE PAST)NIL (TRANSITIVE) (DATIVE) (DATIVE-PREP TO) ))
node:PP52 pos:(PP) name:NIL range:(5 7) semsug

parent:NIL features:((PP (MODIFIABLE) (NUM 3s») mmcezNIL

node:PREP45 pos:(PREP) namezBY range:(5 5) sunsug:NIL
parurtzPP52 features:((PREP)) trace:NIL

node:NPSI pos:(NP) name:NIL range:(6 7) semsugzNIL
parcntzPP52 features:((NP (NUM 3S) (MODIFIABLE))) trace:NIL

node:NC49 pos:(NC) name:NIL range:(6 6) semsugzNIL
parentzNPSI features:((NC (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:NS47 pos:(NS) name:NIL range:(6 6) scrnsugzNIL
parent:NC49 features:((NS (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:DET46 pos:(DET) namczTHE range:(66) semsugzNIL
parmuNS47 features:((DET (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:NOUN48 pos:(NOUN) namc:POOL range:(7 7) sernsugzNIL
parerrtzNP51 features:((NOUN (NUM 3S))) trace:NIL

Come into rule-matching part...
Return from nrlc-rnatching routine. rules: (VP-PP)

Only one rule selected = VP-PP

Suggestion received: dip-CONNECT ol=<abso TAUGHT.2> o2=<abso POOL.3> o3=BY whichasp= NIL
two-way collisions bctw <abso POOL.3> and <abso TAUGHT.2> = 27
# of paths for these two-way collisions = 562

call rate secondary marker passing

Number of original three- -way collisions- - 2

number of collisions before H-heuristics: 2

f of collisions after H~heuristicz 1

Number of full paths bcf it rel ﬁltering = 1

number of full paths after nurn. relation ﬁltering = 1

Number of full paths after hierarchy of end absos = 1

Final paths selected In pickup_ irrterp.

<<>> (<abso PLACE> <abso TEACH) (<abso POOL3> I <abso POOL> D <abso PLACE> C _INV <asp LOC. LOC TEACH) A _INV
<rel LOC -TEACH> A _IN V <asp TEACH. LOC- TEACH> C_ INV <abso TEACH> I <abso TAUGHT. 2>))

—— Trees in CLIST
node:S36 pos:(S) name:NIL range:(l 4) semsug:NIL
parerruNIL features:((S (MAIN))) trace:NIL
0d 8837 SS 1 l :NIL
WWW ‘ us36 121‘... ‘J«“s‘s“>3°3££i °( ’m‘“
;{Jims .(NP) name:NIL range: (2 3) scrnsu
parcrrt:836eatu roan” :((NP (NUM 3S) (MODIFIABL LE) (8UBJ))) trace. NIL

node:NC39 pos: (NC) name:NIL rang c:(2
parerrtzNP38 features: ((NC (NUM 18828 38 )lPZP 253$)»L trace: NIL

node:NS40 pos: (NS) name:NIL range: (2 2)semsu suzNIgL
paresrtzNC39 features: ((NS (NUM IS 28 38 1P 2P P))) trace:NIL

nodezDET4l pas: .(DET) name :THE range: (2 2) semsug :NIL
parurtzNS40 res: ((DET (NUM 18 28 38 1P 2P 3P))) trace: NIL

;odc :NOUN42 pos: (NOUN) name: TEACHER c: (3 3) semsug: NIL
parent .NP38 features: ((NOUN (NUM 38))) tracefNﬁ.

 

 

 

node: AUX43 pos: (AUX) name:NIL range: NIL semsug: NIL
parent: S36 features: ((AUX)) trace: NIL

node:VP44 .(VP) name:NH. range: (4 7) semsug :NIL
parurt:836 mm: ((VP (DATIVE-PREP T0) (DATIVE) (TRANSITIVE) (MAIN) (TENSE PAST) (EN)))

node: V35 pos: (V) name: TAUGHT range: (4 4) semsug: (<abso TEACHER. l> <abso TAUGHT. 2> SUBJ NIL)
parent: VP44 features: ((V (EN) (TENSE PAST) (MAIN) (TRANSITIVE) (DATIVE) (DATIVE- PREP T0) ))

nodezPP52 pos:(PP) name:NIL range:(S 7) sernsug:NIL

190

parem:VP44 features:((PP (MODIFIABLE) (NUM 3S))) trace:NIL

r:ode:PREP45 pos:(PREP) namezBY range:(S S) semsug:NIL

paruttzPPSZ features:((PREP)) trace:NIL

node:NP51pozs (NP) name :NIL range: (6 7) sernmEéabso TAUGHT. 2> <abso POOL.3> BY NIL)
parentzPPSZ features: ((NP (NU M 38) (MODIFIAB ))) trace:NIL

node:NC49 pos: (NC) name: NIL ran g:sernse(66) sernsug).NlL
parurtzNPSI features: ((NC (NUM 18 28 381P2P 3 ))) trace: NIL

node“ :NS47 pos:(NS) name:NIL rang e. (6 6) semsugzNIL

parentzNC49p“ features. ((NS (NUM 188 28 38 1P 2P P))) trace .NIL
node:DET46 (DEF) name: THE range: (6 6) semsug :NIL
parent:NS47 Futures: ((DET (NUM 18 28 38 1P 2P 3P)g)) trace: NIL

;we1NOUN48 pos:(NOUN) namezPOOL range:(7 7) sernsug:NIL
parattzNPSI features:((NOUN (NUM 38))) trace:NIL

Come into rule-matching part...
Retum from rule-matching routine. rules: NIL
No rules are found.

----- justify ﬁnished.“ success ----------
anobjectreadinfromcash=VSO

Come into rule-matching part...

Return from rule-matching routine. rules: (NP- -VPP)

Only one rule selected: NP-VPP

Suggestion received: dir=CONNECF 01=<abso POOL3> 02=<abso PASSED. 4> 03:03] whichasp= 2
two-way collisions betw <abso PASSED. 4> and <abso POOL. 3>- - 17
ll of paths for these two-way collisions = 356

call into secondary marker passing

Number of original three-way collisions = 1

nmnber of collisions before H-heuristics = 1

# of collisions after H-heuristic: 1

Number of full paths bef it rel ﬁltering = 1

number of full paths after nurn. relation ﬁltering = 1

Number of full paths after hierarchy (I end absos = 1

Filter paths having too__abstract_concept....

Final paths selected in pickup_interp: no paths found.

Come into backup...

Come into backup. .original rule: (NP-VPPI)
--== Trees in CLIST
node:SlO pos:(S) name:NIL range:(l 3) semsugzNIL
parentzNIL features:((S (MAIN))) trace:NIL

node .8811 pas: .(SS) name: @ range: (1 1) sernsug:NIL

 

 

pasta); .810 res. ((88)) trace: NIL
node: NP.12 )nameleL range:(2NIL) sem .NIL
parent:810 Futures: ((NP (NUM 3S) (SUBJ))) trace: NS.

node:NC13 pos:(NC) name:NIL range:(2 2) semsug:NIL
parern:NP12 features:((NC (NUM 18 28 38 1P 2P 3P))) trace:NIL

node: N814 pos: (NS) name: NIL rang e: (22) semsug:]‘JIL
parentzNCl3 features: ((NS (NUM IS 28 38 1P 21’ P))) trace:NIL

nodezDET15 :(DET) name .TI-IE marge: .(2 2) semsug:NIL
parartzNSM eaztures ((DET (NUM 18 38 1P 2P 3P))) trace:NIL

;odezNOUNlé pos:(NOUN) name:TEACHER range:(3 3) sernsug:NIL
parentzNPIZ features:((NOUN (NUM as)» trace:NIL

node. 817 Mros;(8) name: NIL range: (4NIL) semsug: NIL

sum eatuzres ((8 (SEC)( muoox)» trace: NIL
Jude: NP18 .(NP) name:NIL ran e: NIL sernsug:NIL
parent:817 muses eatures:((NP (DUMMY) )tracezNIL

nodezAUX19 pos:(AUX) name:NIL rangezNIL sernsugzNIL
parentzs 17 features:((AUX (PASSIVE))) trace:NIL

;odeNPZO Fos:(VP) name:NIL range:(4 NIL) semsug1NIL
parent:Sl7 eatures:((VP)) trace:NIL

node:V9 pos:(V) namezTAUGHT range:(4 4) semsug:(<abso 'I‘EACHER.1> <abso TAUGHT.2> TO 2)

191

parent:VP20 features:((V (EN) (TENSE PAST) (MAIN) (TRANSITIVE) (DATIVE) (DATIVE-PREP TO) (N UM 18 28 38 1P 2P 3P)))
«1081--
node:NP21 pos:(NP) name:NIL range:NIL semsug:NIL
parart:VP20 features:((NP (TRACE))) trace:NPlZ
node:Sl7 pos:(S) name:NIL ran e:(4 NIL) semsug:NIL
parentzN'I;12 features:((S (SEC) ( TRUC-OK») trace:NIL
-.su ..
node:NP18 pos:(NP) name:NIL range:NIL sernsugzNIL
parent:817 features:((NP (DUMMY))) trace:NIL

nodezAUX19 pos:(AUX) name:NIL rangezNIL sernsugzNIL
param817 features:((AUX (PASSIVE))) trace:NIL
node:VP20 :(VP) name:NIL range:(4 NIL) semsug:NIL
parent:817 eatures:((VP)) trace:NIL

node: V9 pos: (V) name: TAUGHTran ranzge .(4 4) sem :(<abso TEACHERJ) <abso TAUGHT.2> TO 2)

[3.13.ng features: ((V (EN) (TENS PAST) (TRANSITIVE) (DATIVE) (DATIVE-PREP TO) ))

node:NP21 pos:(NP) name:NIL rangezNIL semsugzNIL
parent:VP20 features:((NP (TRACE))) trace:NP12

Come into rule-matching part...
Return from rule-matching routine. rules: NIL
No rules are found.

--«-- justify finished.“ success --------------
Save backup pain for sec clause:(Sl7 810)

A word read in read-word: word=BY
No rules are found.

««««« justify finished.as success ..------------

A word read in read-word: wordleIE

Come into rule-matching part....

Return from rule-matching routine. rules: (PARSE-DET)

Only one rule selected = PARSE-DET
This rule executed.

Come into rule—matching part...
Retum from mic-matching routine. nrles= NIL
No mles are found.

«mm ' stify finished.“ success - -----------

A w read in read-word: word=PO0L

Come into rule-matching part...

Retum from rule-matching routine. rules: (NOUN-PARSEl)

Only one rule selected = NOUN-PARSEI
Ibis rule executed.

Come into rule-matching part...
Retum from rule-matching routine. rules: NIL
No rules are found.

-------- ' stify ﬁnished.as success .-----.--...--

A w read it: read- word: word=PASSED

Come' into rule-matching part...

Return from rule-matching routine. rules: (NOUN-PAR8E3)

Only one rule selected: NOUN-PARSE3
This rule executed.

Come into rule-matchin part...
Retum from rule-matching routine. rules=build-PP
Only one rule selected: build-PP

a Tree: in CLIST
node:SlO pos:(S) name:NIL range:(13)semsug:NIL
parentzNIL features:((S (MAIN))) trace:NIL

node:SSl] pos:(SS) narne:@ range:(l l) sesnsugzNIL
pasrstﬁESIO features:((88)) trace:NIL

node:NP12 pos:(NP) name:NIL range:(2 NIL) semsug:NIL
parent:810 features:((NP (NUM 38) (8081») trace:NIL

node:NCl3 pos:(NC) name:NIL range:(2 2) semsug:NIL
parentzNP12 features:((NC (NUM 18 28 38 1P 2P 3P))) trace:NIL

 

 

192

node:NSl4 pos:(NS) name:NIL range:(2 2) sernsug:NIL
parentzNC13 features:((NS (NUM IS 28 38 1P 2P 3P))) trace:NIL

node:DET 15 pos:(DET) name:THE range:(2 2) semsugiNIL
parartzN814 features:((DET (NUM 18 28 38 1P 2P 3P))) trace:NIL

;odezNOUN16 pos:(NOUN) namez'I'EACHER range:(3 3) sernsug:NIL
parentzNPIZ features:((NOUN (NUM as)» trace:NIL

node:Sl7 pos:(S) name:NIL range:(4 NIL) sernsugzNIL
parent:NPl2 features:((S (SEC) (STRUC~OK))) trace:NIL
«SUBL-
nodezNPl8 pos:(NP) name:NIL rangezNIL semsugzNIL
parent:817 features:((NP (DUMMY))) trace:NIL

nodezAUXl9 pos:(AUX) name:NIL rangezNIL sernsug:NIL
parent:817 features:((AUX (PASSIVE))) trace:NIL

node:VP20 pos:(VP) name:NIL range:(4 NIL) sernsugzNIL
parent:Sl7 features:((VP)) trace:NIL

;ode:V9 pos:(V) narnezTAUGHT range:(4 4) semsug:(<abso TEACHERJ) <abso TAUGHT.2> TO 2)
paIrentiVPZO features:((V (EN) (TEN SE PAST) (MAIN) (TRANSITIV E) (DATIVE) (DATIVE-PREP TO) ))
.. ()3 --
node:NP21 pos:(NP) name:NIL rangezNIL semsugzNIL
pathPZO features:((NP (TRACE))) trace:NP12
node:Sl7 pos:(S) name:NIL range:(4 NIL) semsug:NIL
parentzNPIZ features:((S (SEC) (STRUC-OK)» trace:NIL
«SUBL-
node:NP18 :(NP) name:NIL range:NIL sernsugzNIl.
parent:817 eatures:((NP (DUMMY))) trace:NIL

nodezAUXl9 pos:(AUX) name:NIL rangezNIL sernsug:NIL
parent:817 features:((AUX (PASSIVE))) trace:NIL

node:VPZ) :(VP) name:NIL range:(4 NIL) semsugzNIL
pamt:817 eatures:((VP)) trace:NIL

node:V9 pos:(V) name:TAUGHT ran e:(4 4) semsug:(<abso TEACHER.1> <abso TAUGHT.2> TO 2)
parent:VP20 features:((V (EN) (TENSE PAST) (MAIN) (TRANSITIVE) (DATIVE) (DAIIVE~PREP TO) )) trace:NIL
--IOBJ--
node:NP21 pos:(NP) name:NIL range:NIL semsugzNIL
parent:VP20 features:((NP (TRACE))) trace:NP12
nodezPP78 pos:(PP) name:NIL range:(S 7) semsugtNIL
parentzNIL features:((PP (MODIFIABLE) (N UM 3S))) trace:NIL
nodezPREP71 pos:(PREP) namezBY range:(S 5) semsugzNIL
parent:PP78 features:((PREP)) trace:NIL

node:NPTI pos:(NP) name:NIL range:(67)semsu :NIL
parent:PP78 features:((NP (NUM 3S) (MODIFIAB ))) trace:NIL

node:NC75 pos:(NC) name:NIL range:(6 6) semsugzNIL
parent:NP77 features:((NC (NUM IS 28 3S 1P 2P 3P))) trace:NIL

node:NS73 s:(NS) name:NIL range:(6 6) sernsugzNIL

parent:NC7 features:((NS (NUM IS 28 38 1P 2P 3P))) trace:NIL
nodezDET72 pos:(DET) namezTHE range:(6 6) semsugzNIL
paruitzNS73 features:((DET (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:NOUN74 pos:(NOUN) name:POOL range:(7 7) semsugzNIL
parent:NP'77 features:((NOUN (NUM 38))) trace:NIL

Come into rule-matching part...
Return from ﬂue-matching routine. rules: (VP-PP PASSIVE—BY)

come into go_on_several_niles. rules: (PASSIVE-BY VP-PP)

start of branch in go__several_rules = (PASSIVE-BY)

Suggestion received: dir=CONNECF ol=<abso TAUGHT.2> 02=<abso POOL.3> 03$UBJ whichasp= 2
two-way collisions betw <abso POOL3> and <abso TAUGHT.2> = 27

ll of paths for these two-way collisions = 562

call into secondary marker passing

Number of original three-way collisions = 0

No three way collision found.....

This launch is removed.

start of branch in go_several_rules = (VP-PP)

Suggestion received: dir=CONNECT ol=<abso TAUGHT.2> 02=<abso POOL.3> o3=BY whichaspr: NIL
two-way collisions betw <abso POOL.3> and <abso TAUGHT.b = 27

ll of paths for these two-way collisions = 562

193

call into secondary marker passing

Number of original three-way collisions = 2

number of collisions before H-heuristics = 2

drop collision at ﬁlterl: cnode=<rel LOC-ACT>

# of collisions after H-heuristic: 1

Number of full paths bef ll rel ﬁltering = 1

number of full paths after num. relation ﬁltering = 1
Number of full paths after hierarchy (I end absos = 1

Final paths selected in pickup_ interp.
<<>> (<abso PLACE> <abso TEACH> (<abso POOL.3> I <abso POOL> D <abso PLACE> C_ INV <asp LOC. LOC- TEACH> A IN V
<rel [DC -TEACH> A IN V <asp TEACH. LOG-TEACH) C__ INV <abso TEACH) I <abso TAUGHT. 2>))

==—-- Trees in CLIST -—-—-
node:SlO7 pos:(S) name:NIL range:(l 3) sernsug:NIL
parent:NIL features:((S (MAIN))) trace:NIL
node:SSIOS Fos: .(SS) narne:@ range:(l 1) sernsug:NIL
paren.su E38107 eat:ures ((88)) trace :NIL
node:NP109 :(NP) name:NIL range:(2 NIL) sernsu :NIL
parent:8107 eatures:((NP (NUM 3S) (SUBD)) trace:NI

node:NCllO pos: (NC) name :NIL ran gze (2 2) sernsu
parent:NPlO9 features: ((NC (NUM 18 828 38 IP 2P3 3%») trace: NIL

node: N81 11 pos:(NS) name :NIL range: (2 2) )sernsu;:NIL
pawn: NCl 10pm features: ((NS (NUM 18 28 38 1P 2P P))) trace:NIL

nodezDET112 :(DET) narnezTHE range:(2 2) sernsug:NIL
parartzN81ll eatures:((DET (NUM 18 28 3S 1P 2P 3P))) trace:NIL

node: NOUN113 pos: (NOUN) name: TEACHERran ME:- (3 3) sernsug:NIL
pawn: NP109 features: ((NOUN (NUM 38))) trace

node:Sl l4 pos:(S) name:NIL range:(4 NIL) sernsug:NIL
pawnzNP109 features:((S (SEC) (STRUC-OK») trace:NIL

«SUBJ—
node: NP1 15 ros: .(NP) name NIL range .:NILsemsug NIL
parent:8114 eatu:res ((NP (DUMMY))) trace: NIL

node:AUX116 pos:(AUX) name:NIL range:NIL sernsug:NIL
parent:8114 features:((AUX (PASSIVE))) trace:NIL

node:VP117 :(VP) name:NIL range:(4 7) semsugzNIL
parent:8114 eatures:((VP)) trace:NIL

node:V118 pos:(V) namezTAUGHT range:(4 4) sems :(<abso TEACHER.1> <abso TAUGHT.2> TO 2)

paIrOenBtJNPl 17 features:((V (EN) (TENSE PAST) (M (TRANSITIVE) (DATIVE) (DATIVE-PREP TO) ))

node:NP119 pos:(NP) name:NIL range:NIL sernsug:NIL
parent:VP117 features:((NP (TRACE))) trace:NP12

node:PP1(X).(PP) name. NILran e.(5 7) sansug:NIL
parent:VP1 lﬁeatures eatures:((PP (MODIFIEABLE) (NUM 38))) trace:NIL

node:PREP101 pos:(PREP) namezBY range:(S 5) sernsug:NIL
parent:PP1(X) features:((PREP)) trace:NIL

node:NP102 pos:(NP) name:NIL range:(6 7) sernsuf‘éabso TAUGHT. 2> <abso POOL.3> BY NIL)
pawn:PP1(X) features:((NP (NUM 38) (MODIFIAB ))) trace: NIL

node:NC 103 pos:(NC) name:NIL range:(6 6) sernsug:NIL
parent:NPlO2 features:((NC (NUM 18 28 3S 1P 2P 3P))) trace:NIL

node:N8104 pos: (N8) name :NIL rang e: (6 6) sernsug:NIL

pawn: NC103 features. ((NS (NUM 188 28 3S 1P 2P P))) trace :NIL
nodezDETIOS Fos: .(DET) name: THE rang e: (6 6) semsug :NIL
parent:NSlO4 tures: ((DET (NUM 18 28 38 1P 2P 3P))) trace: NIL

node:NOUN106 pos:(NOUN) name:POOL range:(7 7) sernsug:NIL
parent:NP102 features: ((NOUN (NUM 38))) trace:NIL
node: 81 14 (.8) name: NIL range: (4 NIL) semsug: NIL
parersttU 1:13PM?s eazmres ((8 (SEC) (STRUC- OK))) trace: NIL
node:NP115.(NP) name: NIL range: NIL semsug :NIL
parent: 81 14 F05: eature:s ((NP (DUMMY))) trace :NIL

node: AUXl 16 pos:(AUX) name: NIL rang e: NIL semsug :NIL
paratt:8114 features: ((AUX (PASSIVE)))8 trace: NIL

194

;ode:W117 for: .(VP) name:NIL range: (4 7) sernsug:NIL

parart:81 14 eatures(:(VP)) trace:NIL
node:V118 pos:(V) narnezTAUGHT range:(4 4) semsu :(<abso TEACHERJ) <abso TAUGHT. 2) TO 2)
wwr 17 features:((V (EN) (TENSE PAST) ) (TRANSITIV E) (DATIVE) (DATIVE-PREP TO) ))

node:NP119 pos:(NP) name:NIL rangezNIL sernsug:NIL
parent:VP117 features:((NP (TRACE))) trace:NPl2

node:PP100 ros: .(PP) name. NIL range. (5 7) semsug :NIL
parern:VP11 featu eatu:res ((PP (MODIFIABLE) (NUMg 3S))) trace:NIL

node:PREP101 pos:(PREP) name:BY range:(5 5) sernsug:NIL
pawtt:PP100 features:((PREP)) trace:NIL

node:NP102 pos:(NP) nune:NIL range:(6 7) semsug: .(<abso TAUGHT.2> <abso POOL.3> BY NIL)
parentzPPIOO features:((NP (NUM 38) (MODIFIAB E))) trace :NIL

node: NC103 pos:(NC) name: NIL rang e: (6 6) semsug: NIL
pamePlOl features. ((NC (NUM 18828 38 1P 2P 3P))) trace:NIL

node:N8104pos:(NS) name. NIL rang e: (6 6) sernsug:NIL
parentzNC103p08 features. ((NS (NUM 18 28 38 1P 2P P))) trace: NIL

node: DET105 pos:(DET) name. THEran marge: .(6 6) sernsug:NIL
parent :N8104 features: ((DET (NUM 18 3S 1P 2P 3P))) trace:NIL

node:NOUN106 pos:(NOUN) namezPOOL range:(7 7) sernsug:NIL
parentzNP102 features:((NOUN (NUM 38))) trace:NIL

beginning of NEXT_round.

Come into rule-matching part

No rules matched. (just after NEXT _ROUNDI)

------- justify ﬁnished. as success ..-----.----..

End of one rule-process for-.(VP-PP) merit-(<abso PLACE> <abso TEACH) (<abso POOL.3> I <abso POOL> D <abso PLACE> C_INV
<asp LOCLOC-I'EACIb A_INV <rel LOG-TEACH) A_INV <asp 'I‘EACH.LOC-TEACH> C_INV <abso TEACH> I <abso TAUGHT.2>»

startofnmparalleltilloneisforrrrd.#ofbranches:1
if of sern bmches: 1 0 of no sem branches: 0

Backed up branches into ‘back -stack‘ :
Branches to continue to run : -PP)

FINAL One branch chosen after parallel execution: (VP-PP)
only one branch chosen in parallel.

Come rnto rule-matching part...
Retum from rule-matching routine. rules: NIL
No rules are found.

-------- justify ﬁnished.as success ---..---..----
Save backup point for sec clause:(8114 8107)

an object read tn from cash: V120
Come into rule—matching part.
Return from rule-matching routine. rules: (NP- -VPP)

Only one rule selected : NP-VPP

Suggestim received: dir:CONNECI' 01=<abso POOL.3> ok<abso PASSED.4> o3=OBJ whichasp: 2
two-way collisions betw <abso PASSED.4> and <abso POOL.3> : 17
# of paths for these two-way collisions : 356

call into secondary marker passing

Number of original three-way collisions : 1

number of collisions before H-heuristics : 1

# of collisions after H-heuristic: 1

Number of full paths bef # rel ﬁltering : 1

number of full paths after num. relation ﬁltering- - 1

Number of full ths after hierarchy of end absos- - 1

Filter paths havrng too _abstract _...conccpt

Final paths selected tn pickup_ interp: no paths found.

Come into backup...

Come into rule-matching part...

Retum from rule-matching routine. rules: NIL
No rules are found.

4.---.. justify ﬁnished.as success --------------

195

anobjectreadinfromcash=V141
Come into rule-matching part...
Return from rule-matching routine. rules: (MAIN -VERB2)

Only one rule selected : MAIN-VERB2

Suggestion received: dim-CONNECT ol=<abso TEACHERJ) o2=<abso PASSED.4> o3=SUBJ whichasp: NIL
two-way collisions betw <abso PASSED.4> and <abso TEACHER. l) : 30
it of paths for these two-way collisions : 1048

call into secondary marker passing

Number of original three-way collisions : 3

number of collisions before H-heuristics : 3

drop collisiat at ﬁlterl: cnode=<rel ACTOR-ACT)

drop collision at ﬁlter]: cnode=<rel ACI‘OR-TRSACD

# of collisions after H—heuristic: 1

Number of full paths bef # rel ﬁltering : 1

number of full paths after num. relation ﬁltering : 1

Number of full paths after hierarchy of end absos : 1

Final paths selected in pickup_interp.

<<>> (<abso PASS) <abso ANIMAL> (<abso PASSED.4> I <abso PASS) C_INV PASS.PASSER-PA88) A_INV <rel PASSERPASS)
A_INV <asp PASSER.PASSER-PASS> C_INV <abso ANIMAL> D <abso HIGl-l_ ATE> D <abso PERSON> D

<abso TEACH_PROFESSIONAL> D <abso TEACHER) I <abso TEACHER.1>))

Trees in CLIST m
node:8121 pos:(S) name:NIL range:(l 8) sernsug:NIL
parentzNIL features:((S (MAIN))) trace:NIL

node:88122 pos:(SS) name@ range:(l 1)sernsug:NIL
$638121 features:((SS))trace:NIL

node:NP123 pos:(NP) name:NIL range:(2 NIL) sernsug:NIL
pawrt:8121 features:((NP (NUM 38) (SUBJ))) trace:NIL

node:NC124 pos:(NC) nune1NIL range:(2 2) sernsug:NIL
parentzNP123 features:((NC (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:N8125 pos:(NS) name:NIL range:(2 2) sernsug:NIL
parent:NC124 features:((NS (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:DET126 pos:(DET) name:THE range:(2 2) sernsug:NIL
pararuN8125 features:((DET (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:NOUN127 pos:(NOUN) name:TEACHER range:(3 3) sernsug:NIL
pawn:NPl23 features:((NOUN (NUM 3S))) trace:NIL

node:8128 pos:(S) name:NIL range:(4 NIL) sernsug:NIL
pagnﬁlgglﬁ features:((S (SEC) (STRUC-OK) (CLOSED))) trace:NIL
node:NP129 pos:(NP) name:NIL range:NIL sernsug:NIL
pawrt:8128 features:((NP (DUMMY))) trace:NIL
node:AUXl30 pos:(AUX) name:NIL range:NIL semsug:.NIL
parent:8128 features:((AUX (PASSIVE))) trace:NIL

node:VP131 pos:(VP) name:NIL range:(4 7) sernsug:NIL
pawn:8128 features:((VP)) trace:NIL

 

 

node:V132 pos:(V) name:TAUGI-IT range:(4 4) sernsug:(<abso TEACHERJ) <abso TAUGHT.2> TO 2)

palroaétivmﬂ features:((V (EN) (TENSE PAST) (MAIN) (TRANSITIV E) (DATIVE) (DATIVE-PREP TO) ))

node:NP133 pos:(NP) name:NIL range:NIL sernsug:NIL
parent:VP131 features:((NP (TRACE))) trace:NP12

node:PP134 pos:(PP) name:NIL range:(5 7) sernsug:NIL
parent:VPl3l features:((PP (MODIFIABLE) (NUM 38))) trace:NIL

node:PREP135 pos:(PREP) name:BY range:(5 5) sernsug:NIL
parent:PP134 features:((PREP)) trace:NIL

node:NPl36 pos:(NP) name:NIL range:(6 7) semsug:(<abso TAUGHT.2> <abso POOL.3> BY NIL)
parent:PP134 features:((NP (NUM 38) (MODIFIABLE))) trace:NIL

node:NC137 pos:(NC) name:NIL range:(6 6) sernsug:NIL
parentzNP136 features:((NC (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:N8138 :(NS) name:NIL range:(6 6) sernsug:NIL
pawrt:NC13 features:((NS (NUM 18 28 38 1P 2P 3P))) trace:NIL

nodezDET139 pos:(DET) name:THE range:(6 6) sernsug:NIL
parentzNSl38 features:((DET (NUM 18 28 38 1P 2P 3P))) trace:NIL

196

node:NOUN140 pos:(NOUN) namezPOOL range:(7 7) semsugzNIL
pawrtzNP 136 features:((NOUN (NUM 38))) trace:NIL

node :AUX147 pos: (AUX) name:NIL range:NIL sernsug:NIL
parun:8121 features: ((AUX)) trace: NIL

node .VP148 pos:(VP) name:NIL ran ge:(8 8) sernsug:NIL
parent:8121 features:((VP (NUM 18 28 38 1P 2P 3P) (TRANSITIVE) (EN) (TENSE PAST) (MAIN))) trace:NIL

node:V141 pos:(V) namezPASSED range:(8 8) sernsug:(<abso TEACHER. 1) <abso PASSED.4> SUBJ NIL)
pawn:VP148 features:((V (MAIN) (TENSE PAST) (EN) (TRANSITIVE) (N UM 18 28 38 1P 2P 3P))) trace:NIL

Come into rule-matching part...
Return from mic-matching routine. ntles= NIL
No rules are found.

-------- {I EEK ﬁnished.as success --------------
next word : THE

A word read in read-word: word=THE
Come into rule-matching part...
Return from ntle-matching routine. rules: (PARSE-DET)

Only one rule selected : PARSE-DET
This rule executed.

Come into rule-matching part..“
Return from mle-matching routine. rules: NIL
No rules are found.

-------- justify ﬁnished.” success --------------
RRRRRRRRRead next word : TEST

passing markers. instance=<abso TEST.5>

this is regular collision at <abso GENERAL_I'EST>.

Bump into node carrying same ori mark at cnode=<abso STATE)
this is regular collision at <rel STEAL-ER).

this is regular collision at <rel AG-ACQUIRE).

this is regular collision at <rel AG-DISCUSS).

A word read' in read-word: word=TE8T
Come into rule-matching part.
Retum from nrle-matching routine. ntles= (NOUN-PARSEI)

Only one rule selected : NOUN-PARSEI
This rule executed.

Come into rule—matching part...
Return from mle-rnatching routine. rules: NIL
N 0 rules are found.

----- - justify ﬁnished.“ success --------------
RRRRRRRRRead next word : @.

A word read in read-word: word=@.
Come into rule-matching part...
Return from rule-matching routine. rules: (NOUN-PARSE3)

Only one rule selected : NOUN-PARSE3

Trees in CLIST a
node:8121 pos:(S) name:NIL range:(18)semsug:NIL
parent:NIL features:((S (MAIN))) trace:NIL

 

 

node:88122 :(SS) nune:@ range:(l 1) sernsug:NIL
$538121 eatures:((88)) trace:NIL

node:NP123 s:(NP) name:NIL range:(2 NIL) semsu :NIL
parurt:812l eatures:((NP (NUM 38) (SUBD)) trace:NI

node:NC124 pos: (NC) name:NIL ran gze (2 2) sernsug:NIL
parentzNP123 features: ((NC (NUM 18828 38 1P 2P 3 ))) trace: NIL

node: N8125 pos:(NS) name: NIL range: (22)sernsu sernsug:NIL
parentzNC124p°s features: ((NS (NUM 18 28 38 1P 2P P))) trace:NIL

nodezDET126 :(DET) name: THE ran e:(22) semsug :NIL
parentzN8125 eaztures ((DET (NUM 18 38 1P2P 3P))) trace. NIL

node .NOUN127 pos: (NOUN) nune. TEACHERrarr rNIIf: .(3 3) sernsug:NIL
parent:NPl23 features: ((NOUN (NUM 38))) trace:

197

node: 8128 gs: .(S) name:NIL range: (4 NIL) sernsug:NIL

par-enstuzNBP13 features: ((8 (SEC) (STRUC-OK) (CLOSED))) trace:NIL
node: NP129 :(NP) name:NIL range:NIL sernsug:NIL
parent:8128 eatures:((NP (DUMMY))) trace:NIL

node:AUX130 pos: (AUX) name:NIL ran g:.eNILsernsug NIL
pawrt: 8128 features. ((AUX (PASSIVE))) trace .NIL

node:VP13l 1pos: .(VP) name:NIL range:(4 7) sernsug:NIL
parent: 8128 tuzres ((VP)) trace .NIL

node: V132 s. (V) name: TAUGHT range: (4 4) semsug: (<abso TEACHER. 1) <abso TAUGHT. 2) TO 2)
paratt: VP131 features. ((V (EN) (TENSE PAST) (MAIN) (TRANSITIVE) (DATIVE) (DATIVE-PREP TO) ))
«IOBJ—

node:NP133 pos:(NP) name:NIL range:NIL sernsug:NIL

pawn:VPl31 features:((NP (TRACE))) trace:NP12

node:PP134 pos:(PP) name:NILran m6 7)sernsug :NIL
parent:VP131 features:((PP (MOD BLE) (NUM8 3S))) trace: NIL

;odezPREP135 pos:(PREP) name: BY range: (5 5) sunsug'NIL
parent:PPl34 features: ((PREP))tra cezNIL

node:NPl36 pos:(NP) name:NIL range:(6 7) sernsuEé<abso TAUGHT. 2) <abso POOL. 3) BY NIL)
parent:PP134 features:((NP (NUM 3S) (MODIFIAB ))) trace: NIL

node: NC137 pos: (NC) name: NIL range: (6 6) semsug :NIL
parent:NP136 features. ((NC (NUM 18 828 3S 1P 2P 3P))) trace .NIL

node:NSl387pos: .(NS) name:NIL rang e.( (66) sernsug:NIL
parent:NC13 features: ((NS (NUM 18 28 38 1P 2? P))) trace. NIL

node .DET139 pos: (DET) name: THE nrzrge: ()6 6) semsug :NIL
parent .N8138 features: ((DET (NUM 18 38 1P 2P 3P))) trace: NIL

node:NOUNl40 pos:(NOUN) namezPOOL range:(7 7) sernsug:NIL
parent1NPl36 features:((NOUN (NU M 38))) trace:NIL

nodezAUX147 pos:(AUX) name:NIL range:NIL sernsug:NIL

parent:8121 features:((AUX)) trace:NIL

node:VP148 (VP:) name: NIL range: (8 8) semsug :NIL

parent:8121eat:ures ((VP (NUM 18 28 38 1P 2P 3P) (TRANSITIVE) (EN) (TENSE PAST) (MAIN))) trace :NIL

node:V141 pos:(V) name:PASSED range:(8 8) semsu :(<abso TEACHERJ) <abso PASSED.4> SUBJ NIL)
pawn:VP148 features:((V (MAIN) (TENSE PAST) (TRANSITIVE) (N UM 18 28 38 1P 2P 3P))) trace:NIL
node:NP154 pos:(NP) name:NIL range:(9 10) sernsug:NIL
parent:NIL features:((NP (NUM 38) (MODIFIABLE))) trace:NIL

node:NC152 pos:(NC) name:NIL range:(9 9) semsugzNIL

parent:NP154 features:((NC (NUM 18 28 3S 1P 2P 3P))) trace:NIL
node:NSlSO pos:(NS) name:NIL range:(9 9) sernsug:NIL
pawnzNC152 features:((NS (NUM 18 28 3S 1P 2P 3P))) trace:NIL

node:DET149 pos:(DET) name:THE range:(9 9) sernsug:NIL
parent:N8150 features:((DET (NUM 18 28 3S 1P 2P 3P))) trace:NIL

;ode:NOUN151 pos:(NOUN) name:TEST range:(lO 10) sernsug:NIL
parent:NP154 features:((NOUN (NUM 38)» trace:NIL

Come into mic-matching part...
Retum from nae-matching routine. rules: (OBJECT-RULE)

Only one rule selected : OBJECT -RULE

Suggestion received: dir=CONNECI ol:<abso PASSED.4> o2=<abso TEST.5) 03:08] whichasp: NIL
two-way collisions betw <abso TEST .5) and <abso PASSED.4> : 8

4! of paths for these two-way collisions : 171

call into secondary marker passing

if of collisions after H-heuristic: 1

Number offull paths be“! rel ﬁltering :1

number of full paths after nurn. relation ﬁltering : 1

Number of full paths after hierarchy of end absos : 1

Final paths selected in pi cig'tgsinterp.
<<>> (<abso GENERAL_ '1‘) <abso PASS) (<abso TEST. 5) I <abso TEST) D <abso GENERAL_ TEST) C_ INV <asp
PASSEE. PASSEE-PASS) A_ INV <rel PASSEE- PASS) A_ INV <asp PASS. PASSEE-PASS) C_ INV <abso PASS) I <abso PASSED. 4>))

...-:Trees' tn CLIST:=

I98

node:8121 pos:(S) name:NIL range:(l 8) sernsug:NIL
parentzNIL features:((S (MAIN) (STRUC—OK») trace:NIL

node:88122 pos:(SS) narne:@ range:(l 1) sernsug:NIL
paratt:812l features:((88)) trace:NIL

«SUBJ--

node:NPl23 pos: (NP) name: NIL range: (2 NIL) semsug :NIL
parutt:8121 features: ((NP (NUM 38) (SUBJ))) trace:NIL

node:NC124 pos:(NC) name:NIL range:(2 2) sernsug:NIL
pawrtzNP123 features:((NC (NUM 18 28 3S 1P 2P 3P))) trace:NIL

;ode:N8125 pos:(NS) name:NIL range:(2 2) semsug:NIL
parent:NC124 features:((NS (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:DET126 pos:(DET) name:THE range:(2 2) sernsug:NIL
parutt:N8125 features:((DET (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:NOUNl27 pos:(NOUN) name:TEACHER mge:(3 3) sernsug:NIL
pawn:NP123 features:((NOUN (N UM 3S))) trace:NIL

node:8128 pos:(S) nachIL range:(4 NIL) sernsug:NIL

pawsnéglﬁ features:((S (SEC) (STRUC—OK) (CLOSED))) trace:NIL
node:NP129 pos:(NP) name:NIL range:NIL sernsug:NIL
parent:8128 features:((NP (DUMMY))) trace:NIL

node:AUXl3O pos:(AUX) name:NIL range:NIL sernsug:NIL
pawtt:8128 features:((AUX (PASSIVE))) trace:NIL

node:VPl31 pos:(VP) name:NIL range:(4 7) sernsug:NIL
pawtt:8128 features:((VP)) trace:NIL

node:Vl32 pos:(V) name:TAUGHT range:(4 4) sernsug:(<abso TEACHERJ) <abso TAUGHT.2> TO 2)
parent:VP131 features:((V (EN) (TENSE PAST) (MAIN) (TRANSITIV E) (DATIVE) (DATIVE-PREP T0) ))
--IOBJ-

nodezNP133 pos:(NP) name:NIL range:NIL sernsug:NIL

paratt:VP131 features:((NP (TRACE))) trace:NP12

node:PP134 pos:(PP) name:NIL range:(5 7) sernsug:NIL
parent:VPl3l features:((PP (MODIFIABLE) (N UM 3S))) trace:NIL

node:PREP135 pos:(PREP) namezBY range:(5 5) sansug:NIL
parent:PP134 features:((PREP)) trace:NIL

node:NP136 pos:(NP) name:NIL range:(6 7) sernsug:(<abso TAUGHT.2> <abso POOL3> BY NIL)
pawnzPP134 features:((NP (NUM 3S) (MODIFIABLE))) trace:NIL

node:NCl37 pos:(NC) name:NIL range:(6 6) sernsug:NIL
parentzNPl36 features:((NC (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:N8138 s. (NS) name: NIL range: (6 6) semsug: NIL
parent:NC13 fea ttrres: ((NS (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:DET139 pos:(DET) name:THE range:(6 6) sernsug:NIL
pawnzN8138 features:((DET (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:NOUN140 pos:(NOUN) name:POOL range:(7 7) semsug:NIL
parentzNP136 features:((NOUN (NU M 38))) trace:NIL

nodezAUX147 pos:(AUX) name:NIL range:NIL sernsug:NIL

parent:8121 features:((AUX)) trace:NIL

node: VP148 pos: (VP) name. NIL range: (8 10) semsug :NIL

parent: 8121 features: ((VP (NUM 18 28 38 1P 2P 3P) (TRANSITIVE) (EN) (TENSE PAST) (MAIN))) trace: NIL
node:V141 pos:(V) name:PASSED range:(8 8) sernsug:(<abso TEACHER. I) <abso PASSED.4> SUBJ NIL)
”3;:W148 features:((V (MAIN) (TENSE PAST) (EN) ('I'RANSIIIVE) (NUM 18 28 38 1P 2P 3P))) trace:NIL
node:NP154 pos:(NP) name:NIL range:(9 10) sernsug:(<abso PASSED.4> <abso TEST .5) QB] NIL)
pawn:VP148 features:((NP (NUM 3S) (MODIFIABLE))) trace:NIL

node:NC152 pos:(NC) name:NIL range:(9 9) sernsug:NIL
parentzNP154 features:((NC (NUM 18 28 38 1P 2P 3P))) trace:NIL

node:NSISO pos:(NS) name:NIL range:(9 9) semsugzNIL
parcntzNC152 features:((NS (NUM rs 23 as IP 2? 3?)» trace:NIL

node:DET149 pos:(DET) name:THE range:(9 9) sernsug:NIL
parent:N8150 features:((DET (NUM 18 28 38 1P 2P 3P))) trace:NIL

199

node:NOUN151 pos:(NOUN) name:TEST range:(lO 10) sernsug:NIL
pawttzNP154 features:((NOUN (NU M 38))) trace:NIL

Come into rule-matching part...
Return from rule-matching routine. rules: NIL
No rules are found.

------- justify ﬁnished.as success -------------
an object read in from cash : FPUNC153
Come into rule-matching part...

Return from rule-matching routine. rules: (S-CLOSE-AFF)
Only one rule selected = S-CLOSE-AFF

3 Trees in CLIST

,' This tree is the ﬁnal syntactic representation of the input sentence.
node:8121 pos:(S) name:NIL range:(l 11) sernsug:NIL
parent:NIL features:((S (MAIN) (STRUC-OK)» trace:NIL

 

 

node:88122 :(SS) narne@ range:(l 1) sernsug:NIL
prtrcrtSU B38121 eatures:((88)) trace:NIL

node:NP123 :(NP) name:NIL ran ge:(2NIL)semsu :NIL
parent:8121eatzures ((NP (NUM 3S) (8UBJ))) trace .NI

node :NC124 pos: (NC) name:NIL rang c: (2 2) semsug
pawn:NP123 features: ((NC (NUM 18 s28 38 1P 2P 3P))) trace: NIL

node: N8125 pes: (N8) name: NIL ran g:e (2 2) sernsug:NIL
parent:NC124 features: ((NS (NUM 188 28 38 1P 2P P))) trace: NIL

node: DET126 po:s (DET) name: THEran ranzge: .(22) )sernsug :NIL
parmt:NSl25 features: ((DET (NUM 18 3S 1P 2P 3P))) trace: NIL

node:NOUN127 pos: (NOUN) name. TEACHERran NIit: .(3 3) semsug :NIL
parent:NP123 features. ((NOUN (NUM 38))) trace

node: 8128 {on .(S) name: NIL rang“: NIL) semsug :NIL
pawnU:NP13 features: ((8 (SEC) ( UC-OK) (CIDSED))) trace. NIL

node:NP129 :(NP) name:NIL range:NIL sernsug:NIL
parent:8128 eatures:((NP (DUMMY))) trace:NIL

nodczAUX130 pos:(AUX) name:NIL range:NIL sernsug:NIL
parent:8128 features:((AUX (PASSIVE))) trace:NIL

node:VP131 res: .(VP) name:NIL range:(47) sernsug:NIL
parent:8128 tu:res ((VP)) trace: NIL

node:Vl32 pos:(V) name:TAUGHT rang). e:(4 4) semsu :(<abso TEACHER.1> <abso TAUGHT.2> T0 2)
ﬁriwrsr features:((V (EN) (TENSE

node:NP133 pos:(NP) name:NIL range:NIL sernsug:NIL
parent: VP131 features:((NP (TRACE))) trace .NP12

node:PP134 pes: (PP) name:NIL ran g:ge(57)sernsu :NIL
parent: VP131 features. ((PP (MODIFIABLE) (NUMg 38))) trace' .NIL

node :PREP135 pos:(PREP) name: BY range:(5 5) sernsug:NIL
pawn: PP134 features: ((PREP)) trace:NIL

node NP 136 pos: (NP) name:NIL range.( (6 7) semsug-(<abso TAUGHT. 2) <abso POOL.3> BY NIL)
pawn PP:134 features: ((NP (NUM 38) (MODIFIAB ))) trace: NIL

node: NC137 pes: (NC) name. NIL range: (6 6) sernsug>NIL
parent:NP136 features: ((NC (NUM 18 828 38 1P 2P 3 ))) trace :NIL

node:N8138 res: .(NS) name: NIL range: (6 6) semsu§:NIL
parent:NC13 tures: ((NS (NUM 188 28 3S 1P 2P P))) trace .NIL

node:DET139 :(DET) name. THE c:(6 6) g:NIL
parent:N8138 eaztures ((DET (NUM 18 638 1P 2P 3P))) trace :NIL

node:NOUN140 pos:(NOUN) namezPOOL range:(7 7) sernsug:NIL
parent:NP136 features:((NOUN (NUM 3S))) trace:NIL

nede:AUX147 pos:(AUX) name:NIL range:NIL sernsug:NIL
parent:8121 features:((AUX)) trace:NIL

node: VP148 (VP:) name: NIL range e:(8 10) semsug
parun:8121eaturcs:((VP (NUM 18 28 3S 1P 2P 3P) (TRANSITIVE) (EN) (TENSE PAST) (MAIN))) trace: NIL

AST) (M (TRANSITIVE) (DATIVE) (DATIVE-PREP TO)))

node: V141 pos:(V) name: PASSED range: (8 8) semsug: (<abso TEACHER. 1) <abso PASSED. 4) SUBJ NIL)
paws}: .VP148 features: ((V (MAIN) (TENSE PAST) (EN) (TRANSITIVE) (NUM 18 28 3S 1P 2P 3P))) trace: NIL
..03 --

node:NP154 pos:(NP) name:NIL range:(9 10) semsug:(<abso PASSED.4> <abso TEST .5) OBI NIL)
pawn:VP148 features:((NP (NUM 3S) (MODIFIABLE))) trace:NIL

node: NC152 pes: (NC) name: NIL range: (9 9) semsug):NIL
parent:NP154 features: ((NC (NUM 18 28 3S 1P 2P3 ))) trace:NIL

node:N8150 pos:(NS) name:NIL range:(9 9) sernsug:NIL
parent:NC152 features:((NS (NUM 18 28 3S 1P 2P 3P))) trace:NIL

node:DET149 tpes: .(DET) name. THEran rarztge: .(9 9) sernsug:NIL
parent:NSlSO tures: ((DET (NUM 18 38 1P 2P 3P))) trace:NIL

node:NOUN151 pos:(NOUN) name:TEST range:(lO 10) sernsug:NIL
parent:NP154 features:((NOUN (NUM 3S))) trace:NIL

node: FPUNC153 pes: (FPUNC) name:@m range: (11 11) sernsug:NIL
paran: 8121 features: ((FPUNC)) trace:NIL

Come into mic-snatching part...

Retum from ntlc-matching routine. rules: NIL
No rules are found.

-------- justify ﬁnished.” success ---—----------
processing one sentence has ﬁnished.

> (print_semi)

Semantic Interpretation ----- )
,',',' This is a set of knowledge base paths found during the parsing.

(<abso TEST. 5) I <abso TEST) D <abso GENERAL_ TEST) C_ INV <asp PASSEE PASSEE- PASS) A _INV
<rel PASSEE-PASS) A_ INV <asp PASS. PASSEE -PASS) C_ INV <abso PASS) I <abso PASSED. 4))

(<abso PASSED.4> I <abso PASS) C_INV (“KNIMSS .PASSER-PASS) A _INV <rel PASSER- PASS) A _INV
<asp PASSER.PASSER-PASS> C_INV <abso AL) D <abso HIGH _ANIMATE) D <abso PERSON> D
<abso TEACH_PROFESSIONAL> D <abso TEACHER) I <abso TEACHER. 1))

(<abso POOL.3> I <abso POOL> D <abso PLACE> C_INV <asp LOC.LOC-TEACH> A_INV <rel LOG-TEACH)
A_INV <asp TEACH.LOC-TEACH) C_INV <abso TEACH) I <abso TAUGHT.2>)

(<abso TAUGHT.2> I <abso TEACH) C_INV <asp TEACHREC-TEACH) A _INV <rel REC-TEACH) A_INV

<asp RECREC-TEACH) C_INV <abso PERSON> D <abso TEACH_PROFESSIONAL> D <abso TEACHER) I <abso TEACH ER. 1 >)

> (dribble)

Appendix E

List of Sample Rules

The rules that have been mentioned in Chapter 4 and 6 in this thesis will be listed

in this section.

‘a ‘a ‘a ‘a ‘a ‘a ‘a ‘a ‘a \a \a ‘a ‘a ‘a ‘a

\a

a
I

a
I

I
I

V

In the pattern part, the base pattern is indicated by the flag 1

and the raw pattern is indicated by the flag 0.

Because the base pattern should be able to specify a portion of a tree,

it might be necessary to use more that one node chain elements. For example,
in the "passive-by" rule, the base pattern consists of three node chain
elements, "[(np-l 5-1) 1 ()] [(aux-l s-1) 1 (passive)] [(vp-l s-1) 1 ()]".
This is to specify the following shape of a portion of a tree:

8

/ l \

/ I \
NP AUX VP

The first element of each rule is the rule name which should be unique.

This rule creates a PP.

(build-pp ((1 )(h dative-dobj‘PP1)

[ l [ (PIEP-I) 0 () l
[ (np-l) 0 l) l )
( ) l
[ (create pp-l)

(attach prep—1 to pp—l)

(attach np-l to pp-l)

(transfer-features np-l to pp-l)
]

This rule analyzes the main verb of a clause.
(main-verbZ ()

[ ( [ (np-l s-1) 1 ((link subj)) ]
[ (v-1) 0 (main (not aux)) ] )

( [equal (feature—of num np-l) (feature-of num v-l)]

) l
[

(create aux-l)

(attach aux-1 to s-l)

(create vp-l)

(attach v-l to vp-l)

(transfer-features v-l to vp—l)

(attach vp-l to s-l)

(sem connect np-l v-l subj)

(if (have-feature intransitive in v-l)

(add-feature struc-ok to s-l))
; if *dummy_prop_instance* is set by the that-complement rule,
; then the proposition instance in this var is connected to
; the main verb of the secondary clause in the global var
; called *connections sec clause*.
onnect_main_sec2 v-1)— —

o ‘0 ‘0 ‘g \.

H
A

This rule analyzes the NP which starts with a noun.

(ex) "apples"

(noun-parseO ()

201

I

‘0 ‘0 ‘0

so ‘0 ‘0

‘g ‘0 ‘0

[ ( [ (noun-1) 0 () l

)

( [no-ns-na-nc-noun-2] )

l

[ (push-to-cash-last—node) ; to put new nc in front of noun-l.
(create nc-Z on clist) ; this will be dummy nc containing nothing.
(bring-from-cash) ; to get noun-l after new nc.

]

This rule creates an NC when a noun directly follows the NS.
(noun-parsel ()

[ ( [ (ns-l) 0 () l
[ (noun-1) 0 () I )
U l
[ (create—mother nc-l of ns—l)
(transfer-features ns-l to nc-l)

]

This rule identifies the head noun of an NP by noticing that it is
followed by some non-noun element.

(ex) "the book is..."

(noun-parse3 ((l which-diagZ parse-det pronoun-rule)(h ))

[ ( [ (nc—l) 0 () l
[ (noun-l) O () ] ; this noun becomes the end of hp.
[ (not noun) 0 ((word)) ] ) ; this gives the enough context.
( ) l
[ (push-to-cash-last-node)
(create np-l on clist)
(attach nc-l to np-l)
(transfer-features nc-l to np-l)
(attach noun-1 to np-l) ; noun-l is the end of np.
(sem connect nc-l noun-l) ; nc-l alone means the last noun under it.
(sem connect (na nc-l) noun-1 adj-mod); na alone means the last adj underit.
(transfer-features noun-l to np-l)
(if (have-feature wh in np-l)
(bind-whcomp np-l) : then
(add-feature modifiable to np-l)) : else
; deleted from clist by bind-whcomp. specific rules will do.

1

This rule identifies the head noun of an NP by noticing that
a noun with plural form should be the head noun.

(ex) ”the books ..."

(noun-parse33 ()

[ ( [ (nc—l) 0 () l
[ (noun-1) O ((have-feature num 3p) (not sing-plu-same)) ]
;; sing—plu-same indicates that the noun can be both singular
;; and plural. (ex) fish, sheep.
)
U l
[ (create np-l on clist)
(attach nc-l to np-l)
(transfer-features nc-l to np-l)
(sem connect nc-l noun-1)
(sem connect (na nc—l) noun-1)
(attach noun-1 to np-l)
(transfer-features noun-1 to np-l)
(if (have-feature wh in np-l)
(bind-whcomp np-l) ; then
(add-feature modifiable to np-1)) ; else
; deleted from Clist by bind-whcomp. specific rules will do.

This rule analyzes a non-"head noun" in the nominal compound by noticing
that a noun(of singular form) is immediately followed by another noun.

(ex) "the book shelf ..."

(noun-parse4 ()

;;; accomodate noun-1 as part of the np being built.
[ ( [ (nc-l) O () l
[ (noun—l) O ((have-feature num 35)) ] ; becomes modifier of hp.
[ (noun-2) O () ] ; this will become part of the np being built.
)
U
]
[ (sem connect nc-l noun-l) ; it is not clear if this suggestion is
; necessary. more study of nominal compound is required.
(attach noun-1 to nc-l) ]

This rule determines that a noun followed by a plural noun can be
the head noun of a NP and the plural noun will build another NP.
(ex) "the fish eggs ..."
(noun-parseS ()
[ ( [ (nc-l) 0 () l
[ (noun-1) 0 ((have-feature num 35)) ) ; this is the end of np built.
[ (noun-2) 0 ((have—feature num 3p)) ] : if it is not plural form,
; an article(a, the) should come to lead relative cl.
; noun-2 becomes a new np.

)

‘0 ‘0 ‘0

0 l
[ (push-to-cash-last-node)
(create np-l on clist)
(attach nc-l to np-l)
(transfer-features nc-l to np-l)
(sem connect nc-l noun-l)
(sem connect (na nc-l) noun-1)
(attach noun-1 to np-l) ; this is the head noun of np-l.
(transfer-features noun-1 to np-l)
(if (have-feature wh in np-l)
(bind-whcomp np-l) ; this is then part.
(add-feature modifiable to np-1)) ; this is else part.

This rule attaches a PP to an NP which immediately preceding it.
(ex) "a book on the desk"
(np-pp ((l )(h passive-by))
[ ( [ (np-l low-right) 1 (modifiable) ]
[ (PP-1) 0 (l l )
U l
[ (attach pp-l to np-l)
(if (have-feature trace in pp-l)
(sem connect (np pp-l) np-l (prep pp-l) 2) ; if wh object is under pp.
(sem connect np-l (np pp-l) (prep pp-l)) )
(remove-feature modifiable from np-l) ]

‘3 ‘g

This rule analyzes the reduced relative clause when an NP is followed by
a verb with past participle form(the NP is analyzed as the direct object
of the verb.

(ex) "the book torn ..."

V. ‘0 ‘0 \.

(np-vpp ()
[ ( [ (np-l low-right) 1 (modifiable) ]
[ (v-1) 0 (en transitive) ] ; has feature of past participle

)
( ) ]

[ (create s-l in clist)

(add-feature sec to s-l)

(attach s-l to np-l implicitly)
(create np-2)

(add-feature dummy to np-2)

(attach np-2 to s-l as subj)

(create aux-l)

(add-feature passive to aux-l)
(attach aux-l to s-l)

(create vp-l)

(attach v-l to vp-l)

(create np-3) ; this is the trace node for object.
(add-feature trace to np-3)
(make-pointer np-3 to np-l)

(attach np-3 to vp-l as obj)

(sem connect np-l v-l obj 2) : see order of wl and w2 here.
(attach vp-l to 3-1)

(add-feature struc-ok to s—l)
(remove-feature modifiable from np-l)

This rule applies for the same situation for the above "np-vpp" rule.
But the NP is analyzed as the indirect object of the verb.
(ex) "the teacher taught by the professor ...“

‘0 V0 ‘-

(np-vppl ()
[ ( [ (np-l low-right) l (modifiable) ]
[ (v-1) 0 (en transitive dative) ] ; has feature of past participle

)
( ) l
[ (create s-l in clist)
(add-feature sec to s-l)

I
I

I
I

I
I

o
I

I
I
I

I
I

I
I

(attach s-l to np-l implicitly)
(create np-2)

(add-feature dummy to np-2)

(attach np-Z to s-1 as subj)

(create aux-l)

(add-feature passive to aux—1)
(attach aux-1 to s-l)

(create vp-l)

(attach v-l to vp-l)

(create np-3)

(add-feature trace to np-3)
(make-pointer np-3 to np-l)

(attach np-3 to vp-l as iobj)

(sem connect np-l v—l (dative-prep v-1) 2) ; see order of wl and w2 here.
(attach vp-l to s-l)

(add-feature struc-ok to s-l)
(remove-feature modifiable from np-l)

This rule analyzes the direct object of a clause.
(object-rule ()
[ ( [ (v-l vp—l s-1) 1 (transitive) ]
[I(np-ll 0 ((not wh)) l )
(l
[ (attach np-l to vp—l as obj)
(sem connect v-l np-l obj)
(add-feature struc-ok to s-l) ]

This rule analyzes the determiner(a, the) of an NP.
(parse-det ((l )(h noun-parse3))
[ ( [ (det-l) 0 () l
)
( ) l
[ (create ns-l on Clist)
(attach det—l to ns-l)
(transfer-features det-l to ns-l) ]

This rule analyzes the passive clause's real subject of the form "by NP".
(ex) “... eaten by the tiger"
(passive-by ((1 np-pp np-ppl)(h ))
[ ( [ (np-l s-1) 1 () l
[ (aux-1 s-1) 1 (passive) ]
[ (Vp-l s-1) 1 () l
[ (pp-1) o () l )
( [equal-exact (symbol by) (prep-of—pp pp-1)] ) ]
[ (sem connect (v vp-l) (np pp-l) subj 2)
(attach pp-l to vp-l)
;;(replace np-l with pp-l)
;;(remove-from-clist pp-l)

This rule creates the start of a that-complement clause where
"that" is missing.

(ex) “I know he passed the test."
(reduced-that-conp ((l reduced-relative)(h )) ;;; reduced-relative has lower pri.;;;;

[ ( [ (Vp-l s-1) 1 (thatcomp) ]
[ (hp-1) 0 ( (not wh) ) 1
)
( ) l
[ (push—to-cash-last—node) ; move the np to cash.

(create s-2)

(add—feature sec to 5-2)

(add-feature comp to s-2)

(create that-comp-l)

(add-feature dummy to that-comp-l)

(attach that—comp-l to s-2) ; to use as a claudse starter.

(attach s-2 to vp-l as comp implicitly)

(add-feature struc-ok to s-l)
the following action push the instance of (v vp-l) to
the global variable *connect_main_sec_pr0position*.
the instance for the verb of sec clause will be pushed
later by the main verb processing rules.

(c nnect_main_secl (v vp-l) )

HO‘O‘Q‘. \.

This rule analyzes the first NP of a clause to be the subject of the clause.

I
I

I

I
I

‘0 ‘0 ‘-

205

(subject-rule ()

[

[

( [ (*clause-start* s-1) 1 () ] ; ss, relpro, comp, .. ..... .
[ (hp-1) 0 ((not wh)) ] )

() ] ; no additional restrictions

(attach np-l to s-l as subj)

(add—feature subj to np—l) ]

This rule close an affirmative sentence by looking at the final

- punctuation mark ".".

I

(s-close-aff U

[

[

( [ (s-1) 1 ( (not question) main struc-ok ) ]
[ (fpunc-l) O ( (not question) ) ])

() ]

(attach fpunc-l to s-l)

(set-sentence—finish) ]

This rule attaches a PP to the VP of a clause.

[

(ex) "eat with a fork"

(vp-pp ((1 )(h wh-dativ?-np-pp))

( [ (vp-l s-1) 1 ()
I (PP-l) 0 () ] )
U ]
(attach pp-l to vp-l)
(if (have-feature trace in pp-l)

(sem connect (np pp-l) (v vp-l) (prep pp-l) 2) ; if wh object is under pp.

(sem connect (v vp-l) (np pp-l) (prep pp-l)) )

This rule starts the relative clause by noticing that
"which" or "that“ is directly following an NP.

(ex) "the book which the man read ..."

(wh-relative () ; after this rule, subject rule will run.

[

( [ (np-l low-right) 1 (modifiable) ]

()[](relpro-l) 0 () l )

(create s-l on Clist)

(add-feature sec to s-l)

(add-feature rel to s-l)

(attach relpro-1 to s-l) ; relpro-1 will work as clause-starter
(bind-whcomp np-l) ; this means making whcomp reg point to np-l
(remove-feature modifiable from np-l)

(attach s-l to np-l as mod implicitly) ] ; but s-l should still be on clist.

LIST OF REFERENCES

LIST OF REFERENCES

Akmajian, A. and Heny, F. W. (1975) An Introduction to the Principles of T ransforma-
tional Syntax. Cambridge, Massachusetts: The MIT Press, 1975.

Allen, J. F. (1983) "Recognizing Intentions from Natural Language Utterances". in
Brady, M. and Berwick, R. C. (editors). Computational Models of Discourse.
Cambridge, MA: The MIT Press, 107-166.

Allen, J. F. and Hayes, P. J. (1985) "A common sense theory of time." Proceedings of
9th International Joint Conference on Artiﬁcial Intelligence, Los Angeles, 1985,
528-531.

Alterman, R. (1985) "A Dictionary Based on Concept Coherence." Artiﬁcial Intelligence
25(2), 1985, 153-186.

Anderson, J. (1983) The Architecture of Cognition. Harvard University Press, Cam-
bridge, MA, 1983.

Barton, E. and Berwick, R. (1985) "Parsing with assertion Sets and Information Mono-
tonicity." Proceedings of the 9th International Joint Conference on Artiﬁcial Intel-
ligence, Los Angeles, California, 1985, 769-771.

Berwick, R. (1983) "A Deterministic Parser with Broad Coverage." Proceedings of the
8th International Joint Conference on Artiﬁcial Intelligence, Karlsruhe, West Ger-
many, 1983.

Birnbaum, L. and Selfridge, M. (1981) "Conceptual Analysis." in Schank and Riesbeck
1981, 318-353.

Bobrow, D. G. (1968) "Natural Language Input for a Computer Problem Solving Sys-
tem." in Minsky, M. L.(editor), Semantic Information Processing. Cambridge,
Mass.:The MIT Press, 146-226.

206

207

Bobrow, D. G. and Winograd, T. (1977) "An Overview of KRL: A knowledge represen-
tation language." Cognitive Science 1(1), 1977, 3-46.

Bobrow, R. J. and Webber, B. L. (1980) "Knowledge Representation for
Syntax/Semantic Processing." Proceedings of the First Annual National Confer-
ence on Artificial Intelligence. Stanford, August 1980, 316-323.

Brachman, R. J. (1979) "On the Epistemological Status of Semantic Networks." In Asso-
ciative Networks: Representation and Use of Knowledge by Computers. N. V.
Findler(editor), New York: Academic Press, 1979.

Brachman, R. J. (1983) "What IS-A is and isn’t: An Analysis of Taxonomic Links in
Semantic Nets." Computer 16(10) 1983, 30-36.

Brachman, R. J., Fikes, R. E. and Levesque, H. J. (1983) "Kripton: A Functional
Approach to Knowledge Representation." Computer 16(10), 1983, 67-73.

Brachman, R. J. and Schmolze, J. G. (1985) "An overview of KL-ONE knowledge
representation system." Cognitive science, 9(2), 1985, 171-216.

Bresnan, J. W. (editor) (1982) The Mental Representations of Grammatical Relations.
Cambridge, Massachusetts: The MIT Press, 1982.

Burton, R. (1976) "Semantic Grammar: An Engineering Technique for Constructing
Natural Language Understanding Systems." Technical Report 3453, Bolt Beranek
and Newman, Cambridge, Mass, 1976.

Carbonell, J. G. and Hayes, P. J. (1983) "Recovery Strategies for parsing Extragrammat-
ical Language." Computational Linguistics, 9(3-4), 1983, 123-146.

Charniak, E. (1976) "Inference and Knowledge." in Charniak and Wilks 1976.

Charniak, E. (1978) "On the use of framed knowledge in language comprehension."
Artiﬁcial Intelligence, " I 1(3), 1978, 225 -266.

Charniak, E. (1981a) "A Common Representation for Problem Solving and Language
Comprehension Information." Artiﬁcial Intelligence, 16(3), 1981, 225-255.

Charniak, E. (1981b) "The Case-Slot Identity Theory." Cognitive Science, 5(3), 1981,
285-292.

208

Charniak, E. (1981c) "Six Topics in Search of a Parser." Proceedings of the 7th Interna-
tional Joint Conference on Artiﬁcial Intelligence, Vancoubcr, Canada, 1981,
1079- 1087.

Charniak, E. (1982) "Context Recognition in Language Comprehension." in Lehnert
and Ringle 1982, 435-454.

Charniak, E. (1983a) "Passing markers: A Theory of Contextual Inﬂuence in Language
Comprehension." Cognitive Science, 7(3), 1983, 171-190.

Charniak, E. (1983b) "A Parser with Something for Everyone." in King 1983, 117-149.

Charniak, E. (1986) "A Neat Theory of Marker Passing." Proceedings of the 5th
National Conference on Artiﬁcial Intelligence, 1986, 584-588.

Charniak, E. and Mcderrnott, D. V. (1986) Introduction to Artificial Intelligence.
Addison Wesley, 1986.

Charniak, E., Riesbeck, C. K. and McDermott, D. V. (1980) Artiﬁcial intelligence pro-
gramming. Hillsdale, NJ, Lawrence Erlbaum Associates, 1980.

Charniak, E. and Wilks, Y. A. (1976) Computational Semantics: An Introduction to

Artiﬁcial Intelligence and Natural Language Comprehension. Amsterdam: North-
Holland, 1976.

Chomsky, A. N. (1965) Aspects of the Theory of Syntax. Cambridge, MA: The MIT
Press, 1965.

Church, K. (1980) "On memory limitations for natural language parsing." Technical
Report MIT/LCS/l‘R-245, Laboratory for Computer Science, MIT, 1980.

Church, K. and Patil, R. S. (1982) "Coping with syntactic ambiguity or how to put the
block in the box on the table." American Journal of Computational Linguistics,
8(3-4), 1982.

Colby, K. M., Faught B. and Parkinson R. (1974) "Pattern Matching Rules of the recog-
nition of Natural Language Dialogue Expressions." Stanford AI Lab. Memo AIM-
234, 1974.

Collins, A. M. and Loftus, E. F. (1975) "A spreading activation theory of semantic pro-
cessing." Psychological Review, 82(6), 1975, 407-428.

209

Crain, S. and Steedman M. (1985) "On not Being Led up the Garden Path: The Use of
Context by the Psychological Parser." in Studies in Natural Language Processing.
Dowty, D., Karttunen, L. J. and Zwicky, A. M.(editors), Cambridge University
Press, 1985.

Cullingford, Richard (1978). "Script Application: Computer Understanding of Newspa-
per Stories." Research Report No. 116, Department of Computer Science, Yale
University, New Haven, Connecticut, 1978.

Dahlgren, K. and Mcdowel, J.(1986) "Using common sense knowledge to disambiguate
prepositional phrase modiﬁers." Proceedings of the 5 th National Conference on
Artiﬁcial Intelligence, Philadelphia, 1986, 589-593.

Dowty, D. R, Wall, R. E. and Peters, S. (1981) Introduction to Montague Semantics.
Dordrecht: D. Reidel, 1981.

Dyer, M. (1983) In-Depth Understanding. Cambridge, MA: The MIT Press, 1983.

Eiselt, KR (1985) "A Parallel-process Model of On-Line Inference Processing."
Proceedings of the 9th International Joint Conference on Artiﬁcial Intelligence,
Los Angeles, California, 1985. 863-869.

Fahlman, S. E. (1979) NETL: A System for Representing and Using Real-World
Knowledge. MIT Press, Cambridge, 1979.

Feldman, J. A. and Ballard, D. H. (1982) "Connectionest models and their properties."
Cognitive Science, 6(3), 1982, 205—254.

Ferreira, F. and Clifton, C. (1986) "The independence of syntactic processing." Journal
of Memory and Language, 25, 1986, 348-368.

Fillmore, C. J. (1968) "The Case for Case." in Universals in Linguistic Theory. E. W.
Bach and R. T. Harms(editors), Holt, Reinhart and Winston, New York, 1968, 1-
88.

Finin, T. (1980) "Semantic interpretation of nominal compounds." proceeding of The
ﬁrst annual national conference on artiﬁcial intelligence, 1980, 310-312.

Fodor, J. A. (1983) Modularity of Mind, Cambridge, MA: MIT Press, 1983.

210

Fodor, J ., Bever, T., and Garrett, M. (1974) The Psychology of Language. McGrawhill
Book Co., New York, 1974.

Fodor, J. D. and Frazier, L. (1980) "Is the human sentence parsing mechanism an
ATN?" Cognition, 8, 1980, 417-459.

Ford, M., Bresnan, J. W. and Kaplan, R. M. (1982) "A Competence-based Theory of
Syntactic Closure." in Bresnan 1982, 727-796.

Forster, K. I. (1979) "Levels of processing and the structure of the language processor."
In Sentence processing: psycholinguistic studies presented to Merill Garrett. W. E.
Cooper and E. C. T. Walker (editors), Hillsdale, N.J.: Lawrence Erlbaum, 1979.

Frazier, L. (1978) On Comprehending Sentences: Syntactic Parsing Strategies. Ph. D.
Dissertation, University of Connecticut, 197 8.

Frazier, L. and Fodor, J. D. (1978) "The Sausage Machine: a New two Stage Parsing
Model." Cognition, 6(4), 1978, 291-325.

Frazier, L. and Rayner, K. (1982) "Making and correcting errors during sentence
comprehension: Eye movements in the analysis of structurally ambiguous sen-
tences." Cognitive psychology, 14, 1982, 178-210

Frazier, L. (1987) "Theories of sentence processing." In Modularity in Knowledge
Representation and Natural Language Understanding. Garﬁeld, J. L. (editor),
Cambridge, MIT Press, 1987, 291-307.

Gawron, J. M., King, J. J, Lamping J. and Egon E., Paulson, E. A., Pullum, G. K., Sag, I.
A. and Wasow, T. A. (1982)
"Processing English with a Generalized Phrase Structure Grammar." Proceedings
of the 20th Annual Meeting of the Association for Computational Linguistics,
Tronto, June 1982. 74-81.

Gazdar, G. (1981) "Unbounded dependencies and constituent structure." Linguistic
Inquiry, 12, 1981, 155-184.

Gazdar, G. (1982) "Phrase Structure Grammar." in The Nature of Syntactic Representa-
tion. Jacobson, P. and Pullum, G. K. (editors), Dordrecht: D. Reidel, 1982.

Gorell, P. G. (1987) Studies of human syntactic processing: Ranked-parallel versus
serial models. Ph. D. Dissertation, University of Connecticut, 1987.

211

Granger, R. H. (1983) "The NOMAD System: Expectation-based Detection and Correc-
tion of Errors during Understanding of Syntactically and Semantically Ill-formed
Text." Computational Linguistics, 9(3-4), 1983. 188-196.

Granger, R., Eiselt, K., and Holbrook, J. (1986) "Parsing with parallelismza spreading-
activation model of inference processing during text understanding." in Memory,
Experience and Reasoning. J. Kolodner and C. Riesbeck(editors), Hillsdale,
NJzErlbaum, 1986.

Grice, H. P. (1975) "Logic and conversation." in Syntax and Semantics, Vol. III : Speech
Acts. P. Cole and J. L. Morgan(editor), Academies Press, New York, 1975.

Grosz, B. (1983) "TEAM: A transportable natural language interface system." Proceed-

ings of the Conference on Applied Natural Language Processing. Santa Monica,
1983, 39-45.

Hendler, J. (1986) "Integrating Marker-Passing and Problem-Solving: A Spreading
Activation Approach to Improved Choice in Planning." TR-l624, Department of
Computer Science, University of Maryland, College Park, MD, 1986.

Hirst, G. (1984) Semantic Interpretation Against Ambiguity. Ph. D. Dissertation,
Department of Computer Science, Brown University, 1984.

Hirst, G. (1986) Semantic Interpretation and the Resolution of Ambiguity. Cambridge
University Press, Cambridge, England, 1986.

Joshi, A. K., Webber, B. L. and Sag, I. A. (editors) (1981) Elements of Discourse Under-
standing. Cambridge University Press, 1981.

Katz, J. J. and Fodor J. A. (1963) "The Structure of a Semantic Theory." Language,
39(2), 1963, 170-210.

Kimball, J. (1973) "Seven Principles of Surface Structure Parsing in Natural Language."
Cognition, 2, 1973, 15-47.

King, M. (1983) Parsing Natural Language. London: Academic Press, 1983.

Kurtzman, H. (1985) Studies in syntactic ambiguity resolution. Ph. D. Dissertation,
MIT, 1985, distributed by IULC.

212

Lebowitz, M. (1980) Generalization and Memory in an Integrated Understanding Sys-
tem. Ph. D. Dissertation, Department of Computer Science, Yale University. 1980.

Lehnert, W. G. and Ringle, M. H.(editors) (1982) Strategies for Natural Language Pro-
cessing. Hillsdale, NJ: Lawrence Erlbaum Associates, 1982.

Lesser, V. R., Fennel, R. D., Erman L. D. and Reddy, D. R.(l975) "Organization of the
Hearsay H Speech Understanding System." IEEE Transaction on Acoustics,
Speech, and Signal Processing. ASSP-23, 1975, 11-24.

Lorch, R. F. (1982) "Priming and Search Processes in Semantic Memory: A test of three
models of spreading activation." Journal of Verbal Learning and Verbal Behavior,
21, 1982, 468-492.

Lytinen, S. L. (1984) The Organization of Knowledge in a Multi-lingual Integrated
Parser. Ph. D. Dissertation, Department of Computer Science, Yale University,
1984.

Lytinen, S. L. (1986) "Dynamically Combining Syntax and Semantics in Natural
Language Processing." Proceedings of the 5th National Conference on Artificial
Intelligence, 1986, 574-578.

Lytinen, S. L. (1988) "Are Vague Words Ambiguous?" in Lexical Ambiguity Resolution,
S. Small, G. Cottrell and M. Tanenhaus(editors), Morgan Kaufmann Publishers,
California, 1988, 109-128.

Marcus, M. (1980) A Theory of Syntactic Recognition for Natural Language. Cam-
bridge, MA: The MIT Press, 1980.

Marcus, M. Hindle, D. and Fleck, M. (1983) "D-theory: Talking about talking about
trees", Proceedings of the let Annual Meeting of the Association for Computa-
tional Linguistics, Stanford, CA, 1983, 129-136.

Marr, D. (1977) Artiﬁcial Intelligence - A personal view", Artiﬁcial Intelligence, 9,
1977, 37-48.

Marslen-Wilson, W. D and Tyler, L. (1980) "The Temporal Structure of Spoken
Language Understanding." Cognition, 8(1), 1980, 1-71.

McCarthy, J. and Hayes, P. J. (1969) "Some philosophical problems from the stand point
of artiﬁcial intelligence." in Machine Intelligence 4, Meltzer and Michie(editors),

213

Edinburgh, 1969, 463-502.

McDermott, D. (1981) "Artiﬁcial Intelligence Meets Natural Stupidity." in Mind
Design, J. Haugeland(editor), MIT Press, Cambridge, MA, 1981, 143-161.

Milne, R. (1980) "a Framework for Deterministic Parsing using Syntax and Semantics."
DAI Working Paper, Department of Artiﬁcial Intelligence, University of Edin-
burgh, January 1980.

Milne, R. (1982) "Predicting Garden Path Sentences." Cognitive Science, 6, 1982, 349-
373.

Milne, R. (1983) Resolving Lexical Ambiguities in the Deterministic Parser. Ph. D.
Dissertation, Department of Artiﬁcial Intelligence, University of Edinburgh, Scot-
land, 1983.

Milne, R. (1986) "Resolving Lexical Ambiguity in a Deterministic Parser." Computa-
tional Linguistics, 12(1), 1986, 1-12.

Minsky, M. (1968) Semantic Information Processing. MIT Press, Cambridge, MA,
1968.

Minsky, M. (1975). "A Framework for Representing Knowledge." in The Psychology
of Computer Vision. Winston, Patrick (editor), New York, McGraw-Hill, 1975,
211-277.

Norvig, P. (1983a) "Frame Activated Inferences in a Story Understanding Program."
Proceedings of the 8th International Joint Conference on Artiﬁcial Intelligence,
Karlsruhe, West Germany, 1983.

Norvig, P. (1983b) "Six Problems for Story Understanders." Proceedings of 4th National
Conference on Artiﬁcial Intelligence, Washington, DC, 1983.

Norvig, P. (1986). Uniﬁed theory of inference for text understanding. Ph. D. thesis,
Department of Computer Science, University of California: Berkeley, 1986.

Pazzani, M. K. and Engelman, C. (1983) "Knowledge-based Qusetion Answering."
Proceedings of the Conference on Applied Natural Language Processing, Santa
Monica, 1983, 73-80.

214

Pazzani M. K. (1984) "Conceptual Analysis of Garden Path Sentences. " Proceeding of
International Conference on Computational Linguistics, 1984, 486-490.

Perreira, F. C. N. and Warren, D. H. D. (1980) "Deﬁnite clause grammars for language
analysis--- a survey of the formalism and a comparison with augmented transition
networks." Artiﬁcial Intelligence, 13(3), 1980, 231-278.

Quillian, M. R. (1968) "Semantic Memory." in Semantic Information Processing. Min-
sky, M. L.(editor), Cambridge, MA: The MIT Press, 1968.

Quillian, M. R. (1969) "The teachable language comprehender: A simulation program
and theory of language." Communications of the ACM, 1969, 459-476.

Ra, D. and Stockman, G. C. (1989a) "Use of Knowledge for Word Sense Disambigua-
tion." Proceedings of the Annual Conference for the International Association of
Knowledge Engineers, College Park, Maryland, 1989.

Ra, D. and Stockman, G. C. (1989b) "Integrated Natural Language Parsing Based on
Interleaved Semantic Processing." Proceedings of the Second International Sympo-
sium on Artificial Intelligence, Monterrey, Mexico, 1989.

Riesbeck, C. K. (1975) "Conceptual Analysis." in Schank 1975, 83-156.

Riesbeck, C. K. and Martin, C. E. "Direct Memory Access Parsing." Technical Report
354, Department of Computer Science, Yale University, 1984.

Riesbeck, C. K. and Schank R. C. (1976) "Comprehension by Computer: Expectation-
based analysis of Sentences in context." Research Report 78, Department of Com-
puter Science, Yale University, 1976.

Rieger, C. (1976) "An organization of knowledge for problem solving and language
comprehension." Artificial Intelligence, 7(2), 1976.

Ritchie, G. (1983) "Semantics in Parsing." in Parsing Natural language, M. King (edi-
tor), Academic Press, London, 1983.

Rumelhart, D. (1975) "Notes on a schema for stories." in Representation and Under-
standing. D. G. Bobrow and A. Collins(editor), Academic Press, New York, 1975,
211-236.

215

Sabah, G. and Rady, M. (1983) "A deterministic syntactic-semantic parser." Proceed-
ings of the 8th International Joint Conference on Artiﬁcial Intelligence, Karlsruhe,
August 1983.

Sachs, J. S. (1967) "Recognition Memory for Syntactic and Semantic Aspects of Con-
nected Discourse." Perception and Psychophysics, 2, 1967, 437-442.

Sampson, G. R. (1983) "Deterministic Parsing." in King 1983, 90-116.

Schank, R. C. (1975) Conceptual Information Processing. Amsterdam: North-Holland,
1975.

Schank, R. C. (1982a) "Reminding and Memory Organization: An Introduction to
MOPS." in Strategies for Natural Language Processing. W. G. Lehnert and M. H.
Ringle(editors), Lawrence Erlbaum Associates, Hillsdale, N.J., 1982, 455-494.

Schank, R. C. (1982b) Dynamic Memory: a Theory of Reminding and Learning in Peo-
ple and Computers. Cambridge University Press, 1982.

Schank, R. C. and Abelson, R. P. (1977) Scripts, Plans, Goals and Understanding: an
Inquin into Human Knowledge Structures. Hillsdale, New Jersey: Lawrence Erl-
baum Associates, 1977.

Schank, R. C. and Birnbaum, L. (1980) "Memory, Meaning and Syntax." Research
report No. 189, Department of Computer Science, Yale University, New haven,
Connecticut. 1980.

Schank, R. C. and Riesbeck CK. (1981) Inside Computer Understanding: Five Pro-
grams plus Miniatures. Lawrence Erlbaum Associates, Hillsdale, N .J ., 1981.

Seidenberg M., Tanenhaus, M., Leiman, J., and Bienkowski, M. (1982) "Automatic
access of the meanings of ambiguous words in context: Some limitation of
knowledge-based processing." Cognitive Psychology, 14(4), 1982, 489-537.

Selfridge, M. (1986) "Interated Processing Produces Robust Understanding." Computa-
tional Linguistics, 12(2), 1986, 89-106.

Small, S. L. (1980) Word Expert Parsing: a theory of distributed word-based natural
Language Understanding. Ph. D. Dissertation, Department of Computer Science,
University of Maryland, September 1980.

216

Small, S. L. (1983) "Parsing as Cooperative Distributed Inference: Understanding
through Memory Interactions." in King 1983.

Small, S. L., Cottrell, G. W. and Shastri, L. (1982) "Toward Connectionist parsing."
Proceedings of the National Conference on Artiﬁcial Intelligence, 1982, 247-250.

Small, S. L. and Rieger, C. J. (1982) "Parsing and Comprehending with Word Experts(a
theory and its realization)" in Lehnert and Ringle 1982, 89-147.

Sowa, J. F. (1984) Conceptual Structures. Addison-Wesley, 1984.

Stabler, E.(1983) "Deterministic and bottom-up parsing in prolog." Proceedings of the
National Conference on Artiﬁcial Intelligence, 1983.

Swinney, D. A. (1979) "Lexical Access during Sentence Comprehension:
(Re)Consideration of context effects." Journal of Verbal Learning and Verbal
Behavior, 18(6), 1979.

Tomita, M. (1987) "An Efﬁcient Augmented-context-free Parsing Algorithm." Compu-
tational Linguistics, 13(1-2), 1987, 31-46.

Tyler, L. and Marslen-Wilson, W. (1977) "The on-line Effects of Semantic Context on
Syntactic Processing." Journal of Verbal Learning and Verbal Behavior, 16, 1977,
683-692.

Waltz, David L (1982). "The State of the Art of the Natural-language understanding."
in Lehnert and Ringle 1982.

Waltz, D. L. and Pollack, J. B. (1984) "Massively parallel parsing: A strongly interactive
model of natural language interpretation." Technical Report, Coordinated Science
Laboratory, University of Illinois, Urbana, Ill, 1984.

Wanner, E. (1980) "The ATN and the sausage Machine: Which one is baloney?" Cogni-
tion, 8(2), 1980, 209-225.

Weischedel, R. M. and Ramshaw, L. A. (1986) "Reﬂections on the Knowledge Needed
to Process Ill-Formed Language." Report No. 6264, BBN Laboratories Inc. Cam-
bridge, Massachusetts, 1986.

. Weischedel, R. M. and Sondheimer N. K. (1983) "Meta-Rules as a Basis for Processing
Ill-formed Input." Computational Linguistics, 9(3-4), 1983, 161-177.

217

Wilensky, R. (1987) "Some Problems and Proposals for Knowledge Representation".
Report No. UCB/CSD 87/351, Computer Science Division, University of Califor-
nia, Berkeley, 1987.

Wilensky, R., Arens, Y. and Chin, D. N. (1984) "Talking to Unix in English: An Over-
view of UC". Communications of the ACM, 27(6), 1984.

Wilensky, R., Mayﬁeld, J., Albert, A., Chin, D., Cox, C., Luria, M., Martin, J. and Wu,
D. (1986) "UC - A Progress Report". Report no. UCB/CSD 87/303, Computer Sci-
ence Division(EECS), University of California, Berkeley, 1986.

Wilks, Y. A. (1975a) "An Intelligent Analyzer and Understander of English." Commun-
ications of ACM, 18(5), 1975, 264-274.

Wilks, Y. A. (1975b) "A Preferential, Pattem-Seeking Semantics for Natural Language
Inference." Artiﬁcial Intelligence, 6, 1975, 53-74.

Winograd, T. (1972) Understanding Natural Language. New York, Academic Press,
1972.

Winograd, T. (1983) Language as a Cognitive Process. Addison-Wesley, 1983.
Winston, P. H. (1984) Artiﬁcial Intelligence. Second Edition, Addison-Wesley, 1984.

Woods, W. A. (1970) "Transition Network Grammars for Natural Language Analysis."
Communications of the ACM, 13(10), 1970, 591-606.

Woods, W. A. (1972) "An experimental parsing system for transition network gram-
mars." in Natural Language Processing, R. Rustin (editor), Algorithmics Press,
New York, 1972.

Woods, W. A. (1975) "What’s in a Link: Foundations for Semantic Networks." in
Representation and Understanding: Studies in Cognitive Science. Bobrow, D. G.
and Collins, A. M. (editors), New York: Academic Press, 1975, 35-82.

Woods, W. A. (1983) "What’s Important about Knowledge Representation." Computer,
16(10), 1983, 22-29.

Woods, W. A., Kaplan, R. M. and Nash-Webber, B. L. (1972) "The Lunar Sciences
Natural Language Information system: Final Report." Report 2378, Bolt, Beranek
and Newman, Inc., Cambridge, Massachusetts, 1972.