u: .y. Aiu
w}: 5.1! ..

l‘f

SuvfiA ...
?,\L

. v...
‘1'?-

V; ..1

 

 

r r I-v‘v

 

9~ IHIlIHIllHIIIlllllllllllllllll“liflllllllllllllllll

3 1293 01405 7602

 

 

 

 

 

 

 

 

 

This is to certify that the

thesis entitled

DISCOURSE REPRESENTATION THEORY AND PRONOUNS

presented by

Daehee Lee

has been accepted towards fulfillment
of the requirements for

M.S. degree in Computer Science

 

 

(P.

T "
Major fessor

 

Date 6/" 9*fé

 

0-7639 MS U is an Affirmative Action/Equal Opportunity Institution

 

 

  
      

 
 

LIBRARY
M'chlsan State
Unlverslty

PLACE N RETURN BOX to romovo this chockout from your rocotd.
To AVOID FINES mum on or Moro duo duo.

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

MSU!eA-:.“" “ " " " ' Z._.Z‘...I-..

 

DISCOURSE REPRESENTATION THEORY AND PRONOUNS

by

Daehee Lee

A THESIS

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

MASTER OF SCIENCE

Department of Computer Science

1996

 

ABSTRACT
DISCOURSE REPRESENTATION THEORY AND PRONOUNS

BY

Daehee Lee

Pronominal Binding is one of the most complicated
process of natural language understanding, since pronouns
are differently interpreted, depending on which antecedent
they refer to, and how they refer to their antecedents.
This thesis resolves pronominal binding, tackling two
problems: how a discourse must be represented to be
accessible by computational procedures, and how
computational procedures must be controlled in accessing the
representations for antecedent search.

By designing a model in which syntax gives DRT some
information about pronominal disjointness, and DRT searches
for pronominal antecedents, this thesis can resolve many
cases of referential, bound, and E-type pronominal binding,
and pronominal binding in subordinations. This thesis also

implements the theory in Prolog.

Dedicated to my mother, Sochool, my wife, Gemhee, and my
three sons, Jaeyoung, Jaegook, and Jaesung

iii

ACKNOWLEDGMENTS

I thank all the members of my committee, Carl Page,
Barbara Abbott, George Stockman, and Betty Cheng, for their
assistance for this study. Specially, I express my
gratitude to Carl Page and Barbara Abbott.

Carl Page encouraged and helped me to be admitted to
the graduate study of computer science, when I express my
first interest in this field. His support and guidance were
very helpful in my graduate study. He gave me good pieces
of comments and information concerning this research.

I am grateful to Barbara Abbott for her encouragments
and good comments on the linguistic part of this study, and
for her guidance in my academic training in linguistics.

I express my deep and special gratitude to Alan Munn.
Although he was not a member of my committee, he very
happily gave his good comments on this project.

I am also indebted to OPHS (Office of Programs for
Handicapper Students) and Tower Guard for their strong
support. Without their reading service I could not have
finished my graduate study at all. I give my special

gratitude Michael Hudson, a vision specialist at OPHS, for

iv

his support and invaluable information on accessibility on
‘ campus.

I also thank my friends at computer science for their
help and friendship. I give my special thank to Hyunduk
John for many delightful conversations and good discussions
throughout the courses of computer science.

I am also very truly grateful to my family. My mother,
and brother, and sisters’ strong support in pray and pride
in me were good encouragments for me.

I never forget my wife’s sacrificing support for me. I
also thank all my three sons that they gave up their mother

most of time when she helped me.

TABLE OF CONTENTS

INTRODUCTION

1. A Computational Model of Discourse . . . . . .
1.1 .A General Model of Language Production
Comprehension . . .
1. 2 Why Discourse Representations Are Necessary
1. 3 Discourse Representation Theory

2. Pronominal References and DRT
2.1 Deictic vs. Anaphoric Pronouns
2.2 Interactions between Syntax and DRT
2.2.1 A General Constraint in Syntax
2.2.2 More Constraints in Syntax
2.2.2.1 Bound Pronouns
2. 2. 2. 2 E- -type Pronouns
2. 3 Pronouns and Antecedent Search in DRT

3. Computing Discourse Representation Structures
3.1 Phrase Structure Rules and Building DRS’s

.1 Phrase Structure Rules

2 Feature Structures and Unification

3 Representing DRS’ s

4 Parsing Discourses

S Parsing Noun Phrases

6 Parsing Verb Phrases .

dding DRS’ s in Knowledge Base
.1 Simple Statements . .

.2

.3

o

.1

.2

.3

w

gwwwww

w

M

L11
wwLAJ'UNNle-‘Hl-‘Hl-‘H

If—then Sentences

Questions

h ric Pronouns . .

Antecedent Search
Subordinations and Anaphora

The Accommodations and Problems

a

DJ
w
wtpuigtuuiw

4. Conclusion and Further Research
APPENDIX

REFERENCES

vi

83

87

102

INTRODUCTION

Typically, people communicate with computers by means of
artificial languages. These languages are more suitable for
computers rather than for people. In many cases, however, it
will be more convenient if people can communicate with
computers in natural languages. For example, to query some
information from a large database, people should know to
formulate the queries to retrieve the information which they
need. If they can communicate with computers in natural
languages, they can easily access the database and retrieve
the information for their purposes.

For computers to be able to understand the natural
languages, they must have the knowledge of the languages such
as phrase structure rules, argument structures, some
constraints on sentential derivations, some rules for semantic
interpretations, and so on.

Relating to the knowledge about language use, the way in
which pronouns are used and interpreted fix: a sentence or
across sentences seems to be one of the most complicated
processes in natural language understanding. Pronouns are

semantically interpreted in some different manners, depending

2
upon what they refer to, and how they are anaphorically linked
with their antecedents. It is what we may call the problem of
pronominal binding.

The goal of this thesis is to resolve intersentential
pronominal binding in discourse, tackling two main problems:
a representational problem and a computational problem. The
former is related to the question of how discourse knowledge
can be formally represented in a computational model, and the
latter is related to the question. of how to control the
computational system to acces the representation of discourse
for pronominal antecedent search: under which conditions
anaphoric relations between different NP’s and pronouns are
possible. If we can understand and formalize the use of
anaphoric pronouns it will contribute to enhancing natural
language understanding systems.

To achieve the goal, this thesis will, first of all,
classify the pronominal uses, and then investigate the
representations and interpretations of anaphoric pronouns in
a computational model of discourse--Discourse Representation
Theory. The thesis will be organized as follows:

Chapter 1 will discuss Kamp’s (1981) discourse
representation theory (DRT) as a computational model in which
the pronouns are represented and interpreted. Chapter 2 will
classify and characterize pronouns on their use. Further, the
representations and some constraints on pronominal

interpretations will be in detail discussed in syntax and DRT.

3
Chapter 3 will discuss some issues of computing discourse
representation structures in logic programming language.
Chapter 4 will summarize the topic and discuss some issues for

further research.

1. A Computational Model of Discourse

1.1 A General Model of Language Production and

Comprehension

Language use (production.and.comprehension) is one of the
human behaviors for communications. In discourse participants
produce some linguistic forms to deliver their messages or
mental states such as knowledge, thoughts, feelings, ETC. to
share with others. For speakers to deliver messages, first of
all, they build some internal forms or conceptual structures
for the messages via some inference rules. Then they will
produce some linguistic forms by means of some principles and
rules. When they produce the linguistic forms for their
intended meanings, they assume that their hearers can
understand.the inference rules and linguistic rules which.they
use for the production. The hearers now parse the linguistic
forms, and translate them into some internal forms with the
same linguistic rules that the speakers use for linguistic
production. Then the hearers further apply the inference
rules to the internal representations and understand the
speakers’ intended meanings.

The internal or conceptual forms will become a set of the
common knowledge for the participants to share in discourse.
As discourse continues to go further, the participants also

update the common knowledge--adding some new information,

5

resolving some references of anaphora, etc. To understand

language use a very general computational model can be

diagrammed in (1).

(1) (cf. Berwick (1983) Figure 2 p.29)

 

linguistic knowledge
discourse knowledge

 

 

 

 

 

 

Speaker———————-— nearer
speaker's intended speaker’s intended
message(internal @————————>@ message(internal
form) utterance( form)

linguistic/ <
external literal meaning
form)

 

 

 

 

 

 

From a computational point of view it is important to
understand how participants construct their discourse
knowledge. As one of the efforts to understand it, this
thesis aims to enhance the representations of discourse
knowledge, and the way to keep track of discourse information
and compute the pronominal reference in discourse, given the

representations.

1.2 Why Discourse Representations Are Necessary

In this section let us consider why the computational
model requires some internal representations, and what factors
are involved in the computation of discourse. Let us take
some examples for concreteness. Consider that a discourse
starts with the utterances in (2).

(2) 81: Do you know that John bought a car?

6
s2: Yes. It looks nice.

In (2) a speaker sl introduces two entities, John and e

 

car, and specifies their relationship in which John is an
owner of a car. With sl’s utterance the two entities are
available common to s1 and s2 in the discourse. At next time
32 uses a pronoun lg, to talk about the car John bought. Thus
for the discourse to be felicitous, the participants in the
discourse should clearly understand what entities are
available in the discourse, how they can be added to the
discourse, and on which conditions it is possible for pronouns
to be used to refer to previously mentioned entities. If
participants do not know these factors, the discourse will
fail to be felicitous.

According to Kamp (1981) and Heim (1983) a discourse
entity is introduced into discourse by only some types of noun

phrases such as proper nouns like John, deictic expressions

 

like hhat this etc., indefinite NP’s like someone or a car.

 

 

Other types of noun phrases such as definite NP’s like hhe
eeg, anaphoric pronouns like he, ih, heg, etc., and so on
refer to the discourse referents already introduced in
discourse. It reflects two computational procedures: the one
introduces new entities; the other links one form to an
already familiar entity.

Another importance is to capture an appropriate meaning
of a sentence in discourse. The utterance of (3) potentially

has two meanings, as described in (4).

'7
(3) Two students ate a cake.
(4) a. There was one cake, and two students ate it
together.
b. There were two students who each ate a cake.

If a hearer hears (3), he /shewe should be able to
properly choose one of the two, (4.a) or (4.b), in the
context. To do so the hearer should internally disambiguate
the form of (3) as (4) at some representational level, and
keep one of (4.a) or (4.b) in the knowledge base for the
computational system to be able to correctly keep track of
discourse interactions.

Let us consider another utterance of (5).

(5) A man entered the room. He looked like a detective.

Following Sidner (1983) and. Grosz (1986), the
computational system. will be interested. in ‘updating' the
internal representations regarding what a discourse talks
about, and.when topic shift takes place. In (5) the discourse
talks about the man who entered the room, and the pronoun he
presumably refers to the man who just entered the room. If
the computation catches what the discourse talks about, it
will help the computation find what the pronoun refers to.

To represent discourse for intersential pronominal
binding, this thesis will adopt and enhance Kamp's Discourse
Representation Theoryu In next section let us take this model

into consideration for detail.

8

1.3 Discourse Representation Theory

Kamp (1981) proposes Discourse Representation Theory
(DRT) in order to interface between a syntactic level and
semantic interpretation. DRT assumes that there is a
linguistic level between syntax and semantics, which is a
level of Discourse Representation.

Let us consider the properties of DRT which make it
possible for the computational systemrto search.for pronominal
antecedents across sentences in discourse. They will be
described as follows: First, DRT is discourse—based. DRT
constructs the representations of a discourse--what we may
call discourse representation structures (DRS's). A single
DRS accumulates all discourse information. Second, DRT can
explicitly specify the intersentential scope and anaphora-
antecedent relations across sentences such as modal
subordinations, quantificational subordinations, and
pronominal references, providing a unified representation of
a. discourse structure ‘which. contains all intersentential
relations in a discourse. Third, DRT is independent of a
theory of syntax, since the level of DRT is placed after
syntactic analyses. So any syntactic analysis is suitable for
determining meaning. And fourth, DRT contains a theory of
truth-conditions. Truth is defined in terms of embedding
DRS's in a model.

The first and second property make it possible for DRT to

9
be adequate for solving the intersentential pronominal
binding. The third property allows syntax to help DRT to
search for pronominal antecedents in a DRS, providing some
constraints on anaphora-antecedent relations.

At the level of DRT a DRS consists of a pair of two sets,
<U, C> where U is a set of discourse referents, and C a set of
properties and relations, including negation, disjunction,
implication, etc. The former is called a universe and the
latter a condition (or a predication) on a universe.

A.DRS initially consists of a pair of empty sets, <U,C>,
since there is no information available in the discourse. To
start some conversation, the participants should use some
linguistic expressions which can introduce some discourse
referents and their properties or relations into a DRS. As
discourse goes on, however, a sequence of DRS's will be
constructed or revised by some algorithms.

According to Kamp (1981) and Heim (1983) a discourse
referent is introduced into a DRS by some types of NP's only.
That is, proper nouns, deictic expressions, indefinite NP’s
introduce new referents into a DRS, and other types of NP’s
such as definite NP’s, anaphoric pronouns, and so on refer to
the discourse referents already introduced in a DRS.

Let us take some examples for DRT now. Consider (6).

(6) John saw Mary.

In (6) a discourse referent, u, is introduced into a DRS

for John, and another discourse referent, v, is added into a

 

10

universe, U, for Mary. The relation between John and Magy,

 

saw(u,v), is added into a condition, C, which explicitly
constrains the relationship between two discourse referents,
u and v. Thus (6) is linearly represented in (7a) and
schematically represented in (7b).
(7) a. D1 = <{u,v}, {John(u), Mary(v), saw(u,v)}>.
b.

 

u v
John(u)
Mary(v)
saw(u, v)

 

 

 

After (7) is constructed the conversation is assumed to
go further with (8). Then we may have (9) as a DRS.

(8) John owns a donkey, and Mary owns a cat.

 

 

 

 

(9)
u v w x
John(u)
Mary(v)
saw(u, v)
donkey(w)
cat(x)
own(u, w)
own(v, x)
A. DRS is a discourse-based structure. 80 several

subsequent sentences can be accumlatively accommodated in one
DRS. Furthermore, even a sentence can have more than one DRS
if necessary, as described just below.

Kamp (1981) suggests that a DRS may be embedded in
another DRS. Let us take (10) for an example.

(10) a. John does not own a car.

b. not [a x: car x] (John own x). (preferred

ll

 

reading)
c. [a x: car x]not(John own x). (nonpreferred
reading)
(11)
u v
John(u)
car(v)

 

 

a own(u, v)

 

If we represent (10.a) as (11) in the same way as we have
done for (8), the DRS, (11), corresponds to (10.c), whose
reading is not preferred for some reason.

That is, the truth of (10.c) is determined by two
factors: the one is whether a unique car exists in a model,
and the other is whether there is a relation of John owning
that car. So (11) may be false even if there is no specific
car in a model, regardless of whether John owns a car or not.

If a speaker of (10.a) intends to mean (10.b), however,
the representation of (11) is problematic.

To solve the problem with the representation for (10.b),
we may represent it as (12). In comparison.with (11) the DRS,
(12), corresponds to (10.b).

(12)

 

John(u)

 

a car(v)
own(u, v)

 

 

 

 

 

 

In (12) Kamp (1981) calls the outer’ box: or’ DRS a

12
superordinate DRS to the inner box or DRS embedded in it,
(henceforth sup-DRS), and the embedded DRS a DRS subordinate
to the outer DRS (henceforth sub-DRS). A sub-DRS is also used
for if—then constructions, universal quantifiers, pragmatic
accommodations, modal/quantificational subordinationsl, etc.
One important thing is that for anaphor-antecedent
relations the discourse referents in sup—DRS are accessible to
the sub—DRS, but not vice versa.
Let us take an example for how a sub-DRS works for if-
then constructions and universal quantifiers. (13) is an if-
then construction, and can be represented in (14).

(13) If a farmeri owns a donkeyj, then hei feeds itj.

 

 

 

(14)
u v
farmer(u) he = u
donkey(v) -> it = v
own(u, v) feed(u, v)

 

 

 

 

 

 

 

 

 

In (14) the leftside box is a DRS superordinate to the
rightside box, and conversely, the rightside box is a DRS
subordinate to the leftside box.

Following Kamp (1981), a universal quantifier like eyegy
should be represented by an if—then construction in DRT.
Consider (15).

(15) a. Every farmer owns a donkey.

 

11 will discuss pragmatic accommodations with respect to E—
type pronouns in detail in chapter 2.

‘l

13
b. [every x: farmer x] [a y: donkey y] (x owns y).
(preferred reading)
c. [a y: donkey y][every x: farmer x](x owns y).
(nonpreferred reading)

Kamp (1981) and others have discussed a DRS, (16), for
(lS.a), but it corresponds to (15.b) only: Since the sentence
(15.a) is ambiguous as (15.b) and (15.c), although (15.c) is
not preferred for some reason here, DRT shouLd be able to
represent (15.c) as a DRS at the level of DRT.

So (15.c) can be represented as (17) in comparison with

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

(16).
(16)
u v
farmer(u) -> donkey(v)
owns(u, v)
(17)
v
donkey(v)
u own(u, v)
farmer(u) ->

 

 

 

 

 

 

 

 

 

The difference in the interpretation between (11) and
(12), and between (16) and (17) will be much more clear after
we review the interpretation of a DRS below.

If a DRS is constructed by some rules, then it is used to
determine the truth of a discourse. According to Kamp (1981,

P. 278):

14

"A sentence S, or discourse D, with representation m is
true in a model M if and only if M is compatible with m;
and compatibility of M with m, .., can be defined as the
existence of a proper embedding of m into M, where a
proper embedding is a map from the universe of m into

that of M which, .., preserves all the properties and
relations which m specifies of the elements of its
domain."

For this let us consider the truth conditions in the
model-theoretic semantics first. In order for a sentence S to
be true in Model M, there is a mapping which maps the
individuals in S into a subset of individuals in M and which
maps the properties or relations of the individuals in S into
a subset of properties or relations in M. In the same way a
DRS D is true in a model M iff there is such a mapping that
the set of discourse referents in the DRS is mapped into a
subset of the individuals in a model M, and that the
properties or relations in the DRS are mapped into the subset
of the properties or relations in M.

Then let us evaluate the truth of the DRS, (7). The DRS,
(7), with a set of discourse referents, U5={u,v}, and a set of
properties and relations, C5={John(u), Mary(v), saw(u,v)}, is
true in a model M with a set of individuals, U”, and an
assignment function, GM, iff <GM(John), GM(Mary)> is a member
of GM(Saw) and there is a mapping g from UD to UN such that
g(u)=G,, (John) and g(v)=GM(Mary) and <g(u), g(v)> is a member
of GM(Saw) .

Consider the case of (9). Let us partially evaluate John

 

owns a donkey in (9) for simplicity because other parts can be

evaluated in the same way: The truth value is determined.with

15

such an interpretation that the DRS, (9), which partially
consists of <UD,(hp.= <{u,w}, {John(u), donkey(w), own(u,w)}>,
is true in a model M with a set of individuals, UM, and an
assignment function, GM, iff there is any individual x in Up
such that GM(x) belongs to GM(donkey) , and <GM(John) , GM(x)> is
a member of GM(Own) and there is a mapping g from UD to UM such
that g(u)=GM(John) and g(w)=GM(x) and <g(u) , g(w)> is a member
of Gh(Own).

The DRS, (12), can be interpreted in a recursive way.
The DRS, (12), consists of D = <U,), CD> = <{u}, {John(u),
not(D1) }> where D1 = <Um,Cm> = <{v}, {car(v) ,own(u,v) }>. It is
true in a model M with a set of individuals, UM, and an
assignment function” 6%, such that there is a mapping g from
UD to UM such that g(u)=G,,(John) and D1 is false (that is,
not(Dg) is true). Now D1 is interpreted in the same way as
(9) .

The difference between (11) and (12) is that car(v)
should be true (or there should be a specific car) for (11) to
be true, and that car(v) does not need to be true for (12) to
be true. So this simulates the nonspecific use of an
indefinite NP.

Kamp (1981) suggests that if—then.sentences anduuniversal
quantifiers should be treated in the same way for
interpretation. For example, the statement, if A then B, is
true iff B should be true in all situations in which A is

true.

16
In next chapter let us look at pronominal representation

and their interpretation in DRT.

2. Pronominal References and DRT

First of all, let us start the discussion about the
pronominal uses, taking a brief look at Evans’s pronominal
classification because he not only introduced a new pronominal
use, but also reflected the prior pronominal analyses well.

Evans (1980) proposes that the pronouns be classified
into four types, depending upon.their occurrences, as follows:

(1) (i) "Pronouns used to make a reference to an object

(or objects) present in the shared perceptual
environment, or rendered salient in some other
way." (p.337)

(ii) "Pronouns intended to be understood as being
coreferential with a referring expression
occurring elsewhere in the sentence." (p.337)

(iii) "Pronouns which have quantifier expressions as

antecedents, and are used in such a way as to
be strictly analogous to the bound variables
of the logician." (p.337)

(iv) Pronouns which also "have quantifier
expressions as antecedents," but "are not bound
by those quantifiers." (p.338)

Let us take a close look at the pronominal properties and
references, based on the classification of (1) one by one

below.

17

18

2.1 Deictic vs. Anaphoric Pronouns

In the case of (l.i), first of all, a pronoun is used as
a referring expression as in (2). The sentence (2.a) may be
uttered with some gestural or perceptual demonstration to a
student coming into the room. Even without any gestural or
perceptual demonstrations to an object or objects, a pronoun
may be used as a referential expression in the context in
which an object or objects are salient, and a pronoun refers
to them. For example, suppose a boy runs back and forth and
makes noise in a public room for a while and then leaves the
room. Right after that (2.b) is uttered.

(2) a. He is a smart student.

b. He is a spoiled boy.

Such pronouns can be called deictic pronouns. The
deictic pronouns are semantically interpreted as a referring
expression as the pronouns classified as (1.ii) are, but they
are different from (1.11) in the uses and representations.

Let us observe the use of the pronouns of (1.ii) first,
and then discuss (1.1) and (1.ii) comparatively.

Now the (1.ii)-type pronouns can be exemplified in (3).

(3) a. When Johni.had finished, hei left.

b. Johnll was strong, and hei broke in the door.
c. John1 loves hisi mother.
d. Johni thinks that hei is sick.

In case one expression determines the semantic value of the

19
other expression, the former should be a constituent of the

latter. In (3), however, the semantic value of John

 

determines the semantic value of his or he, even though John

 

is not a constituent of his or he. In this case it is said

that the pronoun his or he is used as an anaphor on John, and

 

that gehh is an antecedent of hie or he. The pronoun hie or
he is interpreted as referential as in (2). That is, the
anaphoric pronoun hie or he refers to the same object that the
antecedent gehh refers to. Neale (1990) calls the (1.ii)-
type of a pronoun an anaphor as a referential usage.

At the level of DRT we can distinguish deictic pronouns
from anaphoric pronouns in some different way. Let us
consider a model of discourse first. A model of discourse
consists of discourse referents and their properties or
relations. While a conversation goes on, the discourse
referents and their properties in the model continue to be
updated in such a way that new objects or properties are
introduced into the model, or the discourse referents are
reordered with respect to the times when they are referred to.
Furthermore, the manner in which objects and/or properties are
added to the model is to express them linguistically only.
Although some objects are rendered salient by nonlinguistic
factors such as perceptual or gestural ones, they cannot be
discourse referents in the model of discourse if they are not
expressed linguistically. I will assume that this is a

constraint on the introduction of discourse referents into a

20
discourse model.

The terms "object" and "discourse referent" need to be
made clear. The term "object" refers to the "entities that
belong to the real world" (Kamp 1981 p.283), and the term
"discourse referent" refers to the entities that belong to the
representation of a discourse, which denote the objects in the
real world or a more abstract device for intermediate
maintenance of anaphoric relations.

Then let us assume that the anaphoric relation indicates
the configurational relation between an anaphoric expression
and a discourse referent. That is, only a discourse referent
can be a potential antecedent which an anaphor refers to, and
only the discourse referents at sup-DRS’s or in the same DRS
are accessible to an anaphor.

The pronouns of (1.i) refer to objects and the pronouns
of (1.ii) refer to discourse referents. More generally
speaking, (1.1) is distinguished from (1.ii, iii and iv) in
such a way that (1.1) applies to the pronouns which refer to
objects out of a discourse model, and introduce some discourse
referents into a discourse model, while (1.ii, iii and iv)
refer to acdiscourse referent or referents already existing in
the discourse model. The former may be called deictic
pronouns, and the latter anaphoric pronouns.

Then we can define deictic pronouns in DRT as follows:

(4) Pronouns are deictic if they' are used. to) make

reference to an object or objects which are out of a

21
discourse model or rendered salient in just a
nonlinguistic way.

We can also define anaphoric pronouns as follows:

(5) Pronouns are anaphoric if they are linked. with

discourse referents in a DRS.

With respect to (4) and (5), the pronouns in (2) are
deictic, and the ones in (3) are anaphoric.

Kamp (1981 P.282-283) assumes without any specific
explanation that deictic and anaphoric pronouns be treated
under the same principle:

"...the mechanisms which govern deictic and anaphoric

occurrences of pronouns are basically the same. ..

both deictic and anaphoric pronouns select their
referents from certain sets of antecedently available
entities."

It would be wrong if the deictic and anaphoric uses of
pronouns are treated in the same way just for the reason that
both "select their referents fromicertain sets of antecedently
available entities." (p.283)

Apparently, the mechanisms by which deictic and anaphoric
pronouns find their referents are different:

First, the properties of sets fromwwhich.the two pronouns
select are different. As I have mentioned just above, the
entities which the deictic pronouns refer to are in the real
world; that is, they directly refer to the entities in the
real world. In contrast, the anaphoric pronouns refer to the

entities in the representation of a discourse, and indirectly

refer to the entities in the real world via the discourse

22
referents.

Second, deictic pronouns refer to their objects in a
nonlinguistic way -- perceptual or gestural demonstration —-
but anaphoric ones determine their antecedents in very complex
ways--grammatical agreements, syntactic/DRS configuration (or
maybe thematic roles), a selected topic, and so on.

Third, deictic pronouns introduce new discourse referents
to a discourse representation, but anaphoric ones can never
introduce new discourse referents, and instead reorder the
referents in the discourse.

So, instead of the unified treatment of deictic and
anaphoric pronouns, the distinction between deictic and
anaphoric pronouns is necessary to establish the strategies to
determine their referents properly.

This classification of pronouns in DRT will solve some
problem.‘with. Evans’s classification” ‘With (1.ii) Evans
suggests that coreferentiality takes place in the same
sentence. But intuitively, (6) is semantically the same as
(3.b).

(6) Johni was strong. Hei broke in the door.

Following Evans (1980), the pronoun he in (6) is no more
a (1.ii)-type pronoun, because the latter is not in the same
sentence as the former and violates (1.11.). Evans may answer
this question in such a way that the former sentence makes
gehh salient in a context and he in the latter refers to the

salient object. So the pronoun in (6) can be treated as a

23
(1.i)-type. But it is not adequate if (3.b) and (6) are
differently treated, although they seem to be the same.
In addition to this inconsistency, we may say that the
pronoun in (3.b) can also be classified as (1.1) because the

prior clause makes John salient. Hence the case of (3.b) may

 

be classified as either (1.i) or (1.ii).

However, our analysis treats (3.b) and (6) as anaphoric
pronouns, since the pronouns refer to discourse referents in
a DRS.

Now let us consider the DRS’s of (2) and (3). Kamp
(1981) suggests that a proper noun should be placed at the
top—level of a DRS. Here the top level of a DRS indicates the
highest superordinate level. He did not discuss this rule,
but the reason that the a discourse referent for proper nouns
are introduced at the top—level of a DRS seems to make the
referent globally accessible to the anaphora. Then a deictic
pronoun can be treated in the same way as a proper noun. For
deictic pronouns, (2.a) may be represented as (7), and for

anaphoric pronouns, (3.b) will be represented as (8).

24

 

(7)
u

he(u)

smart (u)
student (u)

 

 

 

 

v
John(u)

strong (u)
door (v)

he=u
break_in(u, v)

 

 

 

2.2 Interactions between Syntax and DRT

2.2.1 A General Constraint in Syntax

Before discussing how DRT searches for pronominal
antecedents, we should take the role of syntax into
considerations.

Syntax provides pronominal contra-indices between
pronouns and some noun phrases for DRT. It means that some
pronouns should not be coindexed with some noun phrases with
the same indices that the pronouns have in syntax. If syntax
specifies some contra-indices between pronouns and noun
phrases, then DRT will exclude the NP’s for' pronominal
antecedent search.

For example, assume that syntax should specify a contra-
index between.he and.gehh in (9.a). Then DRT will allow'a DRS
like (9.b) in which he and gehh are anaphorically linked,

while it excludes (9.c) which indicates that h_e refers to

25
John.

(9) a. He.i likes Johni.

 

 

 

 

b.
u
John1 an
he” (V)
likes (v, u)
c.
u
Johni hfl
11.8.1 = u

 

 

likes (u, u)

 

How can syntax provide such contra-indices between
pronouns and NP’s, then? We assume that syntax should parse
a sentence with some constraints, and that the Principle C of
the binding theory in (10) be one of such constraints.

(10) Principle C of the Binding Theory

A R-expression2 with an index should not be c-
commanded by another expression with the same index
in syntax.

(11) X c-commands y iff all the categories which

dominate x also dominate y3.

For example, in (12) the pronoun he with an index i c-

commands John with the same index i so that it violates the

 

Principle C. Thus Syntax should provide a contra-index between
he and Jghn.

(12) a. Heq_thinks that John.i is rich.

 

2Roughly speaking, R—expressions are the ones excluding
anaphoric categories: pronouns, reflexives, and reciprocals.

3The c-command relationship is related to pure syntactic
configurations within a sentence.

26

he V S’
I/\
thinks C S
l/\
that NP VP
ll
John [is rich]
On the other hand, in (13) he does not c-command.gehh at
all, and syntax does not provide any contra-index between he
and gehh.

(13) a. John.i thinks that hei is rich.

27

/ \

NP VP

John V S’

|/\

thinks C S

I / \
that NP VP
I l
he [is rich]

This syntactic constraint and interaction with DRT will
treat pronouns in a consistent way.

For example, a question about Evans's treatment of (3)
has been raised by Bosch (1983). Following Bosch (1983), the
unified treatment of all the pronouns in (3) is irrelevant.
He proposes that the pronouns in (3.a.-b.) and the ones in
(3 .c. -d.) should be treated differently, since the pronouns in
(3.a.-b.) are pragmatically linked to their antecedents, and
the pronouns in (3.c.-d.) are syntactically linked to their
antecedents. Furthermore, Bosch argues that the syntactically
functioning pronouns are semantically interpreted as
nonreferential expressions,- pragmatically functioning ones are
interpreted as referential expressions like proper names.

To support his argument, he suggests the following as an

evidence under the assumption that if an expression is

28
referential then it is replaceable by another referring
expression:
(14) a. Fredi‘thinks hei is sick. (p.41)

b. John.i looks pale, and Fred thinks hei is sick.

(15) a. ?Fred1L thinks the old malinger; is sick.(p.42)
b. John.i looks pale and Fred thinks the old
malingerq is sick.

Replacing the pronouns in (14) with a referring
expression shows us a difference between (14.a) and (14.b).
Following Bosch, the reason that (15.hn is grammatical is that
the pronoun in (14.b) is referential and so replaceable with
another referring expression; the pronoun in (14.a) is not
referential and cannot be replaceable with another referring
expression.

Bosch argues that in (14.a) the pronoun he is
syntactically linked to its antecedent, depending on the
syntactic agreements, regardless of whether or not the
antecedent is a referring expression. The pronoun h_e in
.(14.b) as a referring expression, however, requires another
referring expression as its antecedent. They refer to two
independently referring expressions which have the same
referent, which may be called coreferential. However, the
pronoun in (14.a) does not require a referring expression as
its antecedent because it functions as a nonreferential

expression. So Bosch proposes that, since syntactically

29
functioning pronouns and pragmatically functioning pronouns
have their own mechanism for linking to their antecedents, the
pronouns in (3.a.-b.) and the ones in (3.c.—d.) be treated
under different principles.

However, the ungrammaticality of (15.a) does not depend
upon whether or not the pronoun in (15.a) is referential, as
Bosch analyzes. Rather, (15.a) violates the condition (c) of
the binding theory specified in (10). To put it in other
words, the phrase the old malingerer in (15.a) is coindexed
with.§;ee which c-commands it. On the other hand, the phrase
the old malinger in (15.b) is coindexed with gehh but not c-
commanded by it. 'Thus the ungrammaticality of (15.a) seems to
be attributed not to the characteristics of pronouns but to

the characteristics of the r—expression.

2.2.2 More Constraints in Syntax

To consider more constraints on pronominal antecedency,
we should classify pronouns into three types in syntax:
referential, bound, and E-type pronouns. Referential pronouns
which refer to referential expressions do not require any
other constraint but the Principle C«described in the previous
section. However, syntax should provide some other
constraints on the second and third types of pronouns. In
this section let us consider the characteristics of these

pronouns and their constraints.

30

2.2.2.1 Bound Pronouns

Now let us take (1.iii) into consideration” .An anaphoric
pronoun can also be interpreted as a bound variable if it is
used anaphoric on quantifiers and bound by those quantifiers:

(16) a. Every boy; thinks hei is smart.

b. Some bo q thinks that hei is smart.
c. No boy} thinks that hei loves Mary.
In (16) the pronouns he are interpreted as a variable

bound by the quantifiers every, some and he in (17),

 

respectively. The pronouns in (16) will belong to (1.iii)-
type pronouns. Neale (1990) also refers to (1.iii)-type of a
pronoun as an anaphor as a bound variable.

(17) a. [every x: boy x](x thinks that x is smart)

b. [some x: boy x](x thinks that x is smart)
c. [no x: boy x](x thinks x loves Mary)

With respect to the pronouns as bound variables, Evans
(1980), Reinhart (1983), May (1985) and others suggest the
syntactic constraint that a pronoun anaphoric on a quantifier
is bound by that quantifier iff that quantifier c-commands
that pronoun. Furthermore, May (1985) has shown that the
syntactic level is the level in which the configurational
constraint of c-command should take place.

May (1985) proposes two principles for quantifier scope
and the pronouns as bound variables:

(18) a. The scope of x is theset of nodes that x

31
c-commands at LF. (p.5)
b. A pronoun is a bound variable only if it is
within the scope of a coindexed quantifier.
(p.21)
The principle of (18) explains why the pronouns in (16) can.be
interpreted as bound variables.
This constraint is not applicable to referential
pronouns. Take the following example.
(19) a. Hisi mother loves Johni.
b.* Hisi mother loves every boyi.

In (19.b) the pronoun his cannot be coindexed with evegy

boy, since it violates (18). However, (19.a) the pronoun he
can refer to John. So syntax should provide a contra-index

between hie and everv boy for DRT.

While syntax simply specifies contra-indices, DRT will
resolve pronominal binding. Let us consider (20).

(20) Every chess set comes with a spare pawn. It is

taped to the top of the box. (Kadmon P. 292)

In (20), on.the preferred reading, the phrase every cheee
takes a wide scope over the phrase a pawn, and also a wide
scope over the pronoun ii if the pronoun i; is coindexed with
the phrase e pawn. If the first sentence in (20) is not
processed through. DRT, it may* be difficult to jproperly
interpret the second.sentence containing the pronoun under the
scope of the prior sentence. The sentence of (20) can,

however, be represented as (21) if Roberts’s (1989) discourse

32
subordination. applies for (20) here. ll: is a case of
quantificational subordination.

(21)

 

 

 

u v w
chess(u) pawn(v)

-> spare(v)
come_with(u, v)
box(w)
taped_to_top(v, w)

 

 

 

 

 

 

 

 

 

2.2.2.2 E—type Pronouns

Here are more complex examples of the anaphoric pronouns
which are not interpreted straightforwardly, however:
(22) a. John bought some donkeysi and Bill vaccinated
themi.
b. [some x: donkey x](John bought x and Bill
vaccinated x)
c. [some x: donkey x] (John bought x) & [the y:
donkey y & John bought y](Bill vaccinated y)
In (22.a) the pronoun them and the quantifier phrase we
donkeys seem to be anaphorically related, but the phrase eeme
donkeye does not c—command the pronoun 121E . Hence the

pronoun them cannot be bound by the quantifier as defined in

 

(18).
Furthermore, the pronoun in (22.a) cannot, actually, be
interpreted as a bound variable, either. For example, if we

translate (22.a) to a logical form as a bound variable

33
interpretation of ghee, then (22.a) will be logically
represented as (22.b). .As Evans (1980) points out, (22.b) can
be true even in case John bought 10 donkeys and Bill
vaccinated 9 donkeys. So the pronoun hhem in (22.a) should
not be interpreted as the case of bound variable like (22.b).

Rather, following Evans (1980), naturally, (22.a) is
interpreted in such a way that John bought some donkeys and
Bill vaccinated all the donkeys that John bought. The
pronouns which are anaphoric on the quantifiers but not bound
by those quantifiers are a (1.iv) or E-type pronoun.

An E-type pronoun is also applicable to an indefinite
singular phrase. In this case it can have two different
readings: either maximal or existential, depending on the
context in which it occurs. Consider Heim's (1982) example:

(23) Every man who owned a slavei owned itsi offspring.

In this sentence the indefinite a elave does not c-
command the pronoun ihe, but can be co-indexed with it. Thus
the pronoun should be understood as an E-type pronoun. This
sentence will read a maximal reading like (22.c) so that if a
man owned three slaves then.he would own all the offsprings of
all the slaves that he owned.

According to Evans, an E-type pronoun and its indefinite
antecedent should be uniquely existential:

(24) John bought a donkey; and Harry vaccinated.it,.

That is, in (24) there must be at most one donkey and at

least one donkey that John bought, and Harry vaccinated that

34
donkey. Such, uniqueness cannot be carried. out by the
existential quantifiers themselves; the E-type pronoun makes
the indefinite antecedent unique.

If Evans’s analysis is true, then the E—type pronouns are
very similar to the Russell’s analysis of definite
descriptions. Following Russell's analysis of definite
descriptions, the singular definite descriptions show
uniqueness. F m'example, Russell (1919) analyzes the sentence
THe author of Waverley was Scotch. as follows: (p.218)

(25) a. x wrote Waverlev is not always false.

b. if x and v wrote Waverley. x end v are identical
is always true.

c. if x wrote Waverley. x was Scotch is always
true.

If E-type pronouns have similar properties to definite
descriptions, can we treat E-type pronouns like definite
descriptions?

Evans argues that E-type pronouns have their own
"referents fixed by descriptions" like proper names. We may
call this analysis E-type pronouns as rigid designators‘.

Evans argues that it is necessary to analyze the E-type

pronouns as rigid designators, since the E-type pronouns

 

‘This term is from Kripke (1972). An expression is a rigid
designator if it makes reference to the same object in every
possible world.in which it exists, although it does not necessarily
exist in all the possible worlds (Kripke 1972, p.111). Kripke
argues that proper names are rigid designators, and the definite
descriptions are nonrigid.designatorx The examples are "Nixon" and
"the president of the United States".

35
cannot be used to refer to an empty set, and do not show the
scope ambiguity over sentential operators that overt
descriptions do. Let us take the following:
(26) *John bought no donkey,, and Harry vaccinated it,.
(27) a. A man, murdered Smith, but John does not believe
that he, murdered Smith.
b [the x: man x & x murdered Smith]
(John does not believe that (x murdered Smith))
c. *John does not believe that
([the x: man x & x murdered Smith] (x murdered
Smith)).
The E-type pronouns should have their references fixed by
description, but it fails in (26). Furthermore, the E-type
pronoun he in (27) has de re reading but no de dicto reading.
If it were definite description, then it should also show the
de dicto reading.
Now let us discuss the representation of E-type pronouns
in DRT.
Kamp (1981), first of all, suggests that (28) be
represented as (29).

(28) Every Farmer who owns a donkeyq beats it,.

36

 

 

 

 

 

 

 

(29)
u v
farmer(u) -> it = v
donkey(v) beat(u, v)
own(u, V)

 

 

 

 

 

Then the DRS (29) with a universe UD and condition CD is
interpreted in such a way that (29) is true in a model M with
a set of individuals UM and an assignment function GM iff there
is a mapping function g from D to M such that for all g in
which g(u) is a member of GM(farmer) and g(v) a member of
(%(donkey) and <g(u) g(v)> belongs to Ch(Own), <g(u) g(v)>
belongs Ctheat). In other words every pair of ieime; and
donkey is truth-conditionally evaluated with respect to the
relations of heeh and eye. This representation demonstrates
Geach's analysis of (29).

Geach (1968) argues that the pronoun i; in (28) must be
regarded as a bound variable, bound by a donkey, which is
given wide scope. So he suggests that (30.a) should be a
logical form of (28).

(30) a. Every,"y [x is a farmerA y is a donkeyA x has y]

[x beats y].
b. everyx [x is a farmer " somey[y is a donkey A x
has y]] [x beats y].
If (30.b) is treated as the logical form of (28), it

might be hard to analyze the pronoun i; in (30.b), because e

denkey cannot C-command it. So, following Neale (1990),

37
(30.a) can be derived from (30.b) as follows: Since (some x p)
-> q and [every x] (p -> q) are logically equivalent where q
contains no free occurrence of x, the form (30.b) is logically
equivalent to (30.a). In this way, the indefinite NP’a donkey
gets universal force.

Although this DRS represents a maximal/universal reading
of an E-type pronouns, this analysis does not reflect an
existential reading of E-type pronouns, as described below.

The first case is that the universal reading of a donkey
sentence will give rise to a problem if we replace eyeiy with
meeh. Let us take Chierchia's (1992) example: (P. 119)

(31) a. Most farmers that have a donkeyq beat it,.

A

b. Most,” [farmer(x) " donkey (y) x owns y] [x

beats y]

A A

c. Mostx [farmer(x) [a y [donkey (y) x owns
y]]] [a y[x beats y]]
If we interpret (31.a) in such a way that the quantifier

most has scope over the pair of farmer and donkey as in

 

(31.b), then it will give us a wrong truth condition in some
situations. For example, let us consider the situation in
which there are five farmers such that one farmer has 99
donkeys and each of the other four farmers has one donkey, and
that only the farmer who has 99 donkeys beats his donkeys. In
this situation (31.b) is true. However, (31.a) is naturally
interpreted as false in that situation. In order for (31.a)

to be true, 3 or 4 of five farmers should beat their own

38
donkeys regardless of the number of donkeys which farmers
have. In this sense the quantifier pee; in (31.a) should be
quantified over farmers only and not over donkeys as in
(31.c). In the same sense we need another representation in
which the quantifier' eyeiy; in (28) is interpreted. as a
quantification over farmer only.

The second problem of (29) is that the representation of
(29) cannot express that the indefinite NP, a donkey, has the
existential force in.Russell’s ternn It will be more clear if
we again look at the example which Chierchia takes: (P. 115)

(32) Every person who has a credit card., will pay his

bill with it,.

In (32) the pronoun i; can.be coreferential with.a credit
peed, and, if so, (32) means that every person will pay his
bill with at least and at most one credit card, and does not
indicate that every person will pay his bill with his all
credit cards. If quantifier eyeiy is quantified over a pair
of a person and a gredit card, (32) will give us a wrong truth
condition.

The third. problem. is ‘that (29) cannot explain the
uniqueness presupposition that the pronoun i_t_:_ in (28) carries.
That is, (28) should be evaluated with regard to the farmers
who have just (and at most) one donkey. Even though a farmer
who has 99 donkeys does not beat his own donkeys at all, it
does not determine the truth condition of (28). The universal

reading of e donkey in (28) cannot express such a uniqueness.

39
As we have seen so far, in addition to the representation
(29), we need another DRS to represent the characteristics
described just above. So we can tentatively represent the
first and the second characteristics for (28) in the following
way:

(33)

 

 

 

u
farmer(u)

 

v -> beat(u, it)
donkey(v)
own(u, v)

 

 

 

 

 

 

 

 

 

 

 

 

Then, how can we maintain the anaphoric relations between
E-type pronouns and their antecedents, since a donkey is too
deeply embedded in the antecedent DRS?

To solve this problem, Kadmon (1990) suggests that the
DRS may contain "implicated, accommodated and contextually
supplied information". (p. 273) Following Kadmon (1990), a
speaker and.a hearer can revise the DRS, based upon contextual
information. In other words, the DRS's are basically
constructed with the syntactic structures as an output of
syntactic analysis, and furthermore, to make an utterance
felicitous in the context, can be revised with the
information which is contextually supplied or specified. Let
us take some examples to look at the contextual accommodation,

as Kadmon (1990) suggests.

4O

(34) a. John owns three donkeys.

b. John. owns three donkeys,. Harry ‘vaccinates
them,.
First of all, (34.a) is represented as (35) by a

construction rule assumed.

(35)

 

u v
John(u)
donkey(v)
|v| = 3
own(u, v)

 

 

 

This DRS expresses that John.owns at least three donkeys.
That is, among a set of the donkeys that John owns, we may
talk about any subset with three donkeys.

However, following Kadmon (1990), (34.a) may be uttered
in a context containing a scalar implicature which implies
that John does not own more than three donkeys. Then (35)
will be revised by a speaker and a hearer as (36).

(36)

 

u v
John(u)
donkey(v)
|v| = 3
own(u, v)

 

 

w
donkey(w) —> w in v
own(u, v)

 

 

 

 

 

 

 

 

 

Now let us look at (34.b), a case of E-type pronouns.
When a speaker and a hearer processes the first part of

(34.b), he constructs a DRS like (35). When he utters or

41
hears the second part, however, (35) cannot be felicitous, if
it simply accommodates the second utterance as (37), since
(37) does not express the uniquenesss that (34.b) implies. So
a speaker and a hearer revise (37) as (38) in processing the

second utterance in order to accommodate the uniqueness.

 

 

 

 

 

(37)
u v
John(u)
donkey(v)
|v| = 3
own(u, v)
Harry(w)
them = v
vaccinates(w, v)
(38)
u v w
John(u)
donkey(v)
|v| = 3
own(u, v)
Harry(w)
them = v

vaccinates(w, v)

 

 

x
donkey(x) -> x in v
own(u, X)

 

 

 

 

 

 

 

 

 

Turning to (28), it should accommodate such a uniqueness
in the DRS, and such accommodation also solves the problem in
which E-type pronouns are not anaphorically linked to the
antecedents. ‘That is, a contextual accommodation can.not only
make a utterance felicitous but also, by recovering some
missing information, provide an antecedent for E-type pronouns
which can be anaphorically linked to it. Then (28) can be

represented as (39).

42

 

 

 

 

 

 

 

 

 

 

 

 

 

 

(39)
u v
farmer(u) donkey(v)
own(u, v)
v
donkey(v) w
own(u,v) donkey(w) -> w =v
-> own(u, w)
w'
donkey(w) -> w =v it = v
own(u, w) beat(u, v)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Now let us take a close look at how (39) can be built.
First a speaker or a hearer builds (40.a) for (28). A.sub-DRS
is embedded into the antecedent DRS, since the quantifier
eyehy here takes quantification over farmer, not over
farmer-donkey pairs. Further, a donkey is unique relative to
every farmer; So the accommodation.modifies (40.a) as (40.b).
Finally, the reconstruction or copy of the sub-DRS in the
antecedent DRS into the consequent DRS will build (39) as a
result.

(40) a.

 

 

 

u
farmer(u)

 

v -> beat(u, it)
donkey(v)
own(u, v)

 

 

 

 

 

 

 

 

 

 

 

 

43

 

 

 

 

 

 

b.
u
v farmer (u)
donkey(v) -> beat(u, it)
own(u, v)
w
donkey(w) -> w =v
own(u, w)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Now let us take into consideration the curious copy
process mentioned just above. Kadmon (1990) has mentioned
that the copy process is required to provide the antecedent
for the E-type pronoun in (39), but has not pointed out any
constraints on the copy: e.g. when the reconstruction should
take place or when it should be blocked. Regarding a
constraint on the reconstruction, we may think of uniqueness
and/or maximality. That is, the copy should take place iff
the antecedent implies uniqueness. Ihi (39) e donkey shows
uniqueness and the copy is applicable to it.

With this constraint we can explain such a bad sentence
as *Eveg farmer who owns no donkey, beats it,i In this
sentence no dpnkey does not imply uniqueness at all, and the
copy cannot take place. If the copy is blocked out, the
pronoun i_t cannot have any antecedent, since no donkey is
deeply imbedded in the antecedent DRS like in (40.a).

Furthermore, we can also explain such a bad sentence as

44

*John bought no donkey,, and Harpy vaccinated it,. We can

represent the sentence as (41).

 

 

(41)
u v
John(u)
w
[donkey(w) |
bought(u, w)

 

 

 

Harry(v)
vaccinated(v, it)

 

 

 

In (41) the DRS containing the antecedent donkey is
subordinate to the DRS containing the pronoun ih, and hence
donkey cannot be accessible to it. Further, no donkey does
not imply uniqueness, and so the copy is also blocked.

In the connection with pragmatic accommodations, (a
question has been raised whether or not pronominal
interpretations are purely’ based. upon such. pragmatic or
contextual accommodations. IEvans (1977) and.Heim (1990) argue
that some syntactic factors significantly have effect on the
interpretation or linking of E-type pronouns and their
antecedents. The pragmatic accommodation approach to the
e-type pronouns cannot explain the following sentence: (Heim
1990 p.165)

(42) a. Every man who has a wifeq sits next to her,.

b. *Every married man sits next to her.

(42.a) can be interpreted in a way that a wife is linked

to hep. Although every married man in (42.b) implies eyeiy

man who hae a wife in (42.a), however, (42.b.) cannot be

ii

(D
)
:2

ot‘x

Q r».
.e“,

-
A m h.

n _I
(D

45
interpreted like (42.a). If we build the DRS for (42.b), we
would get (43)5. Here in (43) the pragmatic accommodation

would. recover' some contextual information. and. provide a

 

 

 

 

 

 

 

 

 

 

 

 

 

 

candidate for the pronoun. But such interpretation is not
natural.
(43) a.*
u v
man(u) woman(v)
marry(u, v)
v
woman(v) -> w
marry(u, v) woman(w) —> w=v
marry(u,v)
w
woman(w) -> w = v her = v
marry(u,w) sit_next_to(u, v)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

5If we assume that there is an implicit argument for married
we may represent (42.b) as (43.b); otherwise we may represent
(42.b) as (43.a). Let us assume that (43.b) is correct. Then a
sentence like John.wee killed.may be represented as (i) rather than
the (ii).

(i) d = <{u}, {John(u), kill(-,u)]>.

(ii) d = <{u} John(u), killed(u) >.

A notation '-' in (i) will be explained later in this section.

46

 

 

 

 

 

 

 

 

 

 

 

 

 

 

b.

u v
man(u) wife(v)
married(u) has(u, v)
v w
wife(w) wife(w) -> w=v
has(u, v) —> has(u, w)

w

wife(w) -> w =v her = v

has(u, v) sit_next_to(u, v)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Now we may take two alternatives into consideration to
explain such sentences as (42). The one is to implement a
linking mechanism of E-type pronouns and their antecedents in
syntax, and the other to put more constraints on pragmatic
accommodations.

Evans (1977) and.Heim (1990) take the former approach.and
proposes that E-type pronouns should be syntactically linked
with.their antecedents and.that the well-formedness conditions
for E-type pronouns should be formulated in terms of the
syntactic antecedent-anaphor relations. Heim (1990) tries to
construct E-type pronouns’ linking with their antecedents at
the syntactic level.

Following Heim (1990), the E-type pronouns are linked
with their antecedents by coindexation at LF‘, and this

coindexation also constrains the range of the function that

 

6The LP is a syntactic level whish is assumed to be posited
after a surface level.

47
interprets the E-type pronouns. For example, the coindexing
in (28) between a donkey and i; constrains the function that
has a set of donkeys as its range. Heim (1990) treats the
E-type pronoun-anaphoric relations with the following
transformational rule (P. 170):
(44)XSYNP,Z=>1234+25
l 2 3 4 5
conditions: 4 is a pronoun
2 is of the form [S NPi S]
6 7

This stipulative rule applies to E-type pronouns at LF
and copies the "antecedent (= term 6) and the antecedent’s
scope (= term 7) into the position of the pronoun (= term.4)".
For example, it converts (45.a) into (45.b).

(45) a. [everyg [farmer(x) that [[a.y donkey(y)]2 [x owns

y]]]]1 [x beats itz].

b. [every, [farmer(x) that [[a.y donkey(y)]2 [x owns
y]]]]1 [x beats [it2 [[ay donkey(y)]2 [x owns
all ] 1] .

As already mentioned before, the syntactic level is
problematic to represent the pronominal references, since
intersentential relations cannot be explained. The
transformational rule (44) is problematic. The rule not only
specifies the linkage between an antecedent and the pronoun
with stipulation, but also has conceptual problem.

Following (44), the reconstruction or copy process can

48
take place, regardless of how far S (term=2) is from the
pronoun (term = 4). It seems to be not true, because
pronominal references take place under some kinds of locality
conditions. Now there is only one alternative left with us:
more constraints on pragmatic accommodation. Let us consider
how we can represent (42.b) in DRT.

There might be two options. The one is to prevent the
theory from accommodating such contextual information as that
in (42.b). That is, a. DRS can contain some types of
contextual information only--e.g. uniqueness as Kadmon (1990)
proposed” The other is to allow the theory to accommodate the
contextual information that (42.b) implicitly has, and to make
some constraints on the introduction of discourse referents.
Leaving for further research the question about which
alternative is more adequate in the theory, but here I
tentatively take the second alternative.

As I have mentioned before, deictic pronouns, proper
nouns, demonstrative NP’s and indefinite NP’s can introduce
discourse referents. Furthermore, if these NP's are
explicitly uttered in a linguistic way, they can add new
discourse referents to a DRS. More specifically speaking, the
discourse referents are introduced not in a contextual or
nonlinguistic way but in a purely linguistic way. If a
speaker does not linguistically express something about an
object, it will never be a discourse referent.

For example, in (42.a) e wife is expressed linguistically

49
and introduces a referent which can be an antecedent for a
pronoun heig However, in the case of (42.b), married.of,eye;y
married man cannot introduce a discourse referent to a DRS,
because:a context or background.knowledge cannot introduce any
discourse referent. Hence the representations of (43.a) and
(43.b) are ill-formed, regardless of the linkage between the
pronoun he; and v in wife(v) or woman(v). So a speaker or a

hearer should construct the DRS (46) for (42.b).

 

 

 

 

 

 

 

 

 

 

 

 

(46) *
u
man(u) -> sit_next_to(u, her)
marry(u, -)
The notation ’-' of marry(u, -) in (46) means don't care.
That is, marry(u, -) means that u is married but we don’t care
if u marries whoever she is. Truth-conditionally speaking,
marry(u, -) is true in a model M iff u belongs to a universe

UM and there is a non-empty set of marry relations u has in M.

2.3 Pronouns and Antecedent Search in DRT

Even though Kamp (1981) does not specifically describe
how the antecedent is assigned to an anaphor, there is said to
be an algorithm to find an antecedent in a set of discourse
referent introduced in a DRS in some way.

One of the important factors in determining possible

antecedents is the grammatical agreement with respect to

50
person, gender and number.

Another factor is precedence. Reinhart (1983) argues
that the linear precedence in a sentence is irrelevant because
of the following examples: (p. 34)

(47) Near him,, Dan, saw a snake.

(48) S

/ I \
PP NP VP
I I |
Near him Dan Saw a snake

So Reinhart (1983) proposes the c-command rule to
interpret the pronominal coreference. She argues that (47) is
structurally represented as (48) , and the antecedent can
c-command its anaphor and be accessible to the anaphor.

However, we cannot explain (49) with c-command rule.

(49) a. His,‘wife loves John,.

b. S
/ \
NP VP
| / \
his wife V NP

loves John
In (49) John cannot c-command hie but they can be
coreferential in fact. So I will propose that the precedence

should be defined in terms of a DRS.

51
(50) a. An antecedent should precede its anaphor in a
DRS.
b. A discourse referent x precedes a discourse
referent y iff:
(i) x < y in a universe U in DRS where < is the
order of introduction, or
(ii.) a DRS which contains x is superordinate to
a DRS which contains y.
(51) A DRS x is superordinate to a DRS y iff
(i) x contains y, or
(ii) x is an antecedent DRS of an If—then DRS whose
consequent DRS is y.
(52) Superordination is reflexive, asymmetric, and
transitive.
In the case of (47) and (49) the proper nouns are
introduced into the top level of that DRS as Kamp (1981)

suggests, and then always accessible to the anaphoric pronouns

 

 

as (53).
(53) a. b.
u v u v
Dan(u) John(u)
snake(v) wife(v)
saw_near(n, v, u) wife_of(v, u)
love(v,u)

 

 

 

 

 

 

Now we might raise a question about the introduction of
a discourse referent into the top level of DRT. If it is

true, how can we explain the ungrammaticality of (54)? As in

(53), Dan places a referent at the top of the DRS, and

 

52
accessible to he. Then we cannot distinguish (53) from (54)
in terms of a DRS at all.

(54) a.* Near Dan“ Ike saw a snake.

 

 

 

 

b.
u v
Dan(u)
snake(v)
saw_near(n, v, u)
Now let us remind ourselves of the level of DRT. As

mentioned in section 1.3 and 2.2, the syntactic analysis is an
input to DRT. .At the level of syntactic analysis the anaphora
like reflexives and reciprocals are determined, and only the
disjoint reference, not anaphoric relation, of pronouns are
represented. In (54) hep_ and. he should have disjoint
reference; otherwise it would violate the principle C. Then
the ungrammaticality of (54) is actually determined at
syntactic level, not at DRT level.

Since the concept of precedence is taken in DRT, it is
necessary to make further modification of Kamp’s DRT.
Precedence implies that the discourse referents are ordered in
some way; so the universe U should be assumed to be not a
simple unordered set of discourse referents but a set of
internally ordered referents. This modification can be
related to the locality to determine the antecedent—anaphoric
relations to be discussed just below.

Another factor to determine the antecedent—anaphoric
relations is locality. There might be several types of

locality conditions, Inn: I will consider only locality in

53

time. .As I described just above, the discourse referents are
placed in the order in.which they are introduced into the DRS,
and the ordering will be kept as the conversation goes on.
When.more than one antecedent is available, it is assumed that
the entity which is referred to the most accessible to the
anaphor. Then we can say that the anaphoric function is to
reorder the discourse referents already introduced in the
discourse by referring to its antecedent.

Let us take an example for this. Let us assume that
there is a context in which three referents, u, v, and w,

have been introduced which refer to John, Bill and Mery,

 

respectively. Further, they are assumed to have been
introduced in that order: Meiy is the most recently
introduced, and so on. Now a speaker uses a pronoun he to
refer to a referent in a context. Then w is the most
accessible to the pronoun.heJ but does not grammatically agree
with it. So the next candidate is v now. If there is no
other factors to prevent it from being coindexed with the
pronoun, the pronoun will refer to that v. Then the
reordering of the discourse referents will take place: u, w,
and v are in that order. If a speaker uses a pronoun he once
more after this, it refers to v again.

Another factor is the common knowledge. Let us take an
example for this. When a speaker utters the sentence (55.a),
the hearer will understand that Bill was injured if we have

the entailment rule (55.b) in the common.knowledge or lexicon:

54
(55) a. John hit Bill, and he was injured.
b. hit(x,y) -> injured(y)

Of course, he may be coreferential with dphh. However,
for a speaker to make a hearer understand (55.a) in that way,
the speaker must provide enough information to override the
common knowledge expressed in (55.b).

Another factor is a modality of a sentence. Roberts
(1989) introduces the modal subordinate constructions with
respect to anaphoric relations as follows: (p.683)

(56) a. If John bought a bookqihe'll be home reading it,

by now.
b. Itflq_is a mystery story.

(57) a. If John bought a book, he’ll be home reading it,

by now.
b. It, will be a mystery story.

The sentences in (57) are an example of a modal
subordination. The truth value of the sentence (57.b) is
determined conditionally in the possible worlds in which
(57.a) is true. That is, the modality of (57.a) takes scope
over (57.b). For this case we may say that (57.b) is modal-
subordinate to (57.a). Since (57.b) is under the modal scope
of (57.a), it is possible for the pronoun ih in (57.b) to be
coindexed with a book in (57.a).

However, the pronoun i; in (56.b) cannot be coindexed
with a book in (56.b), since (56.b) is a factual sentence

without any modal force and cannot be subordinate to (56.a).

3. Computing Discourse Representation Structures

It is not a trivial task to translate surface strings
to discourse representation structures (DRS’s). In this
chapter let us consider some algorithms to construct DRS’s
from English surface strings as an input.

Computing DRS’s requires two separate processes. The
first process is to take as an input a discourse (or a set
of sentences in a discourse), parse with phrase structure
rules and translate the input into DRS's. The second
process is to translate DRS's into Prolog facts and rules
and embed them in a model to simulate the model-theoretic
semantics. In this chapter their implementations will be
discussed in terms of Prolog as a logic programming
language, based upon Johnson and Klein (1986), Covington and
Schmitz (1988), Covington, Nute, Schmitz, and Goodman

(1988), and Covington (1994).

3.1 Phrase Structure Rules and Building DRS's

The phrase structure rules are augmented with two types
of feature structures: syntactic and semantic feature
structures. The syntactic features are used to control the
parsing process, and the semantic features to build DRS's
during the parsing process. Unification allows the feature

structures to be passed from one node to another in a

55

56
parsing tree. Thus the phrase structure rules should
specify the information about word order and the way of
unifying feature structures.
To help understanding the parsing procedure, let us
describe the phrase structure rules, and then the feature
structures and general unification mechanisms, and then, the

rules of DRS constructions by unification.

3.1.1 Phrase Structure Rules

The parser here will accept discourses which observe
the following phrase structure rules. For simplicity the
phrase structure rules here represent only the word order.
The phrase structure rules are written in definite clause
grammar (DCG) notation. The string in brackets is a
terminal string.

(1) a. discourse --> sentence, discourse.

b. sentence --> np, vp.
c. sentence --> np, aux, vp.
d. sentence --> [if], sentence, [then], sentence.
e. sentence --> aux, np, vp.
f. np --> n.
np --> [].
np --> n2, rc.
i. np --> n2.

j.' n2 --> det, n1.

57
k. n1 --> adj, n1.
1. n1 —-> n.
m. vp --> v, np.
n. vp —-> v.
o. vp --> adj.
p. vp --> np.

(1.a) helps the parser scan a sentence one by one until
the discourse is empty. (1.b-e) are the rules to parse
sentences. (1.b-d) represent declarative sentences
including copulars. (1.e) is for questions. (1.d) is a
rule for if—then sentences.

(1.f—l) are for noun phrases including relative
clauses. (1.f) is a rule for proper nouns and pronouns.
(1.g) is for a trace or gap in relative clauses. (1.h) is a
NP rule including relative clauses. (1.i-l) are for common
nouns.

(1.m-p) are the rules for verb phrases. (1.m) is for
transitives, and (1.n) is for intransitives. (1.o-p) are
for copular predicates.

In the subsequent sections we will describe how the
phrase structure rules in (1) accommodate feature structures

and their unification.

3.1.2 Feature Structures and Unification

The parser here implements a unification—based grammar,

case,

58
using GULP 3 (Graph Unification Logic Programming) as an
extension to Prolog written by Covington (1994). In the
unification-based grammar each word in a sentence is assumed
to have a set of features. The features specify the
function of a word in a sentence. By unification the
features are passed through the nodes in the parsing tree.
The feature structures and mechanism to unify (or merge)
features are thus important.
The parse uses the following partial feature structures
to control the parsing process and build DRS’s, where a
feature structure is represented as eih and e is an feature
attribute and h is the value of that feature: (cf. Covington
and Schmitz (1988) p.4)
(2) a. syntactic features7
syn: index:

class:

argl :

argZ:

b. semantic features
sem: in:
out :
res: in:
out :

scope: in:

 

7It is also possible to add other syntactic features such as
number, person to control the parsing procedures.

59
out:
The roles of the features in (2) are summarized as
follows:
index has as its value a discourse referent
instantiated on nouns (common and proper nouns).

class has as its value common/proper, instantiated on

 

nouns, or transitive/intransitive, instantiated on verbs.

eigi and died are the discourse referents of subject
and direct object as arguments of a predicate, being unified
with their syn:index features.

sem:in is a list of one or more DRS’s which are
currently being processed.

semzout is the DRS after processing the current node.

sem:res and sem:scope are used for the logical
structure of the sentence.

In general, following Covington (1994), the
unification-based grammar uses the following mechanism of
the unification of feature structures A and B giving C:

(i) If a feature x occurs either in A or B exclusively,
then it also occurs in C with the same value.

(ii) If a feature x occurs in both A and B, then it
also occurs in C, and its value in C is the unifier of its
values in A and B.

Now feature values are unified as follows:

(i) If both A and B have atomic symbols as their

values, then their values must be the same atomic symbol.

60
(ii) If A has a variable as its value and B has a
variable as its value, then B has the same value as A.
(iii) If both A and B have variables as their values,
then the variables become the same variable.
(iv) If both A and B have feature structures as their
values, they unify by applying this process recursively.
To pass feature structures from one node to another in
a parsing tree, a feature is splitted into two sub-features,
one for input and the other for output. Consider the rule
for example: (Covington (1994) p.12)
(3) S -> NP VP
[sem: [in:X1] ] [sem: [in:Xl] ] [sem: [in:X2] ]
[ [out:X3]] [ [out:X2]] [ [out:X3]]
This rule specifies that the sem of the S has some initial
value of X1. X1 is then passed to the NP, which modifies it
in some way and gives X2 out. Then X2 is passed to the VP
for further modification. The output of the VP is X3, and
it becomes the output of the S. This feature passing is

used to build DRS's in parsing.

3.1.3 Representing DRS’s

A DRS is a set of discourse referents U and a set of
conditions C. It can be implemented by the following Prolog
predicate:

(4) drs(U,C).

61
U is a list of discourse referents and C is a list of
predicates of these referents. Discourse referents are
represented with unique Prolog variables at the level of
DRT. When DRS's are embedded into a model, they will be, if
necessary, skolemized by instantiating uninstantiated Prolog
variables with unique integer number.

The DRS for questions and if—then sentences are
represented as query(drs(U,C)) and if_then(drs(U1,C1),
drs(U2,C2)) where drs(U1,Cl) is the antecedent and
drs(U2,C2) the consequent, although the DRS for declarative
sentences can be simply represented by (4).

In the next section let us consider how the phrase
structure rules accommodate feature structures and their
unification, and how the parser builds DRS's during the

parsing process.

3.1.4 Parsing Discourses

A discourse is a set of sequent sentences. As a
sentence in a discourse is parsed, the parsing rules pass
the discourse information from one rule to the next rule
using unification. We can rewrite (1.a), incorporating
feature structures, as in (5).

(5) parse_discourse(D) -->

{ D = sem:in:X,

S = sem:inzx,

62
S = sem:out:Y,

D1 sem:in:Y,

D1

sem:out:Z,

D = sem:out:Z
parse_sentence(s,[],[]),
endpunct, {l},

parse_discourse(Dl).

parse_discourse(D) -—>
{ D = sem:inzx,
D = sem:out:X }.

endpunct --> ['.’] ; [’?'].

The rule (5) will do the following things:

(i) If the discourse has one or more sentences, then
parse_discourse passes all the semantic features of sem:in
to parse_sentence.

(ii) Next, the rule parse_sentence (to be described
later) modifies sem:in, and passes it into parse_discourse
for further modification.

(iii) If the discourse has no sentence, then its
current semantics will be the output of parse_discourse.

While this procedure is recursively going on, sem:out
of parse_discourse accumulatively collects all information
the parsing rules have modified. Initially, the semantic

feature sem:in will contain an empty DRS in the form of

drs([],[l).

63

The phrase structure rule (1.h) for a declarative
sentence can be written as follows.

(6) parse_sentence(S,H1,H3) -->

{ S = sem:X,

NP = sem:X,

NP = sem:scope:Y,

VP = sem:Y,

NP = syn:index:Z,

VP = syn:arg1:Z
},
parse_np(NP,H1,H2),
parse_vp(VP,H2,H3).

This rule will do the following things:

(i) The rule parse_sentence passes all the incoming
semantic information to the NP rule, parse_np, and at the
same time the output semantics of the NP is the semantics of
the sentence rule.

(ii) The VP is the scope of the NP.8

(iii) The index (or the discourse referent) of the NP
is the same as the subject or argl of the verb. So argl
has the same value of the index of this NP.

We can write a parsing rule (1.e) for a question as
follows:

(7) parse_sentence(S,H1,H3) -->

 

8The scope of the NP is, in fact, that of the determiner. It
will be clear when we describe the semantics of determiners.

 

 

64

FM
(1)
ll

sem:in:X,

><
ll

[drs([],[l)].
NP = sem:in:X,

NP = sem:scope:Y,
VP = sem:Y,

NP = syn:index:Z,
VP = syn:arg1:Z,

NP

sem:out:[DRS],
S = sem:out:[drs([],[query(DRS)])]
}.
parse_aux(_),
parse_np(NP,H1,H2),
parse_vp(VP,H2,H3).

(7) will do the following things in addition to (6):

(i) It checks whether the incoming DRS is empty. That
is, it checks if a question is embedded in a discourse. If
it were, it would fail. Otherwise, this topmost DRS is
added, and passed to the NP rule. The NP and VP rule will
modify this empty DRS.

(ii) The DRS is then embedded into the original DRS,
forming query(drs) to represent a question.

This technique can generally apply to building
sub_DRS's including if-then sentences. The following phrase
structure rule is for if—then sentences.

(8) parse_sentence(S,[],[]) -->

{ S = sem:in:X,

65

$1 = sem:in:[drs([],[])|X],
S1 = sem:out:Y,
32 = sem:in:[drs([],[])lY],

82 = sem:out:[DRSZ,DRSl,drs(U,C)|Sup_drs],
S = sem:out:
[drs(U,[if_then(DRSl,DR82)[C]){Sup_drs]
}.
[if].
parse_sentence(Sl,[],[]),
[then],
parse_sentence(S2,[],[1).

This rule basically works in the same way as (7). It
inserts an empty DRS and then constructs the antecedent by
passing it to parse_sentence. After that it inserts another
empty DRS and builds the consequent in it, passing it to
parse_sentence. IF both the antecedent and the consequent
DRS are constructed, then this rule combines them into a
form of if_then(drs1,drs2), and embeds it in the condition
of the original DRS.

We can implement other sentences in (1) with feature

structures in the same manner as described above.

3.1.5 Parsing Noun Phrases

The following phrase structure rule is to parse the

common noun phrases as in (1.j).

66

(9) parse_n2(N2,H1,H2) -->
{ N1 = syn:class:common,
N2 = syn:X,
N1 = syn:X,

Det = sem:Y,
N2 = sem:Y,
N1 = semzz,
Det = sem:res:z
},
parse_det(Det),
parse_n1(N1,H1,H2).
This rule will do the following things:
(i) The syntactic class is specified as cpmmpn.
(ii) The semantics of the NP becomes the restrictor of
the determiner.
To make (ii) more clear, let us consider the semantics
of determiners. Following Johnson and Klein (1986),
semantically, a determiner has two arguments: restrictor and
scope. For example, the sentence A donkey is old. can be
translated into predicate logic as (a X: donkey(X)) old(X)
where e_h is the quantifier, donkey(X) is the restrictor,
and old(X) is the scope.
Accordingly, the DRS for A donkey is old. is
constructed by passing to the determiner e the NP dpphey for

the restrictor and the VP old for the scope.

The phrase structure rule of the determiner e is

67
written as follows:

(10) parse_det(Det) -->

([al ; [an]).

{ Det = sem:in:X,
Det = sem:res:in:X,
Det = sem:res:out:Y,
Det = sem:scope:in:Y,
Det = sem:scope:out:Z,
Det = sem:out:Z

}.

That is, this rule passes the semantics of a determiner
to the restrictor, then to the scope. The determiner e has
no special representation at all.

However, a determiner like eyeiy constructs a DRS in a
more complex way:

(11) parse_det(Det) -—>

[every],
{ Det = sem:in:X,
Det = sem:res:in:[drs([],[])|X],
Det = sem:res:out:Y,
Det = sem:scope:in:[drs([],[])[Y],
Det = sem:scope:out:
[Scope,Res,drs(U,C)ISup_drs],
Det = sem:out:
[drs(U,

[if_then(Res,Scope)|C])|Sup_drs]

68
}.
This rule works the same way as (8) for if—then
sentences.
Next, a lexical insertion rule for common nouns can be
represented in (12).
(12) parse_noun(N) ——>
[Word],
{ common_noun(Word,I,Sem),

append(Sem,C,NewC),

N = syn:indesz,

N = syn:class:common,

N = sem:in:[drs(U,C)|Sup_drs],

N = sem:out:[drs([I[U],NewC)|Sup_drs]

}.

This rule will do the following things:

(i) It generates a unique variable I for the common
noun.

(ii) It adds I to the list U, and adds the semantics of

the common noun to the list C.

It applies to the entire category of common nouns. The
semantics for a word pep is represented in (13).

(13) common_noun(man,X,[man(X),gender(X,m)] ).

This rule forms two predicates pep and gender. The

latter is used for antecedent search.

3.1.6 Parsing Verb Phrases

69
The phrase structure rule (1.m) for the transitive VP

can be represented as follows:

(14) parse_vp(VP,H1,H2) -->
{ V = syn:class:transitive,
VP = syn:X,
V = syn:X,
NP = sem:Y,
VP = sem:Y,
NP = syn:index:Z,
VP = syn:arg2:Z,

V = sem:W,
NP = sem:scope:W
}.
parse_verb(V).
parse_np(NP,H1,H2).
This rule will do the following things:
(i) It specifies that this VP is a transitive.
(ii) The index of this NP becomes the value of the argz
or direct object of this verb.
(iii) The semantics of this VP is the scope of this NP.
The lexical insertion rule for transitive verbs can be
implemented as in follows:
(15) parse_verb(V) -->
[Word],
{ transitive_verb(Word,A1,A2,Sem),

append(Sem,C,NewC),

70
= syn:class:transitive,
= syn:arglel,
syn:argZ:A2,

= sem:in:[drs(U,C)lSup_drs],

< < <1 :< <
n

= sem:out:[drs(U,NewC)ISup_drs]
}.

The lexical item for a transitive verb pep is
represented as in the following:

(16) transitive_verb(owns, X,Y,[own(X,Y)]).

This rule forms the predicate, using the indices that
are passed to it as syntactic arguments, and insert the
predicate in the condition of the current DRS. Thus the
verb rules depend on other rules to pass the values of argl

and arg2 to them.

3.2 Embedding DRS’s in Knowledge Base

To evaluate the truth of a DRS, we should embed it into
a model or knowledge base. It requires two processes. One
is to clean up some information from the DRS, which is
temporary at the level of DRT for some reason, and the other
to translate the DRS to appropriate Prolog facts and rules.
This section will consider the latter process only: how a
DRS can be translated into Prolog facts and rules, and
embedded into a model, based upon Covington, Nute, Schmitz,

and Goodman (1988).

71

3.2.1 Simple Statements

The statements about facts such as declarative
sentences should be added to the knowledge base for
evaluating/querying them later. However, it has a problem
if we add their DRS’s directly to the knowledge base.

For example, a sentence A farmer owne_e donkey. can be
translated into the following DRS:

(17) DRS([X,Y], [own(X,Y), farmer(X), donkey(Y)]).

We cannot add the DRS (17) itself to the knowledge
base, leaving the variables X and Y free. For Prolog facts
the free variables are implicitly interpreted as universal
quantifiers. So (17) would be interpreted in such a way
that any farmer owns any donkey.

To solve this problem, a unique value must be assigned
to an existentially quantified variable. To do this we can
assign a unique integer to it. That is, we can assign the
free variables x and y [1] and [2], respectively, and embed
the facts:

(18) farmer([1]). donkey([2]). owns([l], [2]).

This is a special process of skolemization in which
existentially quantified variables are replaced with unique

identifiers.

3.2.2 If—then Sentences

ab.

Cu

u.‘

72

A simple if—then DRS is equivalent to a Prolog rule.
For example the DRS for the sentence Every donkey is old.
is represented as:

(19) if_then(drs([X],[donkey(X)]), drs([],[old(X)])).

This is equivalent to the Prolog rule:

(20) old(X) :- donkey(X).

We may have a problem if the consequent, e.g. the
rightside DRS contains more than one predicate, as in the
following:

(21) Every donkey is old and gray.

(22) if_then(drs([x],[donkey(X)]),

drs([l.[old(X),gray(X)])).

A Prolog rule cannot have two or more predicates in its
consequent; hence (23) is not permitted in Prolog.

(23) old(X), gray(X) :- donkey(X).

When this if—then is embedded, the consequent should be
distributed into a series of ordinary Prolog rules. For
example, a rule of the form a. b, c. :- d. e. f is added to
the knowledge base as the three rules:

a :- d, e, f.

b :- d, e, f.

C :- d, e, f.

If the consequent of an if—then DRS introduces a new
variable, it is interpreted as an implicit existential
quantifier in Prolog. This is another representational

problem in Prolog. For example the DRS for the sentence

73
Every farmer owns a donkey. is represented as follows:

(24) if_then(drs([X],[farmer(X)]),

drs([Y],[own(X,Y),donkey(Y)])).

We might translate (24) to (25).

(25) donkey(Y), own(X,Y) ::- farmer(X).9

However, (25) means that every farmer owns every
donkey.

If we assign the donkey a unique integer as in (26),
then it means that every farmer owns the same donkey
identified as [2].

(26) donkey([2]), own(X,[2]) ::- farmer(X).

To represent the meaning that every farmer owns a
different donkey, it should be represented as follows:

(27) donkey([2lX]), own(X,[2lX]) ::- farmer(X).

The donkey has a dummy name that contains X so that the
dummy name depends on the assignment of X. Then if we
assign a farmer X with an integer [1], we can get a donkey
named [2,1]; if we assign another farmer X with [3], we can
then get a donkey named [2,3]; and so on. This is a process

of skolemization.

3.2.3 Questions

It is easy to translate query(drs) for questions into

 

9The symbol ::- is used temporarily to represent the complex
if—then DRS’s at the level of DRT.

74

Prolog. For example, the sentence Does a farmer own a
donkey? is translated into a DRS:

(28) query(drs([X,Y], [own(X,Y), farmer(X),

donkey(Y)l)).

Questions are understood as to evaluate the truth of a
DRS in the model defined by the current knowledge base. So,
we want to know whether there is an assignment of values to
X and Y such that farmer(X), donkey(Y), and own(X,Y) will
all be true. For Prolog queries the free variables have the
meaning of implicit existential quantifiers. So it does not
require any skolemization at all. Thus (28) can be
translated into a Prolog query as follows:

(29) farmer(X), donkey(y), own(X,Y).

3.3 Anaphoric Pronouns

3.3.1 Antecedent Search

As we have discussed in chapter 2, DRT searches in U of
drs(U,C) for an antecedent that agrees with the anaphor in
person, gender and number. DRT begins to search for
antecedents in the current DRS, and then goes up through
superordinate DRS’s of the current DRS.

Consider the following hypothetical DRS:

75

 

 

(30) D1

(12:)

D3 D5
D4 - D6
(:__l > DJ
D7
L]

 

 

 

 

 

 

 

 

 

 

 

 

 

In (30) D1 is superordinate to all DRS’s D2-D7 which it
contains. D3 is superordinate to D4, D5 and D6. D5 is
superordinate to D6. D2, D4, D6 and D7 are not
superordinate to other DRS’s.

If a pronoun is in D1, DRT will search for its
antecedent in D1. If it is in D5, DRT will search in D5, D3
and D1. If it is in D6, DRT will search in D6, D5, D3 and
D1. If it is in D2 DRT will search in D2 and D1. If it is
in D4 DRT will search in D4, D3 and D1. If it is in D7 DRT
will search in D7 and D1.

In this implementation the DRS construction algorithm
works with a list of DRS's. This list begins with the
current DRS under construction, followed by all
superordinate DRS's, so that all discourse referents in the
superordinate DRS’s are accessible to the current DRS.

Consider (30) for instance. At the moment D6 is being
constructed. The active DRS is D6, and three other DRS's,
D1, D3, and D5, are superordinate to it. Thus the DRS list
which the construction algorithm can work with at that

moment will be as follows:

76

(31) [D6, D5, D3([...],[D4]), D1([...],[D2,...])].

Note that D2 and D4 are already constructed, closed and

added into the conditions of D1 and D3, respectively.

Under the DRS constructions The sem:in feature is a

list containing the current DRS plus all DRS’s that are

superordinate to it rather than a single DRS. So the

antecedent of an anaphoric pronoun is found by searching the

list of discourse referents in each DRS in the DRS list.

This is done by the following rule for pronouns:

(32) parse_np(NP,H,H)

[he].

{ NP =

-->

sem:in:DRSlist,

member(drs (U, C), DRSlist),

member(Index,U),

member(gender(Index1,m),C),

Index==Index1,

NP =

NP

NP

NP

}.

syn:
sem:

sem:

sem

index:Index,
sc0pe:in:DRSlist,

scope:out:Result,

:Out :Result

The rule will do the following things:

(i) It selects drs(U,C) from the DRS list in the order

from the nearest superordinate DRS to the farthest.

(ii) It selects a discourse referent Index in U in the

order from the most recently introduced entity to the least.

77
(iii) It checks the gender. If it matches, the rule
will set the index of the current NP equal to the index of

the antecedent just found.

3.3.2 Subordinations and Anaphora

The basic mechanism described in the previous section
can in many cases find an antecedent for pronouns (1.i-iii)
described in chapter 2. However, this antecedent search
mechanism has some problem with an anaphor which is under a
modal/quantificational subdorinations. For example,
consider (33).
(33) Every farmer owns a donkey,. It, is old.
To construct the DRS for (33), the DRS building
algorithm takes the following steps:
(1) An empty DRS drs([],[]) is passed to
parse_discourse. So the list contains
[
drs([], [1)
].
(ii) The antecedent DRS, drs([X],[farmer(X)]) is
constructed, and the list contains
[
drs([], [1)

drs([X], [farmer(X)]),

78
(iii) The consequent DRS, drs([Y], [donkey(Y),
own(X,Y)]) is constructed, and the list contains:
[
drs([], [1).
drs([X], [farmer(X)]),
drs([Y], [donkey(Y), own(X,Y)1),
].

(iv) The antecedent and consequent DRS’s are combined
into the if—then DRS if_then(drs([X], [farmer(X)]), drs([Y],
[donkey(Y), own(X,Y)])), and embedded into the topmost DRS,
and the list contains:

[

drs([]. [if_then(drs([X], [farmer(X)]), drs([Y],
[donkey(Y), own(X,Y)]))l)
J.

The currently constructed DRS can be diagrammed in

 

 

 

 

 

(34).
(34) D1
D2 D3
x Y
farmer(x) -> donkey(y)
own(x. y)

 

 

 

 

 

 

 

 

 

Next, DRT parses the sentence it is old. in the
discourse. In this case DRT meets the pronoun ii, and tries
to find its antecedent in the DRS list. However, the
algorithm fails to find an antecedent for the pronoun ih,

since the DRS list contains only the topmost DRS. Thus the

79
DRS construction algorithm wrongly builds the complete DRS

(35) for (33).

 

 

 

 

 

 

 

 

 

(35) D1
D2 D3
X Y
farmer(x) -> donkey(y)
own(x, y)
it = ?
old(?)

 

 

 

To represent (33) correctly, we expect (36).

 

 

(36) D1

D2

 

 

D3

 

x Y
farmer(x) -> donkey(y)
own(x. Y)

it = y
old(y)

 

 

 

 

 

 

 

 

 

This problem is due to the DRS building procedure
rather than to the search procedure. After the algorithm
processes a sentence it will close all other DRS’s but the
topmost DRS. DRT should be able to reopen the closed DRS
for antecedent search. This is done by the following rules:

(37)

parse_sentence(S,[],[]) -->

{S = sem:in:DRS,

DRS = [drs(U,[if_then(DRSl,DRSZ)IC])ISup_drS],

Sl sem:in:[DR82,DRSI,drs(U,C)lSup_drs],

Sl

sem:out:[DRSBI_].

80
S = sem:out:
[drs(U,[if_then(DRSl,DR83)IC])ISup_drs]
}.
parse_sentence(S1,[],[1).

This rule opens the already closed if—then and its
consequent DRS, makes the consequent DRS a current DRS, and
starts searching for an antecedent. If DRT finds the
antecedent in this DRS, then it will insert the analysis of

this sentence in it.

3.3.3 The Accommodations and Problems

To represent its DRS for the existential reading of E-
type pronouns, we should implement the accommodation
mechanism, as described in chapter 2. However, we have some
problems with the implementation of the accommodation
mechanism in DRT with two respects: The one is that it is
difficult to know which predicates should be accommodated
for the E-type pronouns. The other is to translate into
Prolog the complex if—then DRS’s which the accommodation
process builds.

Regarding the former problem, let us consider the DRS

for a sentence Evegy farmer whp pwns a donkey beats it. The

algorithm first constructs the antecedent as follows:

81

(38)

 

 

u
farmer(u)

 

v
donkey(v)
own(u,v)

 

 

 

 

 

 

 

 

 

After that the consequent DRS is constructed. At this
point the algorithm fails to search the antecedent of the
pronoun ih, since the potential antecedent is deeply
embedded in the antecedent DRS of the if-then DRS. For the
next option DRT reopens the DRS immediately embedded in the
just previous superordinate DRS, and search for a potential
antecedent. Now the algorithm can find the antecedent
donkey for ih. However, we should update the DRS containing
donkey by accommodating the following DRS for the correct
representation and interpretation.

(39)

 

 

u
farmer(u)

 

v
donkey(v)
own(u,V)

 

 

w
donkey(w) -> w =v
own(u, w)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

To accommodate (39), we should know what are the

82
predicates to be accommodated: in this case they are
donkey(V) and own(U,V). However, it is difficult to know
the name of the predicates in Prolog.
If we could build a DRS for the above example, as in

(40), we would meet with the latter problem.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

(40)
u v
farmer(u) donkey(v)
own(u, v)
v
donkey(v) w
own(u,v) donkey(w) -> w =v
-> own(u, w)
w
donkey(w) -> w =v it = v
own(u, w) beat(u, v)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

The DRS (40) has a complex if—then: the antecedent
contains another if-then DRS, and the consequent also
contains another if—then DRS. Such a complex if—then DRS is

hard to be translated into Prolog rules.

4. Conclusion and Further Research

In this thesis I have discussed the pronominal
classification, representation and interpretation in DRT,
and their implementations in Prolog.

First of all, I have discussed Kamp's (1981) DRT,
assuming that DRT is a linguistic level as an interface
between the syntactic analysis and model-theoretic
semantics. Although syntax analyzes some types of scope and
anaphoric relations like reflexive and reciprocal anaphora
within a sentence, it is not adequate to represent
intersentential scope and anaphoric relations such as modal
subordinations, quantificational subordinations, and
pronominal references. On the other hand, DRT is adequate
in building intersentential relations in a discourse,
providing a unified discourse structure.

To understand pronominal phenomena we have reviewed
pronominal uses, and classified the pronouns into deictic
and anaphoric pronouns in terms of discourse referents at
DRT. Deictic pronouns introduce discourse referents, and
anaphoric pronouns select as an antecedent one of the
discourse referents already introduced in a DRS.

The well-formedness of some types of anaphora should be
determined at syntax. If an anaphor is reflexive or
reciprocal, the well-formedness is determined by explicit

constraints on how they should be coindexed, and if an

83

84
anaphor is pronominal the well-formedness is determined by
expressing how they should not be coindexed.

To determine the antecedents of anaphoric pronouns, we
should specify some factors such as grammatical agreements,
a hierarchical relations between discourse referents, a
local relations between an anaphor and its antecedent,
pragmatic accommodations, inferences with common knowledge,
and so on.

Of anaphoric pronouns, specially, the properties of E-
type pronouns are very complicated. To understand the
problems of E-type pronouns clearly, we have reviewed other
researchers’ works in comparison, and discussed the ways in
which the logical forms of E-type pronouns are derived and
interpreted. For E-type pronouns the syntactic analysis of
E-type pronouns is more problematic, and so DRT should be
able to resolve E—type pronouns with the accommodation of
pragmatic or contextual information.

To demonstrate the mechanism of the pronominal
references in DRT, first of all, we have discussed some
algorithms and techniques to translate English surface
strings into a DRS in Prolog, and to embed them into a model
for evaluation. Although this implementation can resolve
pronominal references in many cases, it still has some
problems with E-type pronouns in two respects: a predicate
identification and a complex if—then embedding. Relating to

the former problems, one possible solution is to develop

85
another knowledge representation in Prolog. For example, we
may represent the sentence A farmer owns a donkey in Prolog
as follows:

(1) pred(’farmer’,[1]),

pred(’donkey’,[2]),
pred2(’own’,[1],[2])
rather than:

(2) farmer([l]), donkey([2]), own([l],[2]).

We will leave this development open for further
research.

We have discussed how to represent pronouns and find
their potential antecedents in DRT. But we did not mention
anything about which one of the potential antecedents is a
real antecedent for a pronoun. Following Sidner (1983) and
Grosz (1986), we may be able to find out a real antecedent
for a pronoun if we can determine a focus in a discourse.
The idea is that a pronoun typically refers to a discourse
focus. If it is correct, we can efficiently find a real
antecedent for a pronoun if we develop some algorithm to
find a discourse focus.

Many factors seem to be involved in determining a
discourse focus. Syntax, semantics, inferential knowledge,
pragmatic knowledge, and some additional phonological stress
are required to detect a current discourse focus and focus
shift. Conversely, pronouns sometimes play an important

role in guessing a discourse focus. If DRT can somehow

86
accommodate some of these factors it will have a great
effect on pronominal references. It will be left for

further research.

APPENDIX

This appendix presents some codes to build DRS’s and

search for pronominal antecedents.

/* phrase structure and DRS construction rules */

/* Discourse --> Sentence, Discourse */

parse_discourse(D) -->

{ D

sem:in:X,

S

sem:in:X,

S

sem:out:Y,

D1 = sem:in:Y,

D1 = sem:out:Z,

D = sem:out:Z
l.
parse_sentence(S,[],[]).
endpunct, {l},

parse_discourse(Dl).

parse_discourse(D) -->

87

88

D
II

sem:in:X,

D
II

sem:out:x }.

endpunct --> [’.'] ; [’?’].

/* rules for sentences */
/* We should have implemented the Subj-Verb agreement

* (number and person), but it is not our purpose here. */

/* S --> NP, VP */
parse_sentence(S,H1,H3) -->
{ S = sem:X,

NP

sem:X,
NP = sem:scope:Y,

VP = sem:Y,

NP syn:index:Z,
VP = syn:arg1:Z

parse_np(NP,H1,H2),

parse_vp(VP,H2,H3).

/* S --> NP, Aux, VP */
parse_sentence(S,H1,H3) -->

{ S = sem:X,

89

NP = sem:X,

NP = sem:scope:Y,

VP = sem:Y,

NP = syn:index:Z,

VP = syn:arg1:Z
}.
parse_np(NP,H1,H2),
parse_aux(_),

parse_vp(VP,H2,H3).

/* S —-> [if], S, [then], S */
parse_sentence(S,[],[]) -—>
{ S = sem:in:X,

SI sem:in:[drs([],[])IX],

81 = sem:out:Y,
82 = sem:in:[drs([],[])IY],

$2

sem:out:[DRSZ,DRSl,drs(U,C)ISup_drs],
S = sem:out:[drs(U,[if_then(DRSl,DRS2)|C])|Sup_drs]
}.
[if].
parse_sentence(Sl,[],[]),
[then],

parse_sentence(SZ,[],[]).

/* S --> Aux, NP, VP */

parse_sentence(S,H1,H3) -->

9O
sem:in:X,

[drs([].[])],

#5.
U)
ll

><
ll

NP = sem:in:X,

NP = sem:scope:Y,

VP = sem:Y,

NP = syn:index:Z,

VP = syn:argl:Z,

NP = sem:out:[DRS],

S = sem:out:[drs([l,[query(DRS)])]
l.
parse_aux(_),
parse_np(NP,H1,H2),

parse_vp(VP,H2,H3).

/* pronominal antecedent search in subordinations */
parse_sentence(S,[],[]) -—>
{ S = sem:in:DRS,

DRS = [drs(U,[if_then(DRSl,DR82)|C])|Sup_drs],

Sl sem:in:[DR82,DRSl,drs(U,C)|Sup_drs],

SI sem:out:[DRS3I_].

S = sem:out:[drs(U,[if_then(DRSl,DRS3)|C])|Sup_drs]

}.

parse_sentence(Sl,[],[]).

/* rules for noun phrases */

91
/* NP --> N */
parse_np(NP,H,H) -->
{ N = syn:class:proper,

NP = syn:X,

N = syn:X,

NP = sem:reszY,

N = sem:Y,

NP = sem:in:Z,

NP = sem:res:in:z,

NP = sem:res:out:W,

NP = sem:scope:in:W,

NP = sem:scope:out:Zl,

NP = sem:out:Zl

},

parse_noun(N).

/* pronouns */
/* This routine will search for pronominal antecedents
*which agree in gender. If we implement person and number
* features for nouns, then we modify this, too. */
parse_np(NP,H,H) -->
( [he] ; [him] ) .
{ NP=sem:in:DrsList,
member(drs(U,C),DrsList),
member(Index,U),

member(gender(Index1,m),C),

92
Index == Indexl,

NP

syn:index:Index,

NP = sem:scope:inzDrsList,
NP = sem:scope:out:Drs,
NP = sem:out:Drs

parse_np(NP,H,H) -->
[it],
{ NP=sem:in:DrsList,
member(drs(U,C),DrsList),
member(Index,U),

member(gender(Index1,n),C),

Index == Indexl,

NP = syn:index:Index,

NP = sem:scope:inzDrsList,
NP = sem:scope:out:Drs,

NP = sem:out:Drs

}.

/* We can add more pronouns hereafter. */

/* a trace in relative clauses */
/* NP --> [1 */
parse_np(NP,[rel(Index)|Rest],Rest) -->

U.

{ NP = sem:in:X,

93

NP = sem:scope:inzx,
NP = sem:scope:outh,
NP = sem:out:Y,

NP = syn:index:Index

/* NP --> N2, RC */
parse_np(NP,H1,H3) -->
{ NP = syn:X,

N2 = syn:X,
RC = syn:X,
NP = sem:in:Y,
N2 = sem:in:Y,
N2 = sem:out:Z,
RC = sem:in:Z,
RC = sem:out:W,
NP = sem:outzw
},
parse_n2(N2,H1,H2),

parse_rc(RC,H2,H3).

/* NP --> N2 */

parse_np(N2,H1,H2) --> parse_n2(N2,H1,H2).

/* N2 --> Det, N1 */

parse_n2(N2,H1,H2) -—>

{ N1
N2

N1

Det

N2

N1

Det

}.

94

syn:class:common,

syn:X,

syn:X,
= sem:Y,
= sem:Y,
= semzz,

= sem:res:Z

parse_det(Det),

parse_n1(N1,H1,H2).

/* N1 --> Adj, N1 */

parse_n1(N1,H1,H2)

{ N1

Adj

Nla

N1

Nla
Nla
Adj
Adj

N1

},

-->
= syn:X,

= syn:X,

syn:X,
= sem:in:Y,
= sem:in:Y,

= sem:out:Z,

sem:in:Z,

= sem:out:W,

= sem:out :W

parse_adj(Adj),

parse_n1(Nla,H1,H2).

95
/* N1 --> N */

parse_nl(N1,H,H) --> parse_noun(Nl).

/* a rule for relative clauses */
parse_rc(RC,H1,H2) -->

{RC

syn:index:Index,

RC

sem:X,

S = sem:X
}.
([who];[whom];[which];[that]),

parse_sentence(S,[rel(Index)|Hl],H2).

/* rules for verb phrases */
/* VP --> V, NP */
parse_vp(VP,H1,H2) -->

{ V = syn:class:transitive,

VP = syn:X,
V = syn:X,
NP = sem:Y,
VP = sem:Y,
NP = syn:index:Z,

VP syn:argZ:Z,

V = sem:W,
NP = sem:scope:W

}.

parse_verb(V),

96

parse_np(NP,H1,H2).

/* VP --> V */

parse_vp(VP,H,H)

—->

parse_verb(VP).

{ VP

/* VP --> Adj */

parse_vp(VP,H,H)

syn:classzintransitive }.

-->

parse_adj(VP),

{ VP
VP

VP

syn:
syn:
syn:

/* VP --> NP */

class:adjective,
argl:X,

index:X

parse_vp(VP,H1,H2) -->
{ VP = sem:X,
NP = sem:X,
VP = syn:arg1:Il,
NP = syn:index:Iz,
NP = sem:scope:in:[drs(U,C)|Sup_drs],

NP

}.

sem:

scope:out:[drs(U,[(I1=I2)}C])}Sup_drs]

parse_np(NP, H1,H2),

97

{ VP = syn:class:common }.

/* lexical rules */

/* Aux --> [does] ; [is] */

parse_aux(_) --> [does] ; [is].

/* Det --> [a] ; [an] */
parse_det(Det) -->
([al ; [anl).

{ Det

sem:in:X,

Det = sem:res:in:X,
Det = sem:res:out:Y,
Det = sem:scope:in:Y,
Det = sem:scope:out:Z,

Det = sem:out:Z

/* Det --> [every] */
parse_det(Det) -->
[every],

{ Det sem:in:X,

Det = sem:res:in:[drs([],[])IX],
Det = sem:res:out:Y,

Det = sem:scope:in:[drs([],[])[Y],

98

Det

sem:scope:out:[Scope,Res,drs(U,C)|Sup_drs],

Det sem:out:[drs(U,[if_then(Res,Scope)|C])|Sup_drs]

/* N --> Proper_Noun */
parse_noun(N) -->

[Word],

{ proper_noun(Word,I,Sem),

add_topdrs(I,Sem,DRSlist,DRSlist1),

N = syn:index:I,

N = syn:class:proper,
N = sem:in:DRSlist,

N = sem:out:DRSlistl

add_topdrs(I,S,[drs(U,C)], [drs([IIU],NewC)]) :-
append(S,C,NewC), !.

add_topdrs(I,S,[HIT],[HIL]):- add_topdrs(I,S,T,L).

/* Proper_Noun --> [john] */
proper_noun(john, X, [named(X,’John’), gender(X,m)]).

/* We can add more proper nouns hereafter. */

/* N --> Common_Noun */
parse_noun(N) -->

[Word],

99
{ common_noun(Word,I,Sem),

append(Sem,C,NewC),

N = syn:index:I,

N = syn:class:common,

N = sem:in:[drs(U,C)|Sup_drs],

N = sem:out:[drs([IlU],NewC)ISup_drs]

/* Common_Noun --> [man] */
common_noun(man,X,[man(X),gender(X,m)] ).
common_noun(donkey,X,[donkey(X),gender(X,n)] ).

/* We can add more common nouns hereafter. */

/* Adj --> Word */
parse_adj(Adj) -->
[Word],
{ adjective(Word,I,Sem),
append(Sem,C,NewC),

Adj = syn:index:I,

Adj sem:in:[drs(U,C)ISup_drs],

Adj

sem:out:[drs(U,NewC)|Sup_drs]

/* Word --> [01d] */
adjective(old, X,[old(X)]).

/* We can add more adjectives hereafter. */

lOO

/* V —-> Transitive */

parse_verb(V)

[Word],

-->

{ transitive_verb(Word,A1,A2,Sem),

append(Sem,C,NewC),

V = syn:class:transitive,
V = syn:arg1:A1,
V = syn:argZ:A2,
V = sem:in:[drs(U,C)|Sup_drs],
V = sem:out:[drs(U,NewC)|Sup_drs]
}.
/* Transitive -—> [owns] */

transitive_verb(owns, X,Y,[own(X,Y)]).

transitive_verb(own, X,Y,[own(X,Y)]).

/* We can add more transitive verbs hereafter. */

/* V --> Intransitive */

parse_verb(V)

[Word],

-->

{ intransitive_verb(Word,Arg,Sem),

append(Sem,C,NewC),

V

V

syn
syn

:class:intransitive,

:argl:Arg,

sem:in:[drs(U,C)ISup_drs],

 

101

V = sem:out:[drs(U,NewC)lSup_drs]

/* Intransitive --> [brays] */
intransitive_verb(brays, X,[bray(X)]).
intransitive_verb(bray, X,[bray(X)]).

/* We can add more intransitive verbs hereafter. */

REFERENCES

Allen, J. F. (1995). Natural Language Understanding. 2nd
edition, Menlo Park, CA: Benjamin/Cummings.

Berwick, R. C. (1983). Computational Aspects of Discourse.
In M. Brady and R. Berwick (eds.), Computational Models
of Discourse (p.27-106). Cambridge, MA: MIT Press.

Berwick, R. C. (1989). Computational Linguistics. Cambridge,
MA: MIT Press.

Bosch, P. (1983). Agreement and Anaphora: A Study of the
Role of Pronouns in Syntax and Discourse. NY: Academic
Press.

Chierchia, G. (1992). Anaphora and Dynamic Binding.
Linguistics and Philosophy 15, 111-183.

Chomsky, N. (1981). Lectures on Government and Binding.
Dordrecht: Foris.

Cooper, R. (1979). The Interpretation of Pronouns. In F.
Heny and H. Schnelle (eds.), Syntax and Semantics 10.
NY: Academic Press.

Covington, M. A. (1994). Natural Language Processing for
Prolog Programmers. Englewood Cliffs, NJ:
Prentice-Hall.

Covington, M. A. (1994). GULP 3.1: An Extension of Prolog
for Unification-Based Grammar. ACMC Research Report
1994-06, University of Georgia.

Covington, M. A., Nute, D., Schmitz, N. and Goodman, D.
(1988). From English to Prolog via Discourse
Representation Theory. ACMC Research Report 01-0024,
University of Georgia.

Covington, M. A. and Schmitz, N. (1988). An Implementation

of Discourse Representation Theory. ACMC Research
Report 01-0023, University of Georgia.

102

103

Evans, G. (1977). Pronouns, Quantifiers, and Relative
Clauses. Canadian Journal of Philosophy 7, 467-536.

Evans, G. (1980). Pronouns. Linguistic Inquiry 11, 337-362.

Gazdar, G. and Mellish, C. (1989). Natural Language
Processing in Prolog: An Introduction to Computational
Linguistics. Reading, MA: Addison-Wesley.

Geach, P. (1972). Logic Matters. Oxford: Blackwell.

Grosz, J. (1986). The Representation and Use of Focus in a
System for Understanding Dialogs. In B. J. Grosz, K.

S. Jones and B. L. Webber (eds.), Reading in Natural
Language Processing (p.353-362). Los Altos, CA:
Kaufmann.

Heim, I. (1983). File Change Semantics and the Familiarity
Theory of Definiteness. In R. Bauerle, C. Schwarze,

and A. von Stechow (eds.), Meaning, Use, and
Interpretation of Language (p. 164—198). Berlin: Walter
de Gruyter.

Heim, I. (1990). E-Type Pronouns and Donkey Anaphora.
Linguistics and Philosophy 13, 137— 178.

Hobbs, J. (1986). Resolving Pronoun References. In B. J.
Grosz, K. S. Jones and B. L. Webber, (eds.), Reading
in Natural Language Processing (p.353-362). Los Altos,
CA: Kaufmann Publishers.

Johnson, M. and Klein, E. (1986). Discourse, Anaphora, and
Parsing. CSLI Research Report 86-63, Stanford
University.

Kadmon, N. (1990). Uniqueness. Linguistics and Philosophy
13, 272-324.

Kamp, H. (1981). A Theory of Truth and Semantic
Representation. In J. Groenendijk, T. Janssen, and M.
Stockhof (eds.), Truth, Interpretation, and Information
(p.1-41). Dordecht: Foris.

Kaplan, J. (1983). Cooperative Responses from a Portable
Natural Language Database Query System. In
M. Brady and R. Berwick (eds.), Computational Models of
Discourse (p.167-208). Cambridge, MA: MIT Press.

Kripke, S. (1972). Naming and Necessity. In D. Davidson
and G. Harman (eds.), Semantics for Natural Language
(p.253-355). Dordrecht: Reidel.

104

May, R. (1985). Logical Form: Its Structure and Derivation.
Cambridge, Mass.: MIT Press.

May, R. (1989). Interpreting Logical Form. Linguistics and
Philosophy 12, 387-435.

Neale, S. (1990). Descriptions. Cambridge. Mass: MIT Press.

Pereira, F.C.N. and B. Grosz (1994). Natural Language
Processing. Cambridge, MA: MIT Press.

Reinhart, T. (1983). Anaphora and Semantic Interpretation.
London: Croom Helm.

Roberts, C. (1989). Modal Subordination and Pronominal
Anaphora in Discourse. Linguistics and Philosophy 12,
683-721

Russell, B. (1919). Descriptions. From Introduction to
Mathematical Philosophy. London: G. Allen and Unwin
Ltd.,, 167-180.

Schank, R. and R. Abelson. (1977). Scripts, Plans, Goals,
and Understanding. Hillsdale, NJ: Lawrence Erlbaum
Associates.

Sidner, C. (1983). Focusing in the Comprehension of
Definite Anaphora. In M. Brady and R. Berwick (eds.),
Computational Models of Discourse (p.267-330).
Cambridge, MA: MIT Press.

Webber, B. L. (1983). So what can we talk about now? In
M. Brady and R. Berwick (eds.), Computational Models of
Discourse (p.331-370). Cambridge, MA: MIT Press.

Williams, E. (1977). Discourse and Logical Form.
Linguistic Inquiry 8, 101-139.

MICHIGAN STATE UNIV. LIBRARIES
[HlllllllmWI)Wlllllllllllllllllll“llllllllllllll
31293014057602