l
4".0!‘l. .(A
<0. Incl.

'
.t
x...

I

.1.
It.
2; ._. .
:35...) ‘ . Armand ref?
.1 ZZZI :4: lipkfivb ,
I I.‘ .3 f‘ 90"“-
It k ..
”an L .

 

 

: ‘ . .4! .Kiifi
1:2: .v , ‘ I: ;
L31... \ . . . .fl. “atria“...
it. , .. . . . , «an! .I.

.v. . E
t.-.)

it

1 Iii-it:

1 v! '95-?!-
v10]!!! 1'

{fivisill V
.oJcIO..DID!‘nO:!
v9.1.1 .tl
. vL'DVA:

3.53.
. . Av.
rt v. . ,Y-

 

.lo. 3 \to
.~.f ti.

 

‘ . 4...}.1
‘ ‘ . .4 t";
I I“» G ‘ v , . . y vi ‘3! I
III: . P .. - 4..
Lmh”. xi . . JILL H2”... 91th.
. v. a s :0 6.
MV ‘56:..25. Pa; pf919... a (I:

l-l .. vvl. : 5;...
141.90%“ 1".
' ‘ ‘ _

 

 

..
.424...»

’37.‘- Y1.

 

. . ‘ . I
. I.“ . ‘ ll;v1)" D...“
\ z. ughm “I? I. . . . . (91"111'2 o 0"...lrno‘|u.‘¢.
. . 0‘: I'll. .‘I o-.\ ‘ lu‘
i .. 2.». ”Ma ‘ ,5 I. 3 £33.22 _ : .
It 3A q I; v v ELI). . ..)3-1Jfi3l£.:’t. v.3 . I .0! 9. II 1 ‘ .
.2. on F u v I I» .32.... . ‘91:..1.’ 53:23:13.: 3J1... 2.3?! 1 A! . .7 £25.: . ‘ . |
”MI“ , a Jakarta?" 3 5...}. .. at: I. . : . :1... n r. . 3H. 1... l ) ‘

       

if... , . . a. . . :3an

 

livvnv t
v.

lcl‘.lv,| ‘

 

. Il’a , ‘A‘D>. I.

 

“€315

 

 

 

 

 

 

Illillllllllllllllilllllllllllllllll L

3 1293 00899 2764

This is to certify that the

dissertation entitled
CONCEPTUAL RETRIEVAL FROM CASE MEMORY
BASED ON PROBLEM SOLVING ROLES:
A GENERIC TASK ARCHITECTURE
WITH APPLICATION TO

JUSTIFICATORY REASONING IN TAX LAW
presented by

Vernon Eugene Wallingford II

has been accepted towards fulfillment
of the requirements for

Doctor of Philosophy degreein Computer Science

 

 

6%

Major professor

way 9;);

MSU is an Alfirmuu'w- Action/Equal Opportunity Insn'lution 0-12771

 

 

 

 

F LIERARY
Michigan State
University

‘L A

‘—

 

 

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

DATE DUE DATE DUE DATE DUE

firm ID 3 t ‘
+

/L‘

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

MSU Is An Affirmative Action/Equal Opportunity Institution
ammo-m

 

a

CONCEPTUAL RETRIEVAL FROM CASE MEMORY
BASED ON PROBLEM SOLVING ROLES:
A GENERIC TASK ARCHITECTURE

WITH APPLICATION TO
J USTIFICATORY REASONING IN TAX LAW

By

Vernon Eugene Wallingford H

A DISSERTATION

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

DOCTOR OF PHILOSOPHY

Department of Computer Science

1992

ABSTRACT
CONCEPTUAL RETRIEVAL FROM CASE MEMORY
BASED ON PROBLEM SOLVING ROLES:
A GENERIC TASK ARCHITECTURE

WITH APPLICATION TO
JUSTIFICATORY REASONING IN TAX LAW

By
Vernon Eugene Wallingford 1]

Given an understanding of the domain and the data stored, how can one locate in an
immense data base only those records or documents semantically relevant to a particular
topic of interest? This question frames the conceptual retrieval problem. Rather than trying
to solve the conceptual retrieval problem solely in terms of general knowledge about a
domain, this research advances the view that knowing the use to which retrieved items will
be put provides guidance in developing more useful indexing vocabularies and retrieval
methods. This dissertation focuses on conceptual retrieval for the purpose of case-based
justification in tax law, in particular for the area of captive insurance taxation. In this
context, a theory of conceptual retrieval is presented that elaborates an index vocabulary
and organization based on the roles that cases can play in justification.

This dissertation describes three primary products of this research: (1) a
methodology for the Functional Representation of justifications, (2) a model of conceptual
retrieval —an index vocabulary, an index organization, and a retrieval method —— motivated
by this methodology for representing justifications, and (3) a conceptual memory of
arguments based on this model. This conceptual memory is integrated with a Generic Task
architecture for justifying legal classifications. Finally, these concepts are implemented in a
knowledge-based system called CRISTA.

This research addresses the conceptual retrieval problem from a task-specific

perspective. The result is the identification of an indexing methodology that is closely

related to a particular problem-solving task and a particular case representation. In essence,
this work unites a task-specific theory of problem solving with ideas from case-based
reasoning about indexing in order to achieve a more complete picture of conceptual
memory. To related disciplines, one of the important contributions of this research lies in
its description of how one can employ knowledge of a device and its teleology in

constructing more effective and efficient case memories.

Copyright by
VERNON EUGENE WALLINGFORD II
1992

To my dearest Mary.

If all I ever accomplish
is to be worthy of your precious love,
I will consider myself a successful man.

ACKNOWLEDGMENTS

I now understand why people always recognize the contributions of so many others to their
thesis work. This is a long and grueling process, and I certainly needed a lot of help and
support to make it this far. I thank Jon Sticklen, my advisor, for his guidance and
intellectual stimulation these last five years. So much of what I learned will never appear in
this document. Thanks also to my dissertation committee, Rich Hall, Bill McCarthy, and
Tony Wojcik. Each in his own way has left an indelible mark on my professional and
personal views. Special thanks are due to Bill, for finally convincing me that this problem > i

was worth attacking. He was right.

I owe so much to my parents, Vernon and Linda Wallingford, my brother Anthony, and
my sisters, Nancy and Cathy. They have always believed in me — even when I was not
so sure myself. Their love and support often urged to me to try harder, and I thank them

for their encouragement.

Finally, I thank my wife, Mary. I can truthftu say that, without her love, encouragement,
support, and occasional prodding, this dissertation would never have been completed. She
has buoyed my spirit during down times and brought me incomparable joy during up

times. She is my partner and best friend.

vi

TABLE OF CONTENTS

List of Tables ........................................................ xi
List of Figures ....................................................... xii
Chapter 1. Introduction ............................................ 1
1.1 The Conceptual Retrieval Problem .................................... 1
1.1.1 Conceptual Retrieval for Legal Analysis .................. 2
1.1.2 Conceptual Retrieval in Tax Accounting .................. 3
1.2 Requisites of a Solution ................................................ 4
1.2.1 The Context of Tax Law .................................... 6
1.3 Intellectual Influences ................................................... 7
1.3.1 The Generic Task Approach ................................ 7
1.3.2 Toulmin’s Model of Argument ............................. 8
1.3.3 Functional Device Understanding .......................... 8
1.4 Research Toward a Solution ........................................... 9
1.4.1 Objectives ..................................................... 9
1.4.2 Issues .......................................................... 10
1.5 A Synopsis of This Work ............................................... 11
Chapter 2. The Conceptual Retrieval Problem: An Analysis ..... 13
2.1 Introduction ................................................................ 13
2.2 The Traditional Solution: Key-Word Search .......................... 14
2.2.1 Conceptual Approaches to the Retrieval Problem ......... 16
2.3. Information Retrieval .................................................... 17
2.3.1 An AI-Based Model of Conceptual Retrieval .............. 19
2.3.2 AI-Based Approaches and Their Significance ............. 23
2.4 Cognitive Psychology .................................................... 25
2.4.1 Memory Organization and Access ........................... 27
Retrieval Based on Surface Features ....................... 27
Retrieval Based on Abstract Features ....................... 28
2.5 Case-Based Reasoning ................................................... 30
2.5.1 Issues in Case-Based Reasoning ............................ ' 31
2.5.2 Case Memory and the Indexing Problem ................... 34
The Use of Low-Level Features ............................ 35
The Use of Abstract Indices ................................. 36

vii

2.6 Major Themes in Memory Organization ................................
2.6.1 Case Retrieval for Human Reasoners .......................
2.6.2 Simple and Abstract Features as Indices ....................
2.6.3 Problem-Solving Goals as Abstract Indices ................
2.7 Conclusion ..................................................................

Chapter 3. Justificatory Reasoning in Tax Law: A Review of Past Work 43

3.1 Introduction ................................................................
3.2 A1 Approaches to Legal Analysis .......................................
3.2.1 Rule-Based Approaches ......................................

3.2.2 Exemplar-Based Approaches ................................

3.2.3 Case-Based Approaches ......................................
Ashley and Hypo .............................................

Branting and GREBE .......................................

3.2.4 Status of AI in Legal Analysis ...............................

3.3 Representing Justifications ..............................................
3.3.1 The Structure and Content of Justifications ................

3.3.2 Strategic Argument Representation ..........................

3.3.3 Tactical Argument Representation ...........................
Ashley and Hypo .............................................

Branting and GREBE ........................................

3.4 A Legal Domain: The Taxation of Captive Insurance Arrangements .
3.4.1 Insurance .......................................................

Risk Shifting ..................................................

Risk Distribution .............................................

3.4.2 Captive Insurance Arrangements .............................

3.5 Implications for Conceptual Retrieval ...................................
3.5.1 Integration of Problem Solving and Retrieval ..............

3.5.2 Representation of Domain Concepts and Justifications

3.5.3 Organization of Case Memory by Problem-Solving Roles
3.6 Conclusion ..................................................................

Chapter 4. A Problem Solving Architecture for Legal Justification
4.1 Introduction ................................................................
4.2 A Task Analysis of Legal Justification ..................................

4.2.1 The Subtask of Fact Abstraction ............................

4.2.2 The Subtask of Case Retrieval ...............................

4.2.3 The Subtask of Precedent Application ......................

4.2.4 A Control Strategy for Legal Justification ..................

4.3 The Problem Solving Architecture ......................................
4.3.1 Components ...................................................

4.3.2 Communication ................................................

4.3.3 Problem Solving Methods ...................................
Justification Generator .......................................

43
44
45
47

72
73
74
75

Fact Abstractor ............................................... 83

Situation Data Base .......................................... 86
Case Memory ......................... ' ........................ 86
4.4 Conclusion ................................................................ 86
Chapter 5. A Functional Representation of Justificatory Analysis 88
5.1 Introduction ................................................................ 88
5.2 Motivations ................................................................ 89
5.2.1 Toulmin’s Model of Argument .............................. 89
5.2.2 Functional Device Understanding ........................... 91
5.2.3 Viewing a Legal Case as a Device .......................... 94
5.3 Representing Legal Analysis in the FR ................................ 95
5.3.1 The Legal Case as a Device .................................. 95
Selection of Case Functions ................................ 96
Treatment of Multiple Opinions ............................ 97
5.3.2 Case Context as Device Annotation ........................ 98
5.3.3 Legal Issues as Function Identifiers ........................ 99
5.3.4 Justifications as Behaviors ................................... 102
5.3.5 A Complete Example of the Representation ................ 105
5.4 Conclusion ................................................................. 109
Chapter 6. A Conceptual Memory of Justifications ............... 113
6.1 Introduction ................................................................ 113
6.2 Index Vocabulary .......................................................... 115
6.2.1 Index Terms for Case Citation Queries ...................... 117
6.2.2 Index Terms for Justification Queries ....................... 118
6.2.3 Summary of Index Vocabulary ............................... 119
6.3 Index Organization ........................................................ 119
6.3.1 Index Organization for Case Citation Indices ................ 119
6.3.2 Index Organization for Justification Indices ................. 120
Viewing a Body of Case Law as a Whole .................. 120
The Issue Composition Hierarchy (ICH) ................... 122
The ICH as an Index into Case Memory .................... 123
6.4 The Case Retrieval Algorithm ............................................ 126
6.4.1 Match Knowledge in the ICH ................................. 128
6.4.2 Use of the Retrieval Algorithm for Automatic Indexing 129
6.5 Conclusion .................................................................. 131
Chapter 7. CRISTA: A Computer Program for Conceptual Retrieval 132
7.1 Introduction ................................................................ 132
7.2 The Implementation of CRISTA ........................................ 133
7.2.1 The Software Environment ................................... 133
7.2.2 CRISTA: The Top Level ..................................... 136
7.2.3 CRISTA: The Subagents ..................................... 137

Situation Data Base .......................................... 137

Fact Abstractor ............................................... 140

Case Memory ................................................. 147

7.3 Samples of CRISTA’s Problem Solving .............................. 152

7.3.1 Sample Problem #1 ........................................... 153

7.3.2 Sample Problem #2 ........................................... 157

7.4 Conclusion ................................................................. 159
Chapter 8. Comparisons to Related Work: Extensions and Elaborations 160

8.1 Introduction ................................................................ 160

8.2 A Functional Representation of Legal Justifications ................... 161

8.2.1 Representation of Justificatory Arguments .................. 161

Warrants ....................................................... 161

Function Specification ....................................... 163

8.2.2 The Issue of Case Granularity ............................... 165

8.2.3 Potential Impact on Legal Practice ........................... 166

8.3 Conceptual Retrieval based on Problem-Solving Roles .............. 166

8.3.1 The Use of Abstract Indices .................................. 167

8.3.2 Relationship to Goel and Hafner ............................. 168

Goel and Functional Indices ................................. 168

Hafner and Issue Discrimination ............................ 169

A Synthesis .................................................... 169

8.4 Conclusion ................................................................. 171

Chapter 9. Conclusion ............................................... 172

9.1 Introduction ................................................................ 172

9.2 Contributions of this Research .......................................... 172

9.2.1 Modeling Justifications as Abstract Devices ............... 172

9.2.2 Modeling Conceptual Retrieval

Based on Problem-Solving Roles ......... 173

9.2.3 Extending Generic Task Theory ............................. 174

9.3 Avenues for Future Research ............................................ 174

9.3.1 The Functional Representation of Cases .................... 175

9.3.2 The Model of Conceptual Retrieval .......................... 176

9.3.3 The Problem Solving Architecture ........................... 176

9.3.4 Practical Matters ............................................... 177

9.4 Final Discussion .......................................................... 178

Appendix A. Legal Case References ................................ 180

Bibliography ........................................................... 182

VOQQGUMBUN—i

LIST OF TABLES

A RUBRIC Production Rule .................................................
A Control Strategy for Legal Justification ...................................
Channels of Communication in the PSA ....................................
Viewing a Legal as a Device in the FR ......................................
Language Grammar: A Legal Case ..........................................
Language Grammar: Case Context ..........................................
A Sample Case Context Frame —- Humana [1989] ........................
Language Grammar: Case Issues ...........................................
Language Grammar: Function ...............................................
A Sample Issues Frame — Humana [1989] ................................
Language Grammar: Justification ...........................................
Language Grammar: Warrants ...............................................
A Language for Representing Legal Analysis ..............................
A Complete Case Description — Humana [1989] .........................
The FR for Legal Justifications ..............................................
The FR for Legal Justifications and Toulmin ...............................
A Language for Representing Legal Analysis ..............................
A Case Retrieval Algorithm for Searching the ICI-I ........................
A Case Indexing Algorithm ..................................................
CRISTA’s Control Strategy — Abstract Level .............................
CRISTA’s Control Strategy — Message-Passing Level ...................
Channels of Communication in CRISTA ....................................
A Sample of CRISTA’s Fact Variables ......................................
A Sample Case from CRISTA — Humana [1989] .........................
The Facts of Sample Problem 1 ..............................................
Cases Returned by CRISTA on the First Pass of Sample 1 ...............
CRISTA’s Case Indexing Algorithm ........................................
The Case Description for Harper [1991] .....................................

Hemqamauu.‘

LIST OF FIGURES

A Sample Issue/Case Discrimination Tree ................................... 21
The Four Kinds of Issue Pointer into Case Memory ....................... 22
The Four Kinds of Link Pointer into Case Memory ........................ 23
Toulmin’s Model of Argument ................................................ 55
The Task of Justification ....................................................... 72
The Analogical Method of Justification ...................................... 73
The Task of Justification by Precedent ....................................... 75
A Task Analysis of Legal Justification ....................................... 76
A Problem Solving Architecture for Legal Justification .................... 79
A Sample Structured Matcher ................................................. 84
A Sample Hierarchical Classifier ............................................. 85
Toulmin's Model of Argument ............................................... 90
An Argument in Toulmin's Representation ................................. 91
A Fragment of an ER fora Clothespin ...................................... 93
A Sample Justification — Humana [1989] ................................. 104
The Justification Define Insurance Standard — Humana [1989] ......... 107
A Portion of the FR for the Clougherty Case ............................... 110
The Problem Solving Architecture for Legal Justification ................. 116
Issue Decomposition: Insurance .............................................. 121
A Sample Issue Composition Hierarchy ..................................... 123
The Four Kinds of Issue Node Pointer into Case Memory ................ 126
CRISTA’s Problem Solving Architecture ................................... 135
The Related Party Classifier ................................................... 143
The Captive Insurer Classifier ................................................ 143
The Insurance Provider Classifier ............................................ 143
The Standard Contract Matcher ............................................... 144
The Risk Diversity Matcher ................................................... 144
The Insurance Risk Matcher .................................................. 145
The Arm's Length Transaction Matcher ..................................... 145
The Unavailable Commercial Coverage Matcher ........................... 146
The Business Purpose Matcher ............................................... 146
CRISTA’s Issue Composition Hierarchy (Part 1) .......................... 150
CRISTA’s Issue Composition Hierarchy (Part 2) .......................... 151

gamers;

INTRODUCTION

1.1 The Conceptual Retrieval Problem

The arrival of the Information Age has radically changed how we do — and
conceive of doing — most knowledge-intensive tasks. With the widespread availability of
digital computers and expansive, inexpensive primary and secondary memory, industries
and individuals have accumulated vast stores of raw numeric data, full-text documents, and
interpreted knowledge. In many domains, such as molecular biology, the challenge
involves searching unrefined data for patterns that will lead to an understanding of the
domain itself. These fields require advances in database technology that will support
evolving and conflicting data models and the search for meaningful patterns. Other
domains, while sharing some of these concerns, face another primary challenge: given an
understanding of the domain and the data stored, how can one locate in an immense data
base only those records or documents relevant to a particular t0pic of interest? This
question frames the conceptual retrieval problem.

Traditional approaches to information retrieval (IR) have relied on syntactic
phenomena, employing key-word indices into a data base of full-text or structured-text
items. However, such key-word schemes cannot capture the rich semantic structure of
domains such as the law or medicine, and as a result efficient use of these tools can be
attained only by practitioners experienced in the creation of appropriate key-word queries.
In order to provide effective access to voluminous data for a broader array of users, these

domains require models of the data that incorporate semantic as well as syntactic

2

relationships. Users could then query the data base using the vocabulary of the domain and
rely on the conceptual model to identify documents relevant to the identified concepts;
Such a model will necessarily support a vocabulary that is abstract enough to reflect how
users conceptualize the domain but also is supported by a more concrete understanding of
domain relationships.

The problem of conceptual information retrieval parallels a human information-
processing problem that people solve routinely every day, that of episodic memory
retrieval. People continually face situations in which they are reminded of similar past
experiences; they can often use these experiences as aids for resolving the new situation.
While many such “remindings” involve experiences related only by surface features to the
current situation, other remindings reflect deeper conceptual relationships. This latter
phenomenon is especially marked in focused, problem-solving circumstances. Individuals
highly experienced in certain domains and tasks develop skill in identifying critical
conceptual similarities and differences among situations that determine how best to solve
new problems. Capturing this kind of knowledge about a domain would greatly enhance

the conceptual retrieval of information stored in computers.

1.1.1 Conceptual Retrieval for Legal Analysis

Nowhere is the problem of conceptual retrieval more important than in the task of
legal analysis. Given a description of a situation and a target classification desired for the
situation, one must generate a justification that supports the desired treatment. A
justificatory line‘of reasoning is an essential component of any legal classification. Without
substantiation, a legal claim carries no force in a legal system that is adversarial by design.

In a common law tradition, past classifications serve as the principal source of
justification for new classifications. Common law thus leads to a proliferation of case

records, which are then accumulated and organized for future reference. Specialists in

3

particular case law domains seem to organize their understanding of cases according to the
themes of prevailing lines of reasoning. That is. the use to which they put past cases
affects their understanding of the body of case law. Such specialists cannot themselves
remember all of the cases in a large domain, though, and must then turn to exhaustive
catalogs of past cases, whether in the printed literature or on-Iine computer data bases.
Legal analysis consists of two stages, and case research plays a role in each stage.
First, the researcher analyzes the current situation with respect to all sides of the law. The
goal of this stage is not to justify a particular classification but rather to consider all possible
treatments and lines of precedent. The researcher actively seeks out all past cases that could
reasonably bear on the outcome of the case at hand. From the assembled collection of
relevant statutes and case law, the researcher outlines potential conclusions regarding the
law and the grounds upon which such conclusions might be based. Second, the lawyer
evaluates these lines of the reasoning and writes a brief that justifies a particular treatment
of the case at hand. The brief will be written to emphasize elements of the most favorable
precedents and to downplay elements of the most unfavorable ones. It may also seek to

counter some of the arguments anticipated from the opposing counsel.

1.1.2 Conceptual Retrieval in Tax Accounting

Legal analysis plays a significant role in the practice of tax accounting. Tax
accountants have two primary concerns: the application of tax law in filing tax forms, and
the planning of future transactions to minimize tax effects. In unsettled areas of the law,
such as the taxation of captive insurance corporations, these tasks are made more difficult
by the uncertainty entailed in classifying new situations. Through legal analysis and case
research, the tax specialist can evaluate potential treatments of a situation or proposed

action. Consideration of past cases and the risk involved in adopting a particular

4

classification enables the accountant to make the best choices in the interest of his client,
within the bounds of the law.

Tax law offers an interesting domain in which to consider legal analysis. Like other
areas of the law, tax law evolves as new situations are considered and classified.
However, unlike most areas of the common law, taxation has a large body of statute that
defines and relates concepts in the domain. Tax law statute often does not rely on common-
sense notions of human relationships, as do “natural” law domains such as contracts and
torts. Rather, it creates a set of abstract entities and relations for the purpose of directing
tax policy. This abstraction makes the domain ideal for initial investigations of the legal
analysis task. By considering legal analysis in tax law first, one can focus on the task
itself,within the context of the artificial world in which tax law operates. The models of
legal analysis developed through study of tax law domains can then be applied to the law as
a whole, at which time one will be better equipped to focus on complexities native to more

natural legal domains.

1.2 Requisites of a Solution

Solving the conceptual retrieval problem will likely require moving beyond key-
word approaches to the modeling of semantic structure and task features in particular
domains. This type of domain modeling has long been one of the central themes of the
Artificial Intelligence (AI) community. Though AI research has often focused on the
development of problem-solving systems for tasks such as legal argumentation, some
researchers now believe that AI will make its greatest practical impact, at least initially, on
issues of information reuieval: “In the short run, Al’s principal contribution to society may
be to provide intelligent access to our vast data bases of information, in particular, helping

us to select and organize information that is relevant.” [Ashley 1990, page 6]

5

Yet such a contribution will likely profit from the understanding of task structures
gained through research on problem solving. This research aims to identify vocabularies
that are useful for describing and performing different types of problem solving. One of
the central advances of problem-solving research is the recognition that task-level terms
define particular roles for domain knowledge to play in performing the task. In this way,
these vocabularies provide leverage for organizing knowledge in the domain to support a
particular problem-solving task. The sheer volume of information to be searched in
domains such as the law demands the judicious use of such conceptual vocabularies as a
means for organizing computer memory.

Any useful model of conceptual retrieval will include three components: an index
vocabulary, an index organization, and an algorithm for retrieval. The index vocabulary
specifies the terms in which one formulates queries. In key-word systems, the index
vocabulary consists of all words appearing in a stored document. The index organization
specifies how terms in the vocabulary relate to one another. Key-word systems rely solely
on syntactic relations among words. Finally, the retrieval algorithm specifies how to
identify and select relevant items given a particular query. In key-word systems, search is
conducted using pre-built inverted files of words that point to their exact locations in
documents.

By making explicit the semantic relationships among index terms in a given
domain, one can tailor a retrieval algorithm that most effectively exploits the structure of
domain knowledge. Ideally, this sort of theory will generalize to the class of tasks and
domains characterized by the same types of knowledge structures. In order to support such
abstraction, the model must provide some mechanism for identifying appropriate index
terms in the domain and for identifying appropriate combinations of index organization and

search algorithm.

1.2.1 The Context of Tax Law

One such mechanism applicable to the task of legal analysis involves understanding
jusn'fications and the roles they play in the domain. Justifications serve not only to support
particular treatments of tax actions but also to implicitly define terms left unspecified in
statute. This latter role indicates that a justification may be relevant to queries about
concepts that are refined by it. In this way, justifications offer definitional content to a
domain’s concepts.

Likewise, justifications play a role in abstracting away detail from more settled
issues to allow closer examination of the critical issues at hand. By citing a relevant
precedent that establishes a particular classification, the new justification need not include
the detailed reasoning that supports the classification; it can focus on another issue that is
more problematic in the current case. This role also indicates how an understanding of tax
arguments can positively influence the selection of an index vocabulary. To support such
screening of detail, the index vocabulary should employ abstraction links among
justifications based on their conceptual relationships. In this way, an analysis of
justifications in the domain — coupled with an analysis of how domain specialists
understand the justifications - can provide significant guidance for organizing and
searching cases in a conceptually-organized data base.

Furthermore, knowledge of how specialists solve problems in a domain also guides
the development of effective case organization and retrieval schemes. This idea has
received considerable attention in the study of case-based reasoning, and the investigation
of problem-solving methods in conjunction with case memory offers new ideas for the
solution of the conceptual retrieval problem. The goals of a tax law specialist in classifying
a situation serve as valuable indicators of when certain past justifications will be of use.
Having an understanding of the structure of the problem-solving process itself furnishes

even more information about how cases can best be organized for efficient retrieval. Going

7

beyond design of index vocabularies to this sort of attention to index organization
represents a critical step in the realization of effective and efficient conceptual retrieval ‘

tools.

1.3 Intellectual Influences

This research reflects a variety of influences from a number of related fields, among
them knowledge-based systems, information retrieval, case-based reasoning, cognitive
psychology, AI in the law, and epistemology. The effects of these influences will appear
throughout this volume. However, a smaller number of ideas have had an important
impact on both the framing of the research problem and the conceiving of the proposed

solution. These pivotal ideas are described briefly below.

1.3.1 The Generic Task Approach

Chandrasekaran [1983, 1987, 1990] has advanced a task-specific theory of
problem solving founded on the notion of task/method/sub-task analysis of information
processing. For a given information processing task, characterized by its inputs and
outputs, one identifies different methods available for solving it. Each method is
characterized by a set of objects to be manipulated, a set of operators, and knowledge for
selecting and applying operators to objects. Some operators are not “primitive,” in the
sense that they establish sub-tasks to be solved by other means. This leads to a recursive
decomposition of tasks that halts upon reaching methods for which all operators are directly
applicable. In this sort of analysis, tasks correspond to steps in the solution of a problem,
and methods correspond to ways of realizing particular tasks, whether by decomposition or
by direct action.

The application of this methodology to a variety of real-world tasks and domains

has led to the identification of a set of ubiquitous methods for solving certain tasks that

8

arise in a variety of different circumstances. Chandrasekaran has termed these task/method
pairs generic tasks. Generic tasks have proven especially useful as models of problem-
solving types because they explicitly delineate (1) the types of domain knowledge required
for application of the method and (2) the nature of the control strategy for applying its
operators. Examples of generic tasks include hierarchical classification [Bylander and
Mittal 1986; Sticklen, Chandrasekaran, and Josephson 1987], routine design [Brown
1987], and structured matching [Bylander, Johnson, and Goel 1991]. These problem-

solving types play particular roles in the analysis of tax law argumentation proposed here.

1.3.2 Toulmin’s Model of Argument

A philosopher of science, Toulmin [1958] questioned the usefulness of traditional
work on logic and deductive inference for assessing the sorts of arguments actually made in '
most scientific fields. Recognizing that most argument did not correspond to the notion of
absolute proof, Toulmin abandoned mathematics as the foundation of logic and instead
adopted jurisprudence as his model. This move led to his extending the classical syllogism
to explicate more accurately the nature of persuasive argument — that is, to specify the
various roles that assertions can play in an argument and the relationships among these
roles. Touhnin’s model greatly enriches the vocabulary available for describing an
argument to include such natural terms as data, backing, warrant, qualification, and
conclusion. On this view, logic deals not with techniques of inferring but rather with

retrospective justification of a claim.

1.3.3 Functional Device Understanding
The Functional Representation (FR) of Sembugamoorthy and Chandrasekaran
[1986] provides a language for describing devices at multiple levels of abstraction based on

the device’s known functions or goals. In the FR, devices are decomposed into their

9

components, whose own functions can then be composed in order to understand the
functioning of the composed device. The causal behaviors of each device or sub-device are
indexed according to the functions they realize. Since functions of the higher-level device
can be expressed in terms of the functions of its components, the FR supports an
abstraction of behavioral detail between different levels of the device decomposition. This
sort of representation originally aimed at understanding physical devices but has since been
extended to the comprehension of “abstract” devices such as computer programs [Allemang
1990] and biological ecosystems [Sticklen and Tufankji 1992].

One can view a justification as an abstract device with the function of supporting
some claim given an initial set of assertions. Different components of the justification play
specific roles, such as to rebut a counterargument or to propose a hypothetical situation,
that in concert achieve the main goals of the justification. Considered this way, legal
analysis can be modeled with an FR in a way that is strongly reminiscent of Toulmin’s
model of argument. This functional representation can then serve as the basis for a scheme
to index a memory containing justifications and justification fragments. Goel [1989]
demonstrated the utility of using an FR as an index vocabulary for a case-based memory
used in a task of designing physical artifacts. However, he did not address the problem of
organizing and searching a large data base of designs. The great affinity between the FR
and Toulmin’s model of argument in large part inspired the work reported here aimed at

addressing that problem.
1.4 Research Toward a Solution

1.4.1 Objectives
This research focuses on conceptual retrieval for the purpose of legal analysis in

taxation, in particular, in the area of captive insurance corporations. In this problem-

10

solving context, a theory of conceptual retrieval will be presented that elaborates an index
vocabulary and organization based on the functional roles of justifications. This theory is
then generalized to the class of tasks for which it is appropriate. The present research has

four primary objectives:

(1) To develop a methodology for the functional representation of
justifications based on the ideas of Toulmin.

(2) To develop a model of conceptual retrieval —— index vocabulary,
index organization, and retrieval method — motivated by this
methodology for representing justifications.

(3) To design a conceptual memory of legal cases based on this model
and integrate it with a problem-solving architecture for justifying
classifications in the domain of tax law.

(4) To construct, as a proof of principle, a working knowledge-based

system based on these concepts.

1.4.2 Issues

Rather than trying to solve the conceptual retrieval problem solely in terms of
general knowledge of a domain, this project reflects the view that knowing how retrieved
items willibe used provides extra guidance in developing more useful indexing vocabularies
and methods. Such task-specific structures augment general-purpose indices with a richer
vocabulary for representing and retrieving information. Furthermore, vocabularies derived
from task structures can often elucidate the nature of the terms used in more general-
purpose indices. Certain theoretical questions arise concerning the relationship between

different models for conceptual retrieval, each based on different intuitions and principles.

11

To this end, this research proposes a view of conceptual retrieval that relates general-
purpose and task-specific models. -

The work presented here addresses several issues of importance to two related
research areas, Al in law and case-based reasoning. While this research does not propose a
new theory of legal analysis, it does formalize a view of legal analysis that is widely
accepted in the legal AI community. The generic-task problem-solving architecture thus
proposed embodies particular notions regarding the interaction between case memory and a
legal justification problem solver. The theory of conceptual retrieval advanced in this work
is relevant to many issues in the field of case-based reasoning, in particular the selection of
an index vocabulary, the organization of cases in memory, and the use of cases in problem
solving. With respect to these two research disciplines, the central contribution of this
work lies in its description of how one can employ lmowledge of a device and its teleology

in constructing more useful and efficacious case memories.

1.5 A Synopsis of This Work

The remainder of this volume describes in detail work undertaken to achieve the
objectives outlined above. The dissertation comprises three organizational sections:
problem analysis, problem solution, and theoretical discussion of the problem and solution.

Chapters 2 and 3 present the problem analysis. Chapter 2 analyzes the conceptual
retrieval problem itself and considers research aimed at resolving it from the information
retrieval, cognitive psychology, and case-based reasoning communities. Chapter 3 then
turns to the task of legal analysis in the domain of tax accounting. A review of several AI
approaches to legal classification and justification is given, followed by a conceptual review
of tax accounting and issues involving captive insurance corporations.

A solution to the conceptual retrieval problem in this context is described in

Chapters 4 through 7. Chapter 4 presents a generic-task problem solving architecture for

12

the task of legal analysis. Chapter 5 details a technique for analyzing and representing
justifications using a Functional Representation scheme based on the ideas of Toulmin;
Using this technique as a foundation, Chapter 6 develops a functional model of conceptual
retrieval, from the selection of index vocabulary to the organization and search of case
memory. Finally, this model is integrated with the problem-solving architecture, and
Chapter 7 describes CRISTAI, a computer program that embodies the model of conceptual
retrieval described here. Chapter 7 also presents examples of CRISTA’s problem solving,
focusing on its use of case memory.

Chapters 8 and 9 offer a theoretical discussion of this work. Chapter 8 evaluates
the work presented here with respect to other approaches and discusses how this work
might be united with work on some other active research issues in Al. Chapter 9 then
concludes the dissertation with a consideration of the work’s ultimate contributions and

significant avenues for future research.

 

1 For “Conceptual Retrieval In the Service of Tax Argumeutation."

CHAPTER 2

CONCEPTUAL RETRIEVAL: AN ANALYSIS

2.1 Introduction

In domains with large bodies of technical documents, computer-aided research has
become a standard practice. The law Offers an extreme example of this phenomenon. The
jurisprudential tenet stare decisis mandates that prior decisions be given precedence in
deciding new cases. This rule imparts great significance to the task Of legal research,
leading to the accumulation Of vast numbers Of case documents to be used in analyzing
subsequent cases. In this context, computer-aided research poses a serious challenge: how
can one locate in an immense data base just those records or documents relevant to a
particular topic of interest?

The idea of systems that index and retrieve information using semantic knowledge
of a domain has been termed conceptual retrieval (CR). This chapter discusses the CR
problem and research aimed at solving it. The remainder Of the chapter consists Of five
main parts. Section 2 describes the traditional approach to information retrieval, key-word
search, and its principle shortcomings. Sections 3 through 5 consider three prominent lines
of interdisciplinary work that Offer potential solutions for CR — information retrieval,
cognitive psychology, and case-based reasoning. Drawing on results from these research
areas, Section 6 outlines the contributions that each approach offers toward the realization
of conceptual retrieval systems. Finally, in the conclusion, the common themes that guide

the work described in the rest Of this dissertation are presented.

13

14

2.2 The Traditional Solution: Key-Word Search

Conventional information retrieval (IR) systems implement the “key word in
combination” (KWIC) method introduced by Horty [1962] and his colleagues.1 In the
KWIC approach, the retrieval system creates an index on the full text Of a document that
lists the exact location of each significant word in the document.2 Queries to the system
consist Of one or more text words connected using simple Boolean and adjacency
Operators, and the system retrieves all documents in which the specified words appear in
the combination specified by the query. Thus, this approach defines its index vocabulary,
index organization, and retrieval algorithm in purely lexical terms — the index vocabulary
equals the set Of significant words appearing in the document, the organization Of these
terms consists in an inverted file on these words, and documents are retrieved when they
contain the words specified in the query.

Despite its simplicity, the KWIC method produced an important breakthrough in the
field of information retrieval. KWIC made possible the construction of large IR systems
consisting Of thousands Of pages of text. In addition, it spawned a wealth Of research
activity focused on the efficient generation and use of index files from text files. A large
majority Of the information retrieval research community still devotes its primary efforts to
improving the utility of KWIC approaches.

However, systems based on this methodology suffer from one crucial problem:
there is no necessary connection between a word and the meaning of a text in which the
word appears or does not appear. This truth follows from the duality of expressiveness

and ambiguity in natural language [Krovetz 1985]. Generally, one can express a concept in

 

1 See also the work of Kehl, Horty, Bacon, and Mitchell [1961]. For a general discussion of early retrieval
systems, see Buchanan and Headrick’s seminal paper on Al applications in the law [1970. pages 41-46] .

2 Only words such as articles, auxiliary verbs, and prepositions are considered “insignificant.”

15

many different ways, and each word may have multiple meanings. These semantic
ambiguities can be resolved only in the context of a particular usage.

The result of this problem for KWIC is a nearly inevitable trade-off between
precision (the relevance of retrieved documents) and recall (the retrieval of relevant
documents). A query specific enough to insure that most documents retrieved will be
relevant often leaves many relevant items unretrieved, but a query general enough to insure
high recall will usually also lead to the retrieval of many irrelevant documents. Even in a
field characterized by technical discourse, such as the law, this problem persists. Blair and
Maron [1985] found that attorneys using key-word retrieval systems perceived that recall
from their queries exceeded 75% when only 20% of the relevant items were actually
retrieved.

This shortcoming of the KWIC approach gives rise to a number of practical
problems in its use. Hafner [1981] described the infeasibility of specifying all possible
word combinations that could express a legal issue or relation, thus indicating the difficulty
of achieving nearly-complete recall in practice. On the other end of the spectrum, Bing
[1978] found that lawyers tend to make critical errors in specifying complex key-word
queries. These errors typically involve the incorrect combination of AND, OR, and NOT
operators. The results of Bing’s study dispel hopes for a high degree of precision.

Other practical problems arise, too. If a new legal decision or statute coins a phrase
in discussing a particular issue, relevant cases decided thereafter will likely contain the
phrase, but relevant earlier cases will not. Or one may wish to retrieve documents that only
partially match a given fact pattern, for instance, in trying to find a larger body of
potentially relevant precedent for a new concept. Generating a useful key-word query for
this situation will be difficult, if not impossible. Key words either appear in a document,

or they do not; there is no room for gradation.

l6

Ultimately, these difficulties can be traced to the fact that key-word systems force
users to focus on syntactic phenomena — the lexical details of documents — rather than on '
the meaning of the documents they seek. The user must not only understand the domain
and its semantics but also know how to translate such understanding into lexical queries of
the data base. By placing this burden on the user, the key-word system serves only as a
mechanized index to a body of documents, providing little additional assistance for the task
of computer-aided research.

The rapid growth in the number of documents to be searched, the importance of the
research task in many professions, and the documented shortcomings of KWIC systems,
however, all indicate a clear need for systems that go beyond text search to information
retrieval based on the semantics of the domain and its documents. Such an “intelligent”
research assistant would mediate between the conceptual vocabulary of the user and the
lexical vocabulary of stored items. In doing so, the system would subsume much of the

implementation knowledge involved in generating queries.

2.2.1 Conceptual Approaches to the Retrieval Problem
Researchers from a variety of disciplines have investigated issues relevant to
solving the CR problem. Three prominent lines of this interdisciplinary research are:
0 information retrieval, which has applied its own lessons and some
of those from Al to the particular problem of legal information
retrieval;
0 cognitive psychology, which through its investigation of analogical
reasoning has identified important aspects of how humans retrieve

and use past experiences in solving new problems; and

17

' case-based reasoning. which has developed a large body of results
on how computer systems can index, store, retrieve, and use past
cases in problem solving.
Each of these research tracks addresses the conceptual retrieval problem in its own context,
and consideration of their different perspectives offers a more complete picture of the
problem. By combining their individual contributions, one can arrive at an informed

strategy for constructing CR systems.

2.3 Information Retrieval

A number of researchers in the information retrieval community have pursued the
notion of conceptual retrieval systems. Focusing on particular domain needs, these lines of
study seek either to extend the traditional model of text-based retrieval with knowledge-
based constructs or to develop new techniques for addressing n'aditional problems. Legal
information retrieval has provided an especially fertile area of study because of its great
practical importance and its archetypal nature. Case law encompasses the full range of
natural language problems inherent in full-text retrieval while also resting on a well-
developed and continually evolving body of conceptual knowledge. In addition,
researchers in Al have long been interested in legal problem solving, thus providing
background for representing and manipulating legal knowledge.

Researchers interested in extending the traditional IR model have typically
investigated the utility of front-end tools that assist the user and then create key-word
queries for a full-text system. Bing [1987] has proposed a “norm-based thesaurus”
approach in which the user traverses a hierarchy of norms (legal rules) and identifies
concepts relevant to the situation at hand. From the selected nodes, the system would
automatically generate a key-word query that incorporates different senses of each concept.

For example, the concept spouse can be elaborated to include expressions such as husband,

18

wife, and married. While acknowledging that such a system would still be limited to
providing synonym support, Bing suggests that the use of formalized legal rules will at
least allow automatic generation of some complex queries.

Another system of this sort, RUBRIC, was first developed as a general intelligent
retrieval methodology [Tong and Shapiro 1985] and was then applied to the problem of
legal document retrieval (Tong, Reid, Crowe, and Douglas 1987]. Using a hierarchy of
production rules to represent conceptual relationships such as implication and synonymy,
RUBRIC allows for “fuzzy” definition of concepts in terms of domain text. An example of

a RUBRlC production rule appears in Table l.

 

Table 1. A RUBRIC Production Rule (Adapted from Tong et al. [1987])

( EVIDENCE financing-condition
( (OR (PHRASE “obtain” “sufficient” “funds”)
(PHRASE “sufficient” “financing” “obtained”)
(PHRASE “arrange” “sufficient” “funds”)
(PHRASE “availability” “sufficient” “funds”)
)

0.8 ) )

 

This rule states that the concept of a financing condition is considered very relevant
(0.8 on a 0-1 scale) in the presence of any of the standard legal phrases supplied. Given a
query in natural-language text, RUBRIC can use such rules to identify concepts relevant to
the query. The system then retrieves documents from memory that have been tagged by the
requested concepts.3 RUBRIC’s builders argue that the user must still work hard to

generate a good query, but at least retrieval is based on concepts rather than text.

 

3 The numeric “relevancy factors” are manipulated in much the same way as MYClN’s certainty factors.
They allow RUBRIC to combine concepts and text expressions based on Boolean operators.

19

2.3.1 An AI-Based Model of Information Retrieval

Hafner [1981] conducted the earliest effort to develop techniques grounded in Al
specifically to address the traditional problems of information retrieval. Hafner’s “Legal
Information Retrieval System” (LIRS) employed a model of legal classification knowledge,
in the form of a restricted semantic network, defined on the area of negotiable instruments
law. This domain model consisted of a network of approximately three hundred objects —
agents, events, and various negotiable instruments, their components, properties, and
roles. Connecting these nodes were links denoting membership, constituency, property,
and event-condition relationships.

From this model, Hafner developed a situation description language (SDL) for
describing documents in the data base and for expressing data base queries. The SDL
allowed simple binary relationships among objects and properties and included primitives
for expressing negation of relationships, a particularly well-developed result of this project.
The LIRS data base contained entries for approximately 200 statutes and 200 cases, as well
as entries for official comments on the statute by legal authorities. Each item in the data
base was represented by a set of SDL descriptors and full legal citations. An entry’s
descriptors denoted information regarding legal setting, concept definition, and legal
issues. In order to facilitate retrieval, each concept in the domain model held a set of
pointers to data base items related to the concept in a particular way (for example, by
presence or absence).

LIRS took a query, determined the indices corresponding to the concepts in the
query, composed an appropriate descriptor based on the index types and the referred
concepts, and compared this composed descriptor to the those, of items indicated by the
concepts index list. Entries matching the query descriptor were then returned by the data
base. Complex queries — those including conjunction or disjunction operators — were

decomposed into primitive queries, and the results of processing the primitive queries were

20

then re-combined according to the complex operators. In addition, the data base would
match queries to documents at varying levels of semantic complexity, as specified by the
user. These match types included an exact match as well as direct and inferred extensions
of the query.

While the LIRS project addressed important questions regarding legal knowledge
representation and conceptual retrieval, the project did not directly tackle the issue of how
to organize the data base for efficient retrieval. Hafner [1987] subsequently refined and
extended the LIRS approach to incorporate an issue/case discrimination tree for indexing a
memory of legal cases. In this extended approach, individual case descriptors consist of
constellations of concepts, drawn from a model of domain knowledge and assembled
according to the legal roles they play in the case. Each concept in the domain has an
associated set of decision rules that specifies the conditions under which the concept is
present These rules also constitute the primitives for representing the justifying theory of
the case, one of the critical components of a case descriptor.4

Cases in memory are indexed by an issue/case discrimination tree (Figure l), which
depicts normative relationships among legal issues and relevant domain facts. Issue nodes
denote legal conclusions to be drawn in deciding a case, and factor nodes denote types of
facts that may influence the process of drawing a conclusion. Influence links connect
issues to sub-issues and to factors relevant in deciding the issue. Each link reflects a

positive or negative relationship between the connected concepts.

 

4 The theory of a case decision demonstrates how the legal issues of the case relate to the case‘s
background facts (which define the situation being considered) and the holdings of the court (the decision itself).
One can view the theory of a case as a simple justification of the decision. with little elaboration of the inference.

21

 

 

 

Broker is liable
for buyer damages

‘

 

Broker breached
duty to give
correct information

 

 

Broker had
a duty to
disclose

 

 

 

 

 

 

 

Condition of

property was
misrepresented

 

 

 

     
   
 

Broker had
actual knowledge

 

 

 

Reliance on
broker was justified

 

 

 

 

 
   
     

Buyer had an
opportunity
to discover

 

0 Factor Node

Issue Node

 

 

Influence Link

 

Figure 1. A Sample Issue/Case Discrimination Tree

 

 

22

 

The issue was decided

Positively

Negatively

 

Reliance was justified,
and the case was

Reliance was not
justified, yet the case

 

 

Plaintiff decided for was decided for
Case was the plaintiff. the plaintiff.
decided
for
Reliance was justified, Reliance was not
Defendant yet the 0886 justified, and the case

was decided
for the defendant.

was decided
for the defendant.

 

 

 

 

Figure 2. The Four Kinds of Issue Pointer into Case Memory

 

 

From this structure, two kinds of index into the case memory can be established.
First, each node in the tree has a compound link to the case law collection (Figure 2). This
link points to the set of all cases in which the issue or factor played a role. For example,
the node for justified reliance points to all cases in which the buyer claims to have been
justified in relying on the broker’s presentation of facts. This set of cases is segregated into
four groups, based on how the issue was decided and on how the case was ultimately
decided. This kind of link relates how an issue was decided to the set of cases in which the
issue was important. ‘

Second, each link in the tree also has a compound pointer to the cases in which the
influence relationship was relevant to disposition of the case. This set of cases is also
divided into four groups, along two dimensions: (1) the presence or absence of the factor

or sub-issue in the case, and (2) the effect its presence or absence had on the resolution of

23

the issue. Figure 3 depicts the division of the case set for a particular influence link, the
link between the issue duty to disclose and the factor broker had actual knowledge. This
type of organization allows the system to provide meaningful answers to very specific
queries, such as “In what cases did the broker have actual knowledge of a defect and yet
the court decide that there was no duty to disclose?” Given this sort of link, the system can
answer queries referring not only to relevant concepts but also to the relevance of the

relationships among concepts in a case.

 

 

With respect to the issue, the factor or (sub—issue) was

 

 

Relevant Irrelevant

Broker had actual Broker had actual
Present knowledge, and knowledge, yet this

this resulted in a ' did not result in a

The factor duty to disclose. duty to disclose.

(or subissue)
was

Broker did not have Broker did not have

Absent actual knowledge actual knowledge

and so had no and yet had a
duty to disclose. duty to disclose.

 

 

 

 

Figure 3. The Four Kinds of Link Pointer into Case Memory

 

2.3.2 AI-Based Approaches and Their Significance
Hafner’s model demonstrates the power of pattern matching (through abstraction to
domain concepts) for conceptual information retrieval. This idea is an old one, dating at

least back to Raphael [1968] and his SIR program, and finds application in a variety of

 

24

forms. Zarri [1985] has proposed a methodology for intelligent information retrieval in
which data in the memory are represented as frames composed of sentences from a case
grammar that defines the domain. Queries are then decomposed along the dimensions of
the cases in the grammar, and each access to the memory is guided by the semantic content
of the classes of patterns associated with the case. From an entirely different perspective.
Rose and Belew [1989, also Belew 1987] propose a connectionist approach to conceptual
retrieval. In this approach, the retrieval system begins with a network constructed from an
initial representation of the documents it contains, linking nodes that are instances of the
important classes of objects in the domain. As users interact with the system, it
automatically refines the weights on the network’s connections to account for the utility of
retrieved documents and to reflect the addition of new documents.

AdOpting a view of conceptual retrieval that rests foremost on principles of Al
garners advantages unavailable to the approaches that rely primarily on traditional IR
techniques. Tools such as those proposed by Bing [1987] and Tong [with Reid, Crowe,
and Douglas 1987] suffice because of the nature of the domains in which they were used.
RUBRIC succeeds because it is applied to a case law domain in which concepts are
characterized by complex permutations of a relatively small number of text patterns. In a
domain where concepts might be expressed with a broader range of text patterns,
RUBRIC’s production rules would quickly become too large and too numerous for the
approach to be feasible. Bing’s norm-based thesaurus succeeds because it is applied within
the European model of law, with its strict adherence to statute and not case law. One may
be able to express all the synonyms for a concept that might appear in statute, but the text of
case decisions is typically too diverse for a purely synonym-based approach to offer much
advantage.

These tools remain too faithful to traditional IR to avoid its principal problems. A

text- or synonym-based approach makes generating queries involving relationships among

25

concepts nearly impossible, given the combinatorial explosion of different phrases that
might be used to express the concepts and the relatiOnship. For this reason, Hafner [1987]
argues that the greatest promise for CR in the law lies in (1) understanding how legal
experts remember and classify cases, and (2) converting this domain knowledge into

appropriate data structures and algorithms.

2.4 Cognitive Psychology

Much work in cognitive psychology deals directly or indirectly with the retrieval of
information from semantic memory. Studies of concept formation and recognition involve
determining whether, and under what conditions, a given item in memory will be evoked in
the presence of some input. From the time of Aristotle, concepts were defined by sets of
necessary and sufficient conditions that characterized instances of the concepts. Classic
psychological studies, such as those reported by Rosch [1975], describe a shift from this
Aristotelian notion to a view in which concepts are characterized by prototypical instances.
Related instances are linked to the prototypes according to the degree of similarity, with
more similar instances “closer” than less similar instances.

In order to explicate this sort of theory, cognitive psychologists have been forced to
consider the mechanism that enables one concept to evoke another. Similarly, research on I
how humans understand and resolve experience (for example, using scripts of typical
events) and how they behave in the presence of incomplete knowledge has offered insights
into how memory is organized and accessed. An especially interesting area of psychology
influenced in this way has been the study of analogical reasoning.

Solving problems by analogy — recognizing that an already-solved problem is
similar to a current one, and adapting the old solution to the new situation — plays an
important role in many domains. One can apply old solutions in a variety of ways:

0 by transferring the solution directly,

26

0 by modifying the solution to account for differences between the
problems,
' by transferring the method used to derive the solution and applying

it to the new situation, or
0 by creating an abstraction of the problem and solution and applying
this generalization to the new situation.

According to Greeno [1978], analogies can be characterized more generally as problems of
inducing structure: take the givens of an Old situation, determine the structural relationship
among its elements, and apply this relationship to the new situation. Analogy, then,
requires that an appropriate analog be available before its structure can be induced and
applied in a new situation. Under contrived conditions, such as in standardized testing, the
analog is typically provided to the problem solver. Studies have shown that people are
often quite adept at solving problems by analogy when given an appropriate analog from
which to induce a structural relationship [Gentner 1989, Stemberg 1977]. More generally,
though, a problem solver must remember or otherwise actively retrieve an analog for use.

It is in remembering useful analogs that people have difficulty. Gick and Holyoak
[1980] sought to determine how to overcome this difficulty. They found that hints
consisting of key phrases, such as “divide and conquer,” promote the retrieval of relevant
past experiences for use in a new situation. Along similar lines, Gentner and Toupin
[1986] conducted studies in an attempt to discover the source of difficulty in recognizing
analogs (in this case, stories). They found that subjects had difficulty recognizing the
utility of a past story when it was not presented — or not understood — in a systematic
way. This condition was most readily observable when the causal relationships among
elements of the story were not clear. Thus, one of the central goals for understanding
analogical reasoning involves determining what types of systematicity characterize the

organization of human memory and best facilitate the retrieval of relevant analogs.

27

2.4.1 Memory Organization and Access

Two schools of thought exist among cognitive psychologists regarding memory
organization. One group holds that memory is indexed only by surface features and that all
retrieval is driven by sensory input. On this view, any use of higher-level concepts occurs
after retrieval of analogs sharing similar surface features. The second camp proposes that
concepts do play a role as indices into memory. Members of this camp argue that the use
of such “abstract” features is directly tied to the kind of higher-level cognitive processes
which have traditionally’been hard to isolate in experimentation. The following sections

discuss these two views of memory in greater detail.

Retrieval based on Surface Features A large body of evidence supports the view that,
without outside facilitation of some sort, people generally access analogs based only on
relatively simple surface feature similarities.5 As a result, some researchers have proposed
highly parallel models of memory that emphasize syntactic and low-level semantic
similarity. Thagard and Holyoak [1989] offer a theory of retrieval based on the
simultaneous satisfaction of three types of constraints on memory: semantic similarity,
structural consistency, and pragmatic centrality. These constraints are not necessary
requirements on retrieval but rather pressures that operate on the memory to different
degrees. Waltz [1989] argues that massive parallelism over very low-level syntactic
features best accounts for human behavior. From this viewpoint, concept-level indices are
outputs of, not inputs to, the retrieval process. The point of such architectural models of
memory access is, Waltz says, that people “can retrieve deep, evaluative features rapidly,
given only readily extractable surface features.” [page 41]

In order to account both for ample experimental evidence of retrieval based on

simple features and for experimental and anecdotal indications that higher-level conceptual

 

5 Gentner [1989] describes a variety of such studies, in addition to those cited above.

28

features affect retrieval, some researchers have proposed theories that operate on two
levels. Forbus and Gentner [1991] offer such a theory in their MAC/FAC6 cognitive
architecture. This approach was motivated by the insufficiencies of competing types of
models for memory organization. Al models of memory have tended to be based on clever
indexing methods that seem unlikely to scale to the demands placed on a large, evolving
memory. Conversely, models of memory from cognitive psychology have typically
employed simple representation schemes, such as feature vectors, that allow tractable large-
scale search but which do not reflect the richness of the semantic features that people
possess and use.

MAC/PAC blends the strengths of these types of model in a two-stage computation.
First, the MAC stage uses a flat feature representation to retrieve a set of potential analogs
at little computational cost. This kind of blanket search is often very imprecise, resulting in
the retrieval of many inaccurate items.7 Second, the PAC stage attempts to match the input
“query” with these items based on their structure and higher-level semantic content.
Forbus and Gentner assert that this model strikes a compromise between “seemingly
incompatible intuitions about memory: Access tends to be governed by surface properties,
while inference tends to be governed by relational matches.” [page 68] These and other
researchers, including Gentner [1983] and Burstein [1989], have long been investigating
the nature of the structural and relational mapping necessary for the second stage of such a

bicameral process.

Retrieval based on Abstract Features Despite these claims that readily extractable low-

level features serve as the keys to human memory retrieval, other research has shown that

 

6 For “Many are called. but few are chosen.”

7 By erring on the side of imprecision. this type of search tends to have very high recall. However. as
Martin [1989] points out, even such a broad retrieval strategy cannot be guaranmd to retrieve all relevant items.

29

retrieval is also strongly influenced by pragmatic features, such as the processing task at the
time of storage. Seifert [1988] has conducted a long line of experiments whose results
support this view. These experiments show that the retrieval of stories from memory is
facilitated by attending to the processing goals present when the story was initially
encountered. Motivated in part by Schank’s view that “remindings”8 reveal the way in
which an event was understood, Seifert has investigated the conditions under which such
remindings occur.

Her experiments indicate that encoding a new story does not always activate
thematically similar stories. Rather, similar stories are activated only when the encoding
context presents a functional or strategic purpose for doing so. Cognitive goals thus form a
critical part of the context in which encoding and retrieval operate. Seifert’s studies indicate
that such goals are particularly active in problem solving tasks such as planning and
argumentation [page 365]. In some sense, these results countermand Schank’s emphasis
on automatic remindings, especially if one defines the notion of intentional reminding
broadly, to include not only the consciously intended but also retrievals that occur due to
cognitive biases in processing.

Seifert [1988] and Gentner [1989] argue that processing goals play a role in
retrieval by affecting the features that are most attended to in the process of determining the
similarity between items in memory. On this view, the mind attends to particular features
of a new story based on the cognitive goals active in working memory at the time of
encoding. Stories encoded under similar conditions are likely to be activated in the
process. Interestingly, many past psychological experiments may have obscured this

phenomenon by setting up simplistic contexts in which only one processing goal is

 

8 Schank [1982] coined the term “reminding" to refer to the spontaneous remembering of past experiences
based on a current situation. He also uses the term to denote the result of such remembering, a past episode.

30

present.9 This kind of experiment allows the goal to be implicitly encoded as part of the
story’s representation in memory. (On the other hand, AI studies of problem solving
generally involve tasks with many different processing goals, thus making memory access
based on goals a natural consequence of system building.)

That psychological models of memory organization should account for high-level
features as indices remains a controversial yet active empirical issue. Researchers who
adopt this view believe that it can be reconciled with simple feature-based approaches by
developing theories that explain how and when high-level features participate in
remembering. Such theories typically focus on how strategic purpose affects what is
considered “relevant” in a given situation. Burstein [1989] investigates how queries of
different forms affect the mapping techniques involved in arriving at a suitable analogy.
(For example, he has compared queries that involve determining the functional organization
of a device given its behavior to queries that involve finding a component capable of
performing a particular behavior given a functional specification.) Viewing analogy as a
matter of inducing structure ultimately leads to the idea that such structures, and the features
that may play a role in mapping structural similarities, can serve as useful components in a

theory that explains the phenomena of memory retrieval.

2.5 Case-Based Reasoning

Case-based reasoning (CBR) refers to a general methodology in which previously-
encountered situations are used as aids in interpreting or solving new problem cases. The
motivation for this methodology is to avoid the computational cost of re-solving problems
from scratch each time they arise by caching solutions to each problem confronted; if a new

situation closely resembles a past situation, the problem solving from the past situation may

 

9 Seifert points to the work of Gick and Holyoak [1983], who found in later studies that controlling for

problem-solving context had a material effect on the performance of memory retrieval. See Seifert [1988] for a
detailed discussion of this point.

31

provide a “short cut” in addressing the new situation. Fundamentally, CBR views problem
solving as remembering, as a memory process rather than a deliberative one: “Finding the ‘
right [story of past experience] is what we mean by understanding.” [Schank 1982]
Remembering a similar prior case can not only save recomputation of a right answer but
can also help avoid past failures by recognizing the conditions that led to a failure in an
earlier case.

Though consistent with much that cognitive psychologists have observed regarding
human problem solving in natural settings, CBR research typically places less emphasis on
exact replication of human memory behavior. Rather, CBR emphasizes discovery of the
principles that underlie reasoning with cases and building useful computer programs to
perform tasks in real-world domains. However, in two general senses, CBR reflects the
nature of how people solve problems. First, case-based systems can typically find a good
but not optimal solution to a problem rather quickly, with less risk of a wildly wrong
answer than systems that do not use past cases.10 Tasks that lend themselves to a case-
based approach, then, are typically characterized by multiple solutions in which trade-offs
and the ability to compare solutions are more important than optimization and correctness.
Second, each new problem solved increases the knowledge of the system, either by noting
the success of a solution or by recognizing a failure in the presence of particular domain

conditions.

2.5.1 Issues in Case-Based Reasoning

The process of case-based reasoning can be portrayed as consisting of the
following steps [Riesbeck and Schank 1989]:

l. Characterize the current situation.

2. Retrieve previous cases similar to the situation.

 

10 The truth of this claim critically depends on the ability of a case-based system to recognize similar past
cases effectively. Producing this ability is one of the central research aims of CBR.

32

3. Determine the best match to the situation among the retrieved cases.

4. Adapt the selected case(s) into a solution for the current situation.

5. Apply the solution. If it is satisfactory, assign indices to the new case

based on the goals it achieves; if it fails, use the reasons for failure
to identify indices of the situation predictive of the failure.

6. Store the situation and solution as a case in memory.

Case retrieval and storage constitute the memory processes of CBR, the reactive interface
between case memory and the system. These steps rely specifically on the generation of
indices provided by situation characterization and solution application. At the other extreme
is case adaptation, a deliberative process that employs task- and domain-specific knowledge
to modify a past solution for use in a new situation. The character of the other tluee steps
in the methodology remains a part of active research. Some view these steps as extensions
of the memory (for example, Waltz [1989] and Ashley [1990]), while others recognize that
substantive problem solving may be required to perform these tasks (most notably,
Hammond [1986], Koton [1988], and Kolodner [1991]). This dispute, as it relates to
indexing, is discussed further below. .

Researchers have generally applied this methodology in a number of ways. Some
use the process as a “weak method,”1l seeking to discover knowledge structures
appropriate for applying its steps to various problem-solving tasks [Hammond 1986,
Kolodner. 1991, Birnbaum et al. 1991]. This style emphasizes case adaptation, as prior
cases are used to derive and apply a solution to a new problem instance. In another branch
of work, often called precedent-based CBR, the method uses cases to classifyior evaluate a
new situation based on the treatment of past similar situations. This approach, common

especially in legal and diagnostic domains, de-emphasizes adaptation and stresses the

 

l 1 A weak method is a search technique that employs no domain-specific knowledge to guide the search.

33

ability to characterize and select cases based on similarity [Ashley 1990, Branting 1991,
Bareiss 1989].

Much of the focus of CBR research has been placed on issues of adaptation. Since
differences between the past and current cases may have significant effects on how the
problem should be solved, adapting a case requires complex task- and domain-specific
knowledge to ensure that the new solution will adequately account for the differences.
Generally, this knowledge has been represented in the form of causal rules that define how
to achieve particular goals and how changes to the solution will affect other parts of the
solution [Hammond 1986]. One notable approach at developing a theory of how case
adaptation can be performed involves the use of model-based reasoning. Koton [1988,
1990] and Goel and Chandrasekaran [1989] have described approaches in which device
models, which capture causal understanding of how a device’s structure contributes to its
behavior, provide a principled means for adapting a solution to a new situation. This
approach also assists in the evaluation of case differences by offering recourse to the model
in deciding the effects that particular features have on the functioning of the device.

The other primary focus of case-based reasoning research is indexing. Whatever
can be made of the differences between problem-solving CBR and precedent-based CBR,
or of the differences among CBR approaches to various problem-solving tasks, they all
share an essential dependence on case retrieval and storage [Kolodner 1991]. One of the
goals of CBR is to avoid the need to do deliberative problem solving by remembering and
employing past solutions. For this reason, the design of a case-based system should
minimize the extent to which the system must rely on “compiled knowledge” to adapt and
repair its solutions. The way to minimize deliberative problem solving in such matters is to
retrieve the best possible past case — where “best” means the one that is most uSeful for
dealing with the new situation. As the reasoner experiences more and more situations, it

becomes increasingly better equipped to use past solutions with little or no modification.

34

How the memory of cases is organized and what features the reasoner can use to access the
memory are critical issues both in building useful systems and in understanding the

phenomenon of reasoning with cases.

2.5.2 Case Memory and the Indexing Problem

Most researchers agree on the basic issues underlying the indexing problem
[Kolodner 1991]. Indices should be predictive of a case’s utility in achieving an outcome.
They should be abstract enough to indicate useful cases that are related to but different from
one another, and yet concrete enough to be recognizable without extensive and
counterproductive computation. Each case may have several indices associated with it,
depending on the various roles the case might play in future problem solving (for example,
combinations of factors responsible for or descriptive of its outcome, or combinations of
factors that describe intermediate steps of note in the problem-solving process).

In addition, many researchers (in particular, Hammond [1989] and Birnbaum
[1989]) have noted that an important trade-off exists between the cost and utility of
different forms of index. That is, indices that can be derived from standard input at a low
cost will tend to describe only a very narrow set of cases, while indices that maximize
descriptiveness are likely to be computationally expensive to generate. Given this trade-
off, an indexing vocabulary must strike a practical and cost-beneficial compromise.

With respect to case retrieval, the issue of indexing introduces a set of related issues
[Domeshek 1991a]:

' determining index content and structure,

0 matching indices,

0 organizing and searching case memory, and

0 generating probes to query the memory.

35

Any proposed indexing methodology must ultimately account for how to do each of these
tasks. Not all of these issues have been formally examined in the literature, but the issue of
index content has been the subject of considerable theoretical discussion. Such research
has focused on this question: should abstract, high-level indices be used and, if so, what
sorts of abstractions will offer the greatest power and utility for retrieval? Some
researchers advocate the use of only simple, low-level features taken directly from system
input. However, tracing back to CBR’s roots in the work of Schank, the dominant view of
indices has stressed the use of functional indices, ones defined somehow in terms of the
system’s problem-solving goals. The next two sections discuss these two primary

approaches to the issue of index content.

The Use of Low-Level Features A small but active group of researchers continues to
investigate the utility of surface features as indices for case memory. Waltz [1989] has
argued that “indexing” is the wrong way to think about memory retrieval. His claim has
two major points. One, indices are not needed for tractability, since memory retrieval is
done in parallel. Two, most memory processes use only surface features and basic
experiential context for retrieval. Thus, complex indices may be useful for describing
problem solving, but they are not necessary to a theory of memory retrieval. He cites the
traditional results from cognitive psychology mentioned above as evidence for his theory of
“memory-based” reasoning.l2

Using the Connection Machine, Waltz and his colleagues at Thinking Machines
Corporation have demonstrated the utility of this approach for tasks such as learning to read
text [Stanfill 1988] and predicting protein structure [Zhang and Waltz 1989]. Others have
employed the approach of retrieval based on surface features, albeit on serial machines, for

similar sorts of tasks. Especially interesting work has been done by Lehnert [1987] and

 

12 However. as noted above, these results may have been flawed in that they failed to control for problem-
solving context [Seifert 1988].

36

Golding and Rosenbloom [1991] on word and surname pronunciation, respectively. This
use of surface features not only eliminates the need to compute indices for queries and ‘
storage but also enables the application of a case-based approach to large, existing data
bases.

But the approach also lacks a degree of complexity often required in performing and
explaining problem solving. Ashley [1990] offers a partial compromise in the form of
dimensions, relatively low-level factors that affect a legal claim. Dimensions are not
surface features of a situation but rather are simple fact abstractions. However, the value of
each dimension can be computed from system input data using a single production rule.
Each case in memory is then indexed by the dimensions that characterize it Thus,
dimensions preserve the spirit of retrieval based on surface features, while also enabling the

use of simple domain abstractions in organizing cases meaningfully.l3

The Use of Abstract Indices The foundation of case-based reasoning research has been
organization of memory according to the functional features of stored cases [Schank 1982].
These functional features denote the utility of a case in future situations. Schank discussed
several types of functional feature, including expectation failures, processing goals, and
abstractions (such as the morals of stories). Hammond [1986] describes a theory of
memory closely tied to the processing goals of a planner. In CHEF, Hammond’s system,
past planning episodes are indexed by the goals they satisfied and by explanations of
failures. Failure explanations consist of features of the situation causally responsible for
the plan’s inadequacy. Such features may be direct input or abstractions computed in the

course of the planning process.

 

13’ Ashley asserts that the nature of legal domains precludes the use of abstractions at a higher level than
dimensions, due to the open-ended evolutionary character of legal concepts. Chapter 3 discusses Ashley’s model
of legal argument and the role of dimensions play in greater detail .

37

Hammond [1989] argues that the computational cost of using such complex
features is, in fact, nil, since they must be generated in the course of problem solving
anyway. Going further, Birnbaum [1989] contends that, even if a complex index must be
computed outside the course of problem solving, the cost of this computation must be
offset against the computational savings realized in using the index. If use of the index
results in better case selection, and hence in less computation being required to adapt
retrieved cases, then derivation of the index may prove to be cost-beneficial. This situation
can occur only when there exists an abstract index vocabulary that is likely to result in the
retrieval of more appropriate cases. Research aimed at identifying efficacious task-specific
vocabularies (such as that described by Hammond [1989] and Birnbaum et al. [1991] for
the task of planning) becomes crucial when considered from this perspective.

An even stronger claim is made by Martin [1989], who studied natural language
analogies. Martin considered stories that share a common theme and found that, while the
relational structures of the stories could be mapped effectively [Gentner 1983], the stories
shared no common surface features. These examples demonstrate that “indexing on the
basis of simple features in the hope that a subsequent structural match will have less
work to do cannot handle examples that rely on purely structural constraints.” [page 295,
emphasis added] Thus, on this view, abstract index vocabularies are necessary to the
recognition of some relevantly similar cases.

In general, one can view the kind of abstract index vocabulary typically used in
CBR as being based on the goals satisfied by the generated solution. For example, in
CHEF, indices are computed from the functioning of the plan, either from gOals that the
plan can achieve or from features predictive of goal failures. More recently, researchers
have moved a step further, with the use of explicit models of the solution serving as a
source of the index vocabulary. Goel [1989, also with Chandrasekaran 1989] has

described an approach in which designs are described by the functions they realize. These

38
functions then serve as indices for design cases stored in memory. Use of such an abstract
index not only facilitates efficient retrieval but also expedites identifying the part of a
retrieved case that will be most useful in the current situation, since the functions can also
index the behaviors that allow the device to perform each of its functions. Following a
similar strategy, Sycara and Navinchandra [1991] use behavioral abstractions to organize

memory according to the utility of designed components.

2.6 Major Themes in Memory Organization

Drawing on research in information retrieval, cognitive psychology, and case-based
reasoning, one recognizes that each field is experimenting with the same continuum of
approaches to memory organization. This continuum ranges along the dimension of index
complexity. One end of the spectrum consists of approaches that advocate the use of little
or no indexing at all, relying instead on a content-addressable memory to match purely
syntactic features over massively parallel hardware. At the other end of the spectrum,
complex indexing schemes are based on problem-solving goals and relations. Given this
range of approaches, one can also recognize common themes that pervade all the disciplines
and indicate promising directions for further work toward the goal of conceptual retrieval
systems:

- the need to provide case retrieval support to human reasoners,

0 the need for both simple and abstract features as indices, and

e the utility of problem-solving goals as abstract indices.

2.6.1 Case Retrieval for Human Reasoners
Each of these disciplines is addressing the need to provide useful analogs to human
problem solvers. This research interest results from an understanding that people are often

competent at reasoning with precedents but are not always good at remembering

39

appropriate ones in the first place. Legal information retrieval research is driven by the fact
that the volume of legal cases far outstrips the capabilities of legal researchers to remain
abreast of case law development in all domains [Hafner 1987, Bing 1987, Zarri 1985].
Cognitive pSychologists have ample empirical evidence that this characteristic holds even in
the presence of far fewer episodes [Stemberg 1977, Greeno 1978, Gick and Holyoak
1980]. In the line of case-based reasoning research exemplified by Goel et al. [1991] and
described more generally by Kolodner [1991], the notion of augmenting human memory
by providing cases at appropriate times for human use in a broader problem-solving context
has become a primary goal.

Case-based reasoning offers a pragmatic suggestion in the attempt to solve the
conceptual retrieval problem: index cases first by their utility, and only then by their surface
similarity. Systems that retrieve the most useful cases will presumably be more useful both
as assistants to human problem solving and as components of larger computer systems

capable of performing difficult information-processing tasks.

2.6.2 Simple and Abstract Features as Indices

In the debate over low-level and high-level features as indices, each discipline
seems to reach the conclusion that both are necessary in any realistic model of memory.
The results of Seifert [1986, 1988], McDougal, Hammond, and Seifert [1991], and
Gentner [1983, also with Toupin 1986] demonstrate that human memory incorporates
elements of both surface and abstract features in the course of storing and retrieving
episodes. Indeed, Ashley’s [1990] use of dimensions could be considered an
implementation of Forbus and Gentner’s [1991] MAC/PAC cognitive architecture. Cases
are indexed by relatively low-level domain features and then retrieved en masse based on
direct feature matching. The retrieved cases are then sifted and selected according to the

implicit roles they can play in the assembly of arguments. Sirnoudis and Miller [1991] also

40

employ such a strategy but use a form of model-based validation to select the causally
relevant cases from among those retrieved. ‘

Hafner’s [1987] issue/case discrimination tree uses abstract factors to structure a
collection of legal cases according to important domain concepts and their relationships.
This structure provides access to cases at varying levels of conceptual and relational
abstraction. However, Hafner also notes that access to cases via purely syntactic surface
features — the text in the stored documents — will always be needed. As legal theory
evolves, researchers will require the ability to retrieve cases dealing with new concepts and
with new senses of existing concepts. These concepts, at least initially, will not be in the
issue/case discrimination tree. Thus, the user will need access to documents containing

particular text phrases that are known to characterize the new concepts.

2.6.3 Problem-Solving Goals as Abstract Indices

Finally, a principal theme running through studies of memory organization holds
that problem solving context, especially goals, in large part determines which features are
salient in any given situation [Seifert and Hammond 1989]. In cognitive psychology,
research offers only a vague idea of the nature and source of these goals, and consequently
the theme is still interpreted rather broadly [Gentner 1989, Seifert 1988]. Experience
derived in building case-based reasoning, however, has provided clearer indications of the
character of these goals. In particular, understanding the structure and behavior of the
intended solution seems to provide valuable information for retrieving useful cases later.
This has been the focus of several different lines of research, including Hammond’s [1986]
theory of planning, Domeshek’s [1991a, 1991b] theory of social advice, and the use of
model-based reasoning by Goel [1989] and Koton [1988, 1990]. Ultimately, these
approaches all support the general notion that problem-solving goals serve as an important

key to understanding and remembering cases.

41

Goel et al. [1991] present a thorough analysis of the sorts of indices that should be
used to organize a conceptual memory for the task of conceptual design in architecture. '
This kind of analysis could be applied to any task and domain. Indeed, their short
discussion of memory organization offers the seeds of a methodology for designing
conceptual retrieval tools in different domains and for different tasks. Goel et al. address
the question of how the subtasks that a conceptual designer performs — such as design
adaptation and evaluation — affect and mutually interact with case memory structures.
Unfortunately, they do not generalize their approach to other tasks or domains. A need for

such a generalization clearly exists in regard to conceptual retrieval tools.

2.7 Conclusion

Researchers in several disciplines have investigated issues relevant to the notion of
conceptual retrieval, and several common themes characterize their results. These themes
suggest that one of the vital open issues for further research is to identify appropriate index
vocabularies for describing cases in memory. However, Domeshek [1991a] reminds
researchers that such a theory must address not only index content but also the organization
and search of the index space. Developing abstract index vocabularies in isolation leaves
system builders to identify the principles of data base organization and search that typify a
vocabulary each time it is applied in a new domain. Ad hoc methodologies also offer little
justification for believing that constructed systems will be useful or robust.

Among the most important of these themes is the idea that problem-solving goals
can serve as especially useful indices for organizing a conceptual memory. Many
researchers have focused their efforts on the development of index vocabularies for
particular problem-solving tasks, but few have taken advantage of the contributions made
by knowledge-based systems research into the nature of problem solving. The rest of this

dissertation outlines a principled theory of indexing — vocabulary, organization, and

42

search — for the task of legal analysis. This research is done within the context of the
Generic Task Approach to knowledge-based systems, which offers a set of languages that
describe particularly useful problem-solving tasks and methods. By basing the theory of
indexing on such a task-specific approach, case memory can be organized and searched

according to the specific needs of a problem solver performing a particular task.

CHAPTER 3
JUSTIFICATORY REASONING IN TAX LAw:

A REVIEW OF PAST WORK

3.1 Introduction

Conceptual retrieval does not take place in a vacuum, but rather in the context of
some Other mental activity. Such activity, whether story understanding or directed problem
solving, establishes the goals that retrieval must satisfy. As discussed in Chapter 2,
research in cognitive psychology and case-based reasoning indicates that this phenomenon
is especially true for complex problem-solving tasks: Problem-solving goals play an
important role in how problem solvers understand problems and solutions and in how they
organize their memories. This dissertation examines conceptual retrieval in the context of a
particular problem-solving task, the task ofjustificatory reasoning in the domain of tax law.
In particular, the taxation of captive insurance arrangements is considered.

. The remainder of this chapter consists of four main parts. Section 3.2 examines
several AI techniques that have been proposed for the task of legal analysis, which involves
the justification of a legal classification. In section 3.3, several approaches to representing
justifications are reviewed. Section 3.4 presents a summary of a particular taxation issue,
that of captive insurance corporations, and its place in the tax law domain. This domain
serves as the test bed for developing a theory of conceptual retrieval based on problem-
solving roles. The purpose of these reviews is to establish the task and domain knowledge

with respect to which conceptual retrieval is considered in later chapters. Finally, section

43

44

3.5 discusses the implications for the idea of conceptual retrieval to be found in this

analysis of a specific task and a specific domain.

3.2 A1 Approaches to Legal Analysis

Legal reasoning consists, in large part, of two tasks, analysis and planning.
Analysis involves the classification of a situation1 as an instance of one or more legal
concepts. This classification must be then be justified by identifying relevant statutes and
case precedent that support the proposed categorization. In planning, the lawyer views the
client’s situation prospectively and recommends actions that satisfy the client’s goals while
minimizing the risk of unfavorable legal consequences. According to Buchanan and
Headrick [1970], the analysis task is essentially one of argument formation, and planning
consists of argument formation followed by risk assessment. In this sense, the two tasks
of legal reasoning interrelate in a fundamental way and share many of the same subtasks.

Researchers in Al have long recognized the law as an attractive domain for study.
The legal profession offers a tradition of examining its own reasoning processes, in the
specialty of jurisprudence, as well as a stylized method for problem solving, in the form of
stare decisis. Furthermore, the law offers a large volume of accessible and well-structured
text to serve as a source of experimental data. AI research in the law (AIL) has typically
addressed analytic problems — those of classification and justification — in an effort to
identify useful representation and problem-solving techniques that can later be applied to
more difficult planning problems. This body of research has evolved through several
stages of techniques, from rule-based approaches, through exemplar-based approaches, to

case-based approaches that more accurately reflect the legal enterprise.

 

1 The situation is often termed a fact situation. to emphasize that the situation is defined by the basic facts
that characterize it.

45

3.2.1 Rule-Based Approaches .

The earliest computer programs aimed at autOmating parts of the legal analysis task
were systems that represented legal knowledge primarily as rules. Popp and Schlink
[1975] described a prototypical system of this generation, Judith, which assisted in the
decomposition of legal issues. Given a legal issue, Judith provided a set of propositions
that define the issue. For each such subissue, the lawyer could either assert a truth value or
ask Judith to decompose the subissue recursively into its constituent propositions. If the
analysis “bottomed out” by reaching a subissue for which the lawyer desired more
information but the system had no decomposition knowledge, Judith generated a simple
query to a legal information retrieval system based on the propositions considered during
the current session. Other systems of the day employed similar approaches and embodied
similar motivations. For example, CORPT AX [Hellawell 1980] determined whether a
stock redemption qualified for favorable tax treatment under a particular provision in the tax
law, and Sprowl’s [1979] program used a decomposition of a statute to create a simple
legal form generator.

Though interesting as initial research vehicles, these early projects suffered from the
serious limitations of their basic approach. Responsibility for all of the significant legal
analysis lay with the system’s programmer and user. Hellawell’s work required an
exhaustive decision-tree analysis of stock redemption and attribution provisions in the
Internal Revenue Code, and Sprowl’s document drafting system required a similar level of
analysis by the user to decompose a statute and to produce “boilerplate” text. Judith did
offer a more flexible goal/subgoal representation of legal rules but was limited by its
inability to provide any assistance in searching the goal hierarchy. The only problem-
solving method it possessed was the “weak method”2 of issue decomposition (a user-

directed mixture of depth- and breadth-first search).

 

2 A weak method is a search technique that employs no domain-specific knowledge to guide the search.

46

Attempts to overcome these limitations within the rule-based framework produced
two notable systems, McCarty’s TAXMAN [1977] and Meldman’s legal analysis system -
[1977]. TAXMAN analyzed corporate stock reorganization transactions. Given a set of
high-level assertions about a corporation, its stock ownership, and stock transfers,
TAXMAN generated as output the proper tax classification for the transaction.3 McCarty
wrote TAXMAN in MicroPLANNER, a predicate logic-based language with general
procedural programming power. By encoding heuristic search rules in MicroPLANNER,
McCarty was able to give TAXMAN the ability to perform efiicient conceptual abstractions
based on experiential knowledge from taxation law. In a similar fashion, Meldman
implemented a network of assault-and-battery concepts in a predicate logic language,
creating a system that was able to classify fact situations as one or more types of intentional
tort. Meldman sought to explicitly represent different forms of legal knowledge —
concepts, rules, and generalized cases — using predicate logic rules as the epistemic
primitive.

From a Generic Task viewpoint, both McCarty’s and Meldman’s work can be
described as simple forms of hierarchical classification. TAXMAN consisted of a set of
small classification hierarchies, where each class was represented as a sub-class of its
parent type with additional constraints. These additional constraints constituted the
category’s establishment knowledge and was implemented as a set of structured pattem-
matching rules for abstracting input data into appropriate taxation concepts. Meldman’s
system also employed a classification hierarchy of tort types and sets of ordered pattem-
matching rules for determining the applicability of the type to the current facts at hand. In
both systems, the pattem-matching rules closely resemble the generic task of structured
matching, providing for a “bottom up” matching of fact patterns to the specifications of

more abstract concepts.

 

3 According to Chapter I, Subchapter C, of the Internal Revenue Code of 1954.

47

The notion of encoding legal concepts and heuristic search knowledge as rules for
classificatory analysis of a situation has been adopted by a variety of researchers and
practitioners in a variety of tasks and domains.4 The idea of matching fact situations
against templates of event and state classes, with instantiation of variables across relations,
permits these systems to address concepts beyond the scope of earlier propositional
systems. Yet McCarty [1983, 1987] argued that TAXMAN, the most highly developed
example of this approach, could not deal with basic concepts in particular domains of the
law because it lacked recourse to a deeper “understanding” of the domain. This limitation
follows from the fact that general legal rules abstract too much detail away from concept
instances, with aresulting inability to recognize fundamental differences between two fact
situations. This criticism led to a second generation of legal analysis systems stressing the

representation of legal concepts at a more basic level.

3.2.2 Exemplar-Based Approaches

Through case analysis, builders of early rule-based systems realized that the rule-
based approach would be insufficient for representing amorphous legal concepts. These
systems relied on logical templates for expressing concepts, but the domains in which they
operated were generally characterized by well-structured semantic relations. This feature
was especially present in the tax domains that were chosen. Areas of developing law,
though, resist this sort of formulation. Instead, they require an approach based on
reasoning from examples.

Levi [1949] presents the classical statement of a legal concept as a set of mappings

 

4 In the general law, these include two systems for determining defendant liability and litigation value.
LDS [Peterson and Waterman 1985] and SAL [Waterman, Paul, and Peterson 1986]. and a system for the assembly
of routine wills and revocable trusts, EPS [Schlobohm and Waterman 1987]. In the tax law. the approach has been
applied to the identification of compliance issues with respect to the inclusion of prizes and awards in gross
income [MacRae 1985], the determination of constructive ownership of stocks [Schlobohm 1985], and a system
for the analysis of prospective pension plans [Grady and Patil 1987]. The most publicized system implemented
for use in actual tax practice was ExperTAX. a program that identifies tax accrual, compliance. and planning issues
described by Shpilberg, Graham and Schatz [1986].

48

between fact patterns and a prototypical instance.5 On an exemplar-based theory, one
seeks to show that a situation is an instance of a particular concept by demonstrating a set
of value-preserving mappings from the instance to a known instance of the concept. In
order to accomplish this kind of reasoning in a computational manner, one requires a
scheme for efficiently and adequately representing concept exemplars and the mappings that
can be applied to them. This kind of approach holds great promise for the non-
computational legal theorist as well, since it enables the development of formal models for
the evolution of legal concepts and for the process of analogical reasoning in the law.

Work on the exemplar-based approach quickly focused on the development of
languages for representing primitive legal concepts. McCarty’s [1983] TAXMAN II
project aimed to create a language for describing legal fact patterns in terms of the rights
and obligations of the parties involved. One result of this project has been LLD, a deontic
logic that supports the representation and manipulation of mappings between fact patterns
[Mch 1985]. Using LLD, situations can be represented as collections of assertions
about the rights and obligations of involved parties, and mappings can be represented as
transformations that change certain features of a situation while maintaining other critical
features. Mappings can be decomposed into more finely-grained mappings along relevant
dimensions in the domain.6 The conceptual distance between two situations then depends
on the number and type of mappings (or “deformations”) needed to transform one into the
other. Finally, important concepts in the domain are represented as prototypical fact

patterns.7

 

5 Such a theory of legal construction was offered at least as early as the turn of the century by both Oliver
Wendell Holmes and Benjamin Cardozo. eminent legal scholars of the time.

6 In taxation domains. these dimensions might include concepts such as liquidity. risk. and control.

7 McCarty terms his approach “prototype and deformation.” Along similar lines, Cross and deBessonet
[1985, also deBessonet 1983] have developed a language called CCLIPS for representing mappings that involve
the primitive legal notion of causality.

49

Exemplar-based approaches have been tried in a variety of legal domains.
Schlobohm [with McCarty 1989] has developed a program, EPS II, that creates estate
plans tailored to a client’s goals. EPS II selects a prototype plan and then effects a series of
mappings from the prototype to a fully-instantiated plan, with each mapping preserving
features important to the client. These mappings are expressed in LLD using the primitive
relations of right and obligation. Gardner [1985, 1987] describes a legal analysis system
that uses generalized fact patterns to assist in classifying situations from contract law. In
essence, these generalized fact patterns correspond to exemplars of particular concepts,
such as offer, acceptance, and counteroffer, which can be instantiated based on the facts of
an input situation. This system has been used to classify typical situations from a first-year
law school contracts course.

The use of exemplars in legal analysis overcomes much of the brittleness found in a
purely rule-based approach. Mapping fact situations to prototypical instances of a concept
enables a more flexible consideration of a situation’s facts. However, some problems
remain. Exemplars, while possessing more detail than a single generalized rule, still
abstract away details from each instance of the concept. Levi’s [1949] discussion of
reasoning from examples in the law actually deals with particular cases that have (or have
not) been adjudged to be instances of a concept. As a concept evolves, different sets of
facts may become relevant to determining whether a situation is an instance of the concept.
Abstraction, however painstakingly done, may omit facts of a situation that could later be

relevant in classifying an instance.

3.2.3 Case-Based Approaches
Motivated by this concern, the current generation of AI research modeling the legal
analysis task has focused on case-based reasoning. Ironically, this move to the

representation of low-level instances has also enabled legal analysis systems to augment

50

their classifications with justifications. These justifications conform to legal standards by
citing relevant case precedent. This line of research is best characterized by the work of

Ashley [1990] and Branting [1989, 1991].

Ashley and Hypo Like other legal analysis systems, Hypo takes as input a set of facts
that define a particular legal situation (in this case, in trade secret law). As output, Hypo
generates one or more “three-ply arguments” to support different classifications of the
situation. These arguments are based on the favorable and unfavorable precedents that
most closely match the input situation. To assist the user with other aspects of legal
analysis, Hypo’s output also includes a complete list of all precedent cases that are relevant
to the situation at hand. These cases are stored in a library according to the facts that
characterize them.

The input situation is analyzed in terms of dimensions, abstract legal concepts
relevant to arguing the legal merits of a claim. Each dimension contains knowledge of a
particular domain factor, the values it can take, and rules that test whether the factor applies
to a situation. These dimensions play a variety of roles in the conduct of argument:

0 they characterize situations,

0 they index the memory of cases according to their presence or

absence from the case,

0 they differentiate among cases in determining the most relevant of

the cases retrieved from memory, and
0 they serve as the keys in constructing arguments and in generating

useful hypothetical situations.

For Hypo, arguments consist, quite simply, of citations to relevant cases. The first

ply of an argument involves citing the case that shares the greatest number of dimensions

51

with the input situation. The second ply of the argument consists of a response, made on
behalf of the opposing party, that takes advantage of some weakness in the mapping '
between the case cited and the input situation.8 Finally, the third ply offers a rebuttal on
behalf of the original party, countering the response made by the opponent in ply two.
Hypo generates such a three-ply argument for each case in memory that shares a maximal
subset of dimensions with the input situation.9

By maintaining a full set of cases from the domain, Hypo can classify a situation by
reference to specific precedent. This approach conforms to stare decisis and permits
analysis in domains having few or no generalized classification rules. However, Hypo
does all classification by case precedent. All the classificatory knowledge in the system
resides in the dimensions — they ultimately determine what features of a situation are
considered relevant in the domain. To the extent that Hypo is given a complete set of
dimensions and a complete set of cases, the system can provide adequate classification by

analogy. But Ashley offers no recourse to other classification knowledge.

Branting and GREBE Branting [1989, 1991] attempted to flexibly integrate rule-based
and case-based knowledge in a single legal analysis system. His program, GREBE, takes
an input fact situation and generates as output a single classification, justified by reference
to appropriate statutory (or common law) rules and precedent cases. A justification
generated by GREBE consists of a chain of reasoning from (some subset of) the input facts
to a legal classification of the situation. Each inference in the chain is annotated by the rule

or case that justifies the concluding assertion.

 

3 For example. by citing a case that shares a larger number of dimensions with the input situation but that
was decided for the opposing party.

9 That is, no other case has a set of facts that subsumes the set of facts shared by the input situation and the
case in question.

52

Rules and cases complement each other in the justification process. For example,
the use of a rule to justify a particular inference may be impossible if one of the rule’s
preconditions is a generalized predicate. Such a generalized predicate will likely not appear
as one of the base facts describing the situation. But if a precedent case exists in which the
input situation is classified as an instance of the generalized predicate, then the case may be
used to show the applicability of the rule to the situation. Thus cases may operationalize
abstract predicates from legal rules.

Conversely, a precedent case might be useful for supporting a particular inference,
except that the facts of the precedent do not match the facts of the current situation closely
enough. GREBE could use a legal rule to derive a required fact from the known facts of
the situation, thus facilitating a case analogy. Or there may exist a rule that shows one or
more of the situation’s facts to be a manifestation of a missing abstract term. This term
could then be used as a fact about the situation, again facilitating application of case
precedent. Branting terms this use of rules case elaboration. (He also describes the use of
rules for term reformulation, another means by which the use of cases is made possible

through use of a rule application.)

3.2.4 Status of AI in Legal Analysis

Case-based approaches such as Ashley’s [1990] and Branting’s [1989, 1991]
enable legal analysis that relies on specific facts of past classifications. The idea of using
multiple forms of knowledge (rules, exemplars, and cases) in legal classification is an old
one, tracing back to Meldman [1977] and McCarty [1977]. But only with the advent of
case-based analysis systems has the use of particular instances of past classifications been
implemented. The move from purely abstract knowledge to a mix of abstract and concrete
classification knowledge is essential for doing legal analysis that meets domain standards.

This move has also enabled these systems to provide more adequate justification of their

53

analyses by referring to case precedent when appropriate. As a consequence, interest in the
representation and manipulation of justifications has attracted more attention in the AIL
community.

Of course, even a case-based approach cannot circumvent the ultimate problem of
abstraction. Each case corresponds to specific classification instance and thus contains
every fact used to describe its legal situation. However, every case representation abstracts
from the “real” situation by including only a subset of the facts that characterize the
situation (that subset considered relevant by the encoder). This is true even of full-text
opinions generated by the courts. One can take the greatest care in describing a case within
some formalism, but there can be no guarantee that the description will be considered
complete in some later context. This problem is not specific to legal analysis, though, since
it applies to all representation. All other things being equal, cases provide an extra level of

detail over more general classificatory knowledge in the law.

3.3 Representing Justifications

With the development of more sophisticated techniques for automated legal analysis
has come a corresponding growth in interest in the generation of justifications for legal
classifications. Legal analysis must ultimately be supported by citation of relevant statute
or case precedent. In a very important sense, these justifications —— more so than the
classifications they suppOrt — are the key products of analytic problem solving in the law.
Several researchers have proposed techniques for representing justifications or justificatory
arguments. Though these researchers do not always offer precise methods for
manipulating justifications, their investigations provide insight into how different
representations promote the understanding and use of justifications. .

These approaches fall into two natural classes. In the f'nst, research is aimed at

capturing the content and structure of justifications — what kinds of assertions compose

54

justifications, and how these assertions are related. Such work generally seeks to broaden
understanding of human argumentation in various domains. In the second class of
approaches, justifications are represented so as to facilitate a particular computational
method in which justifications play a role. This class can be further subdivided according
to the type of argument being investigated:

0 tactical argument, concerned with moves that achieve specific

argument goals, and
0 strategic argument, concerned with moves that establish goals and

direct the flow of the argument.

3.3.1 The Structure and Content of J ustifications

One of the most thorough and influential analyses of justifications was performed
by the philosopher of science Toulmin [1958].10 Toulmin wished to model scientific
arguments from a variety of disciplines, but he found the traditional tools of logical analysis
inadequate for the task. To remedy this problem, Toulmin devised a model of justificatory
argument in which the roles played by particular assertions could be made explicit.
Furthermore, this model also made explicit the nature of the relationships among these
roles. Figure 4 depicts Toulmin’s structural model of justifications.

This model defines specific roles that assertions may play in a justification:

° ' the data that characterize a state of the world,

0 the claim to be concluded,

- the warrant that justifies the inference,

0 the backing that supports applicability of the warrant,

- a qualification on the inference, and

 

10 See Chapter 5 for another discussion of Toulmin’s work. in the context of a computational
representation of justifications.

55

 

 

<DATA> b so <QUALIFIER> <CLAIM>

since unless
<WARRANT> <REBUTFAL>

on account of
<BACKING>

Figure 4. Toulmin's Model of Argument

 

 

 

- a rebuttal that contradicts the claim, under the conditions of the
qualifier.
These roles either explicate the classical syllogism (by requiring that the nature of the
premises be stated explicitly as data, warrant, or backing) or extend it (by providing slots
for roles not directly filled by the syllogism’s propositions). Such a structure allows one to
capture a greater amount of the semantic content carried in a justification.

Several researchers interested in characterization of and computational support for
legal analysis or argumentation have adopted Toulmin’s rich model as the basis of their
approach. Marshall [1989] proposes a general framework that integrates Toulmin’s
representation of the logical structure of justification with a pragmatic structure of argument
goals and methods for achieving them. Within such a framework, justifications can be

represented and manipulated in terms of case facts and argument goals.ll Lutomski [1989]

 

11 Unfortunately, Marshall only outlines the integration of logical and pragmatic structures within her
framework. Her paper instead offers an excellent description of the use of Toulmin structures for representing
chains of inference and for annotating arguments with explanatory comments. This paper provided one of the
initial motivations for the work described in this dissertation.

56

argues that the use of Toulmin’s structure, with its focus on the functional roles of
assertions, provides a principled method for assimilating statistical evidence into complex
abductive arguments in the law. This approach also facilitates the teaching of the use of
statistical evidence to novice lawyers. Others who have employed Toulmin structures
include Lowe [1985] and Dick [1987]. The generality and completeness of Toulmin’s

model makes it widely applicable.

3.3.2 Strategic Argument Representation

Birnbaum’s [1982] argument molecules were an early attempt to model strategic
relationships among argument assertions. This approach arose out of research in natural
language processing aimed at building a computer program that could understand and
engage in arguments. In this approach, arguments consisted of networks of propositions
connected by support and attack relations. The goal of the approach was to identify
commonly- occurring patterns of support/attack relations and to investigate their use in
planning justificatory discourse. Once useful argument molecules were cataloged, the
program could use them to guide its selection of processing strategies. When the program
recognized a particular molecule in the course of understanding an argument, it could
generate an appropriate response based on its knowledge of the structure.

An example of an argument molecule was the contrastive positions structure. This
relation centers on a mutual attack relation between the two main propositions that sum up
alternative viewpoints of the arguers [Flowers, McGuire, and Birnbaum 1982]. Upon
recognizing a contrastive positions structure, the program could focus the selection of its
strategy based on knowledge of the structure. One of two goals would be established:
strengthening the program’s main assertion or undermining the opponent’s.12 In this way,

the molecule plays a role in understanding the arguments in which it appears.

 

12 In either case, the “net strength” of the mutual attack relations would be shifted in favor of the program‘s
key assertion. either by addition or by subtraction.

57

Argument molecules reflect the strategic roles played by assertions in justifications.
By adding strategic content to Toulmin-like structures, each molecule can initiate
expectations regarding the meaning of prior assertions and the utility of possible future
assertions. Additionally, Birnbaum, Flowers, and McGuire [1980] maintain that these
structures, and the expectations they engender, provide a useful way to organize case
memory in support of argumentation. Indexing cases by the expectations that they fulfill or
disappoint provides the arguer with a way to retrieve cases at times most advantageous to
their use in an argument. This enables cases to assist the user in achieving argument goals
indicated by the expectations and in avoiding potential pitfalls based on the opponent’s

opportunities.

3.3.3 Tactical Argument Representation

The work of both Ashley [1989, 1990] and Branting [1988, 1991], discussed
above, involves tactical arguments —— the assembly of a justification that achieves a given
goal, to classify a situation. Rissland [1985] and Ashley [1.985, 1990] describe a set of

moves that an arguer can make in the course of tactical argument:

0 citing the most similar favorable precedent case,
0 citing a counter-example,
° distinguishing a case, by claiming that a difference between a fact

situation and a precedent case justifies a different classification for
the situation, or
0 shifting the focus of argument, by presenting a related hypothetical
situation whose classification depends on a factor more favorable to
the arguer.
These moves, like Birnbaum’s argument molecules, provide knowledge about arguments

that is required for constructing justifications in a domain.

58

Ashley and Hypo In Ashley’s model, justifications consist merely of citations to cases
that share one or more dimensions with the input situation. Each case consists only of a set
of dimensions to which an outcome (a classification) has been assigned. Thus, Hypo
employs a trivial Toulmin structure, with base data connected to a conclusion via a case
warrant. This lack of structure suffices for two reasons. First, Hypo is intended only as a
first-level research assistant; all reasoning needed to flesh out the argument will be done by
the system’s attorney-user. Second, Hypo incorporates significant domain-specific
knowledge for comparing fact situations and for selecting the best cases to cite based on a
given fact situation. Most of this knowledge resides in the dimensions themselves, in the
form of rules for evaluating the magnitude and value of relevant factors.

However, this approach limits Hypo to drawing simple factual analogies between
two cases, since arguments carry no structural information as to why a particular outcome
follows from a particular set of facts. Hypo’s arguments are “flat” — they do not explicitly
reflect the inferential relationships among constituent assertions. While Hypo can generate
argument moves that achieve particular goals, it cannot represent or manipulate goals
explicitly. Nor can Hypo delineate the warrants that authorize specific conclusions in the

. domain.

Branting and GREBE Branting’s approach more closely parallels that of Touhnin than
Birnbaum’s or Ashley’s. In GREBE, justifications are networks of assertions in which
links carry citations to rules or cases that warrant the corresponding inference. As GREBE
attempts to construct a justification, rules or cases may be used to decompose a target
inference into a chain of simpler inferences, or they may be used to facilitate the inference
by mediating levels of abstraction. A case in memory consists, in part, of the justification
for the case situation’s classification. These justification structures can be used explicitly in

generating new classifications.

59

In many ways, GREBE represents the most advanced research system for legal
analysis. It incorporates multiple forms of knowledge and uses them all in each step of the
classification process. By explicitly representing the justifications, Branting is able to gain
maximum leverage from prior classifications in addressing new fact situations. The use of
Toulmin-like graphs facilitates the use of methods for justification construction that most

accurately reflect reasoning in the domain.

3.4 The Taxation of Captive Insurance Arrangements

While the law has provided a fruitful domain for Al research, a particular area of the
law — taxation — has proven especially attractive and fertile. Tax laws permeate all
business dealings. The financial effects of taxation play a such a large role in the long-term
profitability of businesses that many accounting and legal resources are devoted to tax
compliance and planning. In addition, the complexity and abstractness of taxation law
make it an important topic of legal research. McCarty [1983] argues that taxation is a
suitable initial target for AI precisely because of its complexity and abstractness. Its
artificiality allows AI researchers to investigate important issues germane to legal reasoning
without having to tackle the full breadth of human experience. Other areas of the law, such
as torts and contracts, rely almost wholly on the understanding of human relationships.13

As in other areas of the law, claims for a particular tax classification must be
justified by citation of relevant statute and case law. The need to justify tax classifications
becomes most acute for unsettled concepts, since each new case offers a potentially new
circumstance to be classified. One area of the tax law actively undergoing definition within
the courts at this time concerns the issue of captive insurance arrangements, in which

subsidiaries purport to provide insurance coverage to companies that hold substantial

 

13 Conversely, McCarty notes, law students are taught topics such as torts and contracts first. saving more
complex areas such as commercial and corporate law for later. This approach allows new students to rely on their
common sense understanding of the world as they obtain a solid foundation in legal principles.

60

ownership interest in them. Investigation of the issues raised by such corporations and
their tax classification has also led the courts, more generally, to refine many of their
conclusions regarding what does and does not constitute “insurance.” The work presented
here, in considering conceptual retrieval for the task of justificatory reasoning in tax law,
focuses further on the taxation issue of captive insurance arrangements and the body of

case law that is currently evolving around it.

3.4.1 Insurance

Insurance is a social device for accumulating reserves to meet uncertain losses. The
insured party (the “insured”) shifts the risk of a loss to the insurer by paying a fixed
premium in exchange for coverage in the event of a loss. The insurer assumes the risk of
many independent individuals under the assumption that the total losses suffered by the
body of insureds will be less than the total premiums paid. To this end, a vast and diverse
insurance industry has developed. The concerns of insurers are like those of any
businessmen: to sell their products (assumption of casualty risks) at a price (premium) that
enables the insurer to make a reasonable profit (the difference between total premiums and
total losses).

Because of the potential negative impact of catastrophic loss, the United States tax
code allows for the deduction of insurance premiums as a business expense [Duer,
Horvitz, and Coberly 1988]. Section 162(a) of the Internal Revenue Code offers the only
statutory provision regarding insurance. In part, it states:

“[T]here shall be allowed as a deduction all the ordinary and necessary

expenses paid or incurred during the taxable year in carrying on any trade or

business...”
Tax Regulations Section l.l62-l(a) then states that included among these ordinary and

“

necessary business expenses are ...insurance premiums against fire, storm, theft,

61

accident, or other similar losses in the case of a business...” While these pieces of statute
make clear that insurance premiums are deductible as business expenses, the Code goes no
further in defining what counts as an “insurance premium.”

A body of case law has evolved to better define what does and does not count as
insurance and insurance premiums. For example, it is an established matter of law that
payments made into a reserve against contingent losses constitute self-insurance but do not
constitute deductible insurance premiums [Spring Canyon Coal Co. versus the
Commissioner of Internal Revenue 19301.14 The notion of self-insurance contravenes the
tax accounting principle that the liability underlying a deductible expense must be both fixed
and ascertainable [Tax Regulations Section l.461-1(a)(2)]. In the case of an insurance
contract, while the expense is taken in anticipation of an expected loss, the insurance
contract fixes the liability to the amount of the premium15 and eliminates the possibility of a
greater loss. Since self-insurance lacks this important feature, the courts have held that
such payments are not deductible.

Other court cases have provided positive criteria for the presence of insurance. In
Le Gierse versus Helvering [1941], the Supreme Court held that insurance exists when
premiums are paid pursuant to the exchange of an “insurance risk”:

“Historically and commonly insurance involves risk-shifting and risk-

distributing.... That these elements of risk-shifting and risk-distributing

are essential to [an] insurance contract is agreed by courts and

commentators.” (Emphasis added.)

Thus, whenever one can ascertain that the risk of loss has been shifted and distributed, one
can claim that insurance exists. This two-prong test leaves open the question of what

constitutes risk shifting and risk distribution.

 

14 All case references appear as a separate bibliographic listing, Appendix A.

15 And perhaps a deductible of a fixed amount.

62

Risk Shifting Again, subsequent court decisions have refined the definition of these open
terms (for example, Steere Tank Lines versus the United States [1978] and Commissioner

of Internal Revenue versus Treganowan [1950]). Risk shifting typically involves at least

three elements:
- . the risk of loss is transferred from the insured to the insurer,
- a premium is paid from insured to insurer that is less than the

amount of the prospective loss, and
0 the insurer possesses the financial capacity to bear the risk and
indemnify the insured from loss.
In the analysis of risk shifting, the focus lies on the insured party and its contract with a
legitimate insurer. Factors indicating that the insured party no longer bears the risk of a

prospective loss provide positive evidence of risk shifting.

Risk Distribution The criterion of risk distribution focuses instead on the insurer and its
ability to share risk across a large number of parties (that is, to fulfill the social function of
insurance). Generally, risk distribution involves three features [O’Brien and Tung 1973]:
0 mass — The group of insured parties has sufficient exposure to risk for
the “law of large numbers” to take effect.
0 homogeneity — The exposures are sufficiently similar that the risk of loss
for each is approximately equal.
' independence — Each exposure is distinct enough from the others that the
probability of a single event resulting in losses for many insureds is
small.
Under these conditions, the law of large numbers holds that the risk borne by the insurer is

less than the sum of the risks borne by the insured parties. This property guarantees that

63

the individual risks of the insureds will be shared among the pool of insureds through the

medium of the insurer.

3.4.2 Captive Insurance Arrangements

For a variety of reasons, the cost and availability of adequate business insurance
coverage have become more unfavorable over the last twenty years. “[I]nsurance costs
now rival payroll and occupancy as the largest operating expenses for any business.”
[Duer, Horvitz, and Coberly 1988] One potential solution to this problem is to establish a
captive insurance arrangement. In this strategy, a business either purchases an existing
insurance company or incorporates a new insurance company for the purpose of obtaining
insurance coverage from the “captive” insurer. Such a captive insurance company may also
handle part or all of the insurance needs of other related companies, such as other
subsidiaries of the captive’s parent.

The range of possible captive arrangements is quite broad. The captive may be
owned by a single parent corporation or by a group of corporations, or a large portion of its
stock may be owned by a single parent with the rest distributed among many third-party
shareholders. The captive may insure only the parent, the parent and other related
companies, or both related companies and third-party customers. The captive may be
incorporated as a US corporation, perhaps in a state with Special captive insurance
laws16, or as a foreign corporation. Behind all such arrangements, the unifying theme is
the same: the parent company is now able to obtain (reasonably priced) insurance coverage
that would otherwise be unavailable.

The advantages of captive insurance can be significant. First and foremost, the
availability of insurance preserves the deductibility of premiums and provides a tax benefit.

If the captive is incorporated in a foreign jurisdiction, the captive may also pay little or no

 

“5 Such as the Colorado Captive Insurance Act. C.R.S. Section 10-6-101-130. 1973.

64

tax on premium revenue.17 The non-tax benefits of such arrangements are also attractive.
The company can obtain insurance that is not available in commercial markets, whether due
to excessive or unusual risks or just to an unfavorable past claims history. By preserving
the deductibility of premiums, the company can stabilize its reported earnings over the long
term.18 Furthermore, the captive’s status as a bona fide insurance company may allow it
access to reinsurance markets in which coverage is less expensive than direct insurance.
Access to lower-cost coverage can result in lower total insurance costs for the parent and its
related companies. Finally, a captive arrangement enables the parent corporation to retain
control over the defense of claims made against the firms. Otherwise, independent counsel
hired by the insurer decides how and when to defend claims of damage brought by outside
parties.

Historically, the Internal Revenue Service (IRS) has viewed captive insurance
arrangements with a high degree of skepticism. Its central concern involves the similarity
between captive arrangements and self-insurance in which the reserve against losses has
been incorporated as a separate entity. Ifthe substance of a captive arrangement does not
differ materially from that of self-insurance, then the IRS will eliminate the tax advantages
of the arrangement. These advantages are significant enough that corporations have been
eager to defend their captive arrangements in the courts. Their main goal is to distinguish
the arrangements from self-insurance.

The criteria for the existence of insurance outlined in Le Gierse stand as the basic
test of captive insurance. New factors come into play, though, in determining whether risk

has been shifted or distributed. For many years, the IRS argued that a captive insurer was

 

17 Indeed, many countries. such as Barbados and Bermuda, have created lax tax laws for the express purpose
of attracting foreign capital in the form of insurance premiums.

13 Without insurance, the (potentially large) expense of a casualty loss will fall in a single taxable year. or
perhaps in a small number of years. With insurance, the cost of such risks is spread over all taxable years in the
form of a fixed insurance premium. The result is that the company’s net profit will be stable with respect to the
cost of such losses.

65

part of the same “economic family” as its parent and sibling corporations. Under this
economic family concept, all risk of loss remains within the family and so, by definition,
risk could not be shifted. Thus, payments made to a captive insurer, whether by parent or
sibling, were not deductible as insurance expense. The courts were sympathetic to this
argument for some time. The decisions in Beach Aircraft Corporation versus the United
States [1986] and Steams-Roger Corporation versus the United States [1985] were based
largely on acceptance of the economic family argument.

More recently, the courts have begun to recognize that this argument in effect
violates the Corporate Entity Doctrine, first enumerated in Moline Properties versus the
Commissioner of Internal Revenue [1943]. This doctrine holds that corporations having a
legitimate business purpose are to be treated as separate entities for the purposes of
taxation. Following this doctrine, the determination of a captive’s validity must be based
on a consideration of the captive as a separate corporate entity; the existence of an
ownership relationship between a captive and a related company should not be used, a
priori, to deny the existence of insurance between the companies. A recent spate of cases
has seen the courts broaden their definition of insurance along these lines.

In Humana versus the Commissioner of Internal Revenue [1989], the court ruled
that premiums paid by a sibling corporation to a captive are deductible since, relative to
these two entities, risk can be shifted and distributed. This decision still disallowed the
deductibility of premiums paid by the parent to the captive, on the grounds that the parent
still bore the ultimate financial risk of a loss. Thus, risk was not shifted from parent to
captive. However, in Sears, Roebuck, and Co. versus the Commissioner of Internal
Revenue [1991] the court acknowledged that in practice risk could be shifted from parent to
captive under the right conditions. These conditions require attention to the basic elements

of risk shifting and risk distribution. If the elements for risk distribution are satisfied, then

66

the captive’s status as a bonafide insurer can play a role in determining whether risk of loss
has shifted from the parent insured to the pool of companies insured by the captive.

On this new line of reasoning, the determination of a captive’s validity must be
based on substantive issues regarding the existence of risk shifting and distribution. These
issues are numerous, but the courts have pointed to several factors of particular interest.
Among these are:

0 the nature of the insurance contract between the insured party and

the captive insurer,

' the degree of ownership interest held by the parent in the captive,

0 the adequacy of the captive’s capitalization for meeting potential
losses,

0 the captive’s history as a bona fide insurance provider,

0 the volume of unrelated company risks that are insured by the
captive.

These factors have been identified through the consideration of new fact situations that have
incrementally extended those considered in precedent cases. As the court argued in
Clougherty Packing Company versus the Commissioner of Internal Revenue [1987], at
some point along the dimensions that define these factors, the term “captive” may no longer
be appropriate to describe these arrangements. Thus, decisions regarding the validity of
captive insurance arrangements ultimately depend on the law’s standing view of insurance

itself.

3.5 Implications for Conceptual Retrieval

Chapter 2 describes the importance of understanding task and domain features for
conceptual retrieval. As Hammond [1989] argues, “Memory retrieval is a subtask of other

methods. First, we analyze the task and its functional needs, and then we determine the

67

role of memory in these terms.” [page 52] The preceding examination of the task of legal
analysis and the domain of captive insurance taxation suggests three key requirements on
conceptual retrieval for this task and domain: (1) integration of problem solving and
retrieval, (2) representation of justification structures, and (3) organization of case memory

according to the roles cases play in problem solving.

3.5.1 Integration of Problem Solving and Retrieval

With great foresight, Popp and Schlink [1975] declared that “the future of
information retrieval systems for the legal profession seems to require that they form an
important part of a more complex legal information system rather than being a research tool
in their own right.” [page 305] Their system, Judith, integrated legal analysis with
information retrieval by allowing a user to generate a data base query from the context of an
analysis session. Not only did this approach take advantage of the focus provided by the
analysis tool, but it also released the attorney-user from having to generate a complex key-
word query. Judith could generate a query that was directly “on point” with the analysis
done by the attorney thus far, while successfully managing the complexity that leads to the
pitfalls described by Bing [1978].19

McCarty [1983, 1987] has also argued forcefully for the idea that knowledge-based
systems in the law should ideally be hybrids that incorporate analysis, planning, and
information retrieval components. He offers two reasons for this integration. First, the
information retrieval component will have access to problem-solving knowledge and
patterns of inference in the domain, thus permitting more [concise and robust retrieval.
Second, the problem-solving components will have direct access to case materials that are
necessary to justify their conclusions. Given the evolutionary nature of the law, the

conceptual retrieval system should ultimately be integrated with both a problem-solving

 

'9 See Section 2.2 for a discussion of Bing’s findings.

68

system and with a full-text system, since the latter permits access to the most detailed

representation of the facts and reasoning that characterize past cases.

3.5.2 Representation of Domain Concepts and Justifications

Yet integration alone cannot insure adequate problem solving and retrieval skills.
Many of the rule-based legal analysis systems would still prove to be inadequate because of
limitations in the types of knowledge that they possess. Truly robust legal systems will
require, in some sense, an explicit representation of legal concepts and relationships among
them [Cross and deBessonet 1985]. This realization was one of the impetuses for the
move to exemplar-based systems, with their notion of prototype and content-preserving
mappings to other fact situations. A legal analysis or retrieval system in the domain of
captive insurance taxation would require some representation of concepts such as
insurance, risk distribution, and unrelated risks in order to efficiently reason about and
retrieve situations in which the concepts were relevant.

Due to the critical standards of problem solving in the law, legal analysis systems
must justify their classifications. And, as Branting [1988, 1991] showed, performing the
task of legal classification requires access to the justifications of past classifications. Thus,
conceptual retrieval systems must provide access to representations of the justifications
underlying precedent cases. Furthermore, such systems will almost certainly need to use
such representations in order to perform efficient, focused retrieval. This lesson led to the
recent generation of case-based approaches to legal analysis and has engendered a more
useful kind of analysis system — one that uses and produces justifications as a part of its

problem solving.

69

3.5.3 Organization of Case Memory by Problem-Solving Roles

The desire for a case memory of past justifications leads to a final implication: the
need to organize that memory according to the problem-solving roles that the justifications
can play in later justifications. This realization has arisen from work on argument both in
the law and in other domains. Research into argument molecules [Birnbaum, Flowers, and
McGuire 1980; Birnbaum 1982] was directed at natural language understanding and
processing of arguments in various domains. One of its key results was that assertions and
groups of assertions play particular roles in the conduct of argument and that these roles are
useful for indexing memory in a way that enables arguments to be retrieved at the most
opportune times — when they are relevant to understanding or generating a later argument.

The work of Ashley [1990] and Branting [1989, 1991] supports this notion as well
in the legal domain. Ashley argues that propositions and their meanings should be
interpreted in the context of their force in justifying claims. Branting’s GREBE system
relies directly on the ability to retrieve cases when they will be most useful in constructing a
new justification. However, Hypo and GREBE employ only feature-based indices. The
next step in this progression is to take advantage of such roles in indexing memory. This
approach would facilitate direct retrieval of cases according to their utility in problem
solving and make for a more coherent organization of memory. These roles are already
part of the vocabulary of the problem solver and thus offer a natural source of indices into
memory. In the end, this kind of memory organization offers increased support for

realistic justificatory problem solving in many domains.

3.6 Conclusion

In order to establish a problem-solving context for conceptual retrieval, this chapter
examines three distinct but related topics:

' Al techniques for the task of legal analysis,

70

0 the representation of justifications, and

0 the domain of captive insurance taxation.

The first of these three topics introduces many of the task features that characterize legal
analysis and justification. The second describes approaches for representing the content
and structure of justifications that can be used to support problem solving and information
retrieval. Finally, the domain review delineates important concepts and lines of reasoning
that typify reasoning about captive insurance cases.

The chapters that follow draw from these reviews in order to offer a theory of
conceptual retrieval for justificatory reasoning. Chapter 4 expands on the discussion of
legal classification, offering a full task analysis of legal justification and describing a
problem-solving architecture for the task. Chapter 5 then expands on the discussion of
representing justifications, offering a formalism for representing justificatory arguments
that directly supports conceptual retrieval in the context of the problem-solving architecture
from Chapter 4. Finally, Chapter 6 describes a case memory based on the formalism from
Chapter 5, using the domain of captive insurance taxation as a testbed for the case memory

methodology.

CHAPTER 4
A PROBLEM SOLVING ARCHITECTURE

FOR LEGAL ANALYSIS

4.1 Introduction

The previous two chapters establish a context in which to consider the conceptual
retrieval problem. Chapter 2 defines the problem and considers work in three disciplines
aimed at solving it. From this work arises the unifying theme that memory can fruitquy be
organized according to the roles that stored items might fill in problem solving. Chapter 3
then analyzes past work on the task of legal justification, in which conceptual retrieval
plays an integral part. Also described in Chapter 3 is the taxation issue of captive insurance
arrangements, a developing area of the tax law that serves as an interesting experimental
testbed for the conceptual retrieval problem.

This chapter proposes a problem solving architecture, in the technical sense of
Sticklen [1990], for the task of legal justification. In this architecture, Generic Task
problem solvers are integrated with a case memory organized by the roles that past
justifications can play in assembling future justifications. The remainder of the chapter
consists of two parts:

a a task analysis [Chandrasekaran 1990] of legal justification, and

0 a problem solving architecture derived from this task analysis.

Discussion in this chapter focuses on components of the architecture other than the case
memory, which will be described more completely in following chapters. The focus here
is on the task of legal justification and, in particular, the functional requirements that this

process places on case memory.

71

72

4.2 A Task Analysis of Legal Justification

The task of justification poses a straightforward information processing problem:
given the facts of a situation and an assertion about the situation, produce a justification for
believing the assertion1 (Figure 5). Ordinarily, one expects the justification to establish an
inferential chain leading from (some subset of) the facts to the assertion. Such an
inferential chain consists of a sequence of true assertions, each linked to its antecedents in
the chain by a warrant [Toulmin 1958] that justifies its inference. In different domains of
discourse, different types of warrant and backing are considered “valid”, that is, as leading
to acceptable conclusion. Methods of generating and evaluating justifications depend

largely on the nature of warrants and their backing.

 

A set of facts An assertion
describing a situation about the situation

Justify
Assertion

A justification for
believing the assertion

Figure 5. The Task of Justification

 

 

 

 

1 In practice. this task becomes one of creating multiple (and competing) justifications for the assertion
[Ashley 1990. Branting 1991]. These justifications can then be evaluated and ranked by their persuasiveness.
This project considers the specific task of generating and evaluating a single justification. If desired, knowledge
of the type described by Ashley and Branting could be added to this analysis for consideration of the more general
problem.

73

In the law, past cases serve as the primary warrants and backing for new

justifications, and so the standard method for justification is based on analogy [Levi 1949,

Ashley 1990]. This method sets up three subtasks: abstract the facts into appropriate legal

terms, retrieve relevant cases from memory, and use the precedents to justify the target

assertion (Figure 6). Ashley describes a control regime for this method that is essentially

linear — abstract, retrieve, and apply — but typically these subgoals can be interleaved in

complex ways. For example, in applying a precedent, the problem solver may find that a

portion of the precedent’s justification fails for the current situation. This difficulty can be

addressed by retrieving another past case and using its justification.

 

Abstract facts Retrieve Use precedent

 

TASK
Justification

METHOD
Analogy

J
I I I

SUBTASK SUBTASK SUBTASK

precedents

Figure 6. The Analogical Method of Justification

 

 

4.2.1 Fact Abstraction

The subtask of fact abstraction takes as input a set of data and gives as output one

or more abstractions of the data. For the domain of tax law, these data correspond to facts

describing a situation. Abstraction of facts is critical to the justification task. The power of

74

a justification lies in its ability to persuade. The appropriate use of a domain’s technical
vocabulary increases the precision and conciseness of the argument, thus making the
justification more persuasive. Additionally, the retrieval of relevant cases may depend on
the availability of important abstractions (Chapter 2).

One method for this task often used in the law involves matching patterns of facts to
abstractions that characterize the patterns [Meldman 1977, McCarty 1977]. To employ this

method, a problem solver must possess knowledge in the form of D1, D2, Dn —> A
rules, where the Dis are data and A represents an abstract term. A variation on the method

uses precedent cases in place of pattern match rules [Branting 1991]. In this variation, the
problem solver retrieves past situations that share features with the current situation and
notes the abstractions that characterized these precedents. Branting refers to this variation
as “operationalizing” abstraétions. This variation proves especially useful when a problem
solver requires an abstraction in order to use some other knowledge but lacks the compiled

pattern match rules needed to infer the abstraction.

4.2.2 The Subtask of Case Retrieval

The subtask of case retrieval can be characterized as follows: given a set of data,
produce a set of relevant cases. In Chapter 2, several different views of this task are
described and methods for performing it are presented. The input data can include facts,
abstractions of the facts, or goals of the problem solver (such as an assertion to be
justified). One family of methods for performing case retrieval relies on an associative
memory that requires no search knowledge. However, the methods most studied in Al use
explicit indices to organize memory. To use these methods, a problem solver requires
knowledge for differentiating among (classes of) cases based on the features that

characterize them.

75

 

A set of facts Abstractions An assertion One or more

describing a situation of these facts about the situation precedents

 

   

  
   
   

Use
Precedent
J ustifications

A justification for
believing the assertion

Figure 7. The Task of Justification by Precedent

 

 

 

4.2.3 The Subtask of Precedent Application

Finally, the subtask of using precedent justifications produces a justification for the
target assertion (Figure 7). This subtask can be performed by adapting an input
justification so that it applies to the current situation. Such a case adaptation method
requires knowledge for evaluating justifications, determining the rcason(s) for a
justification’s inadequacy, and modifying a justification to overcome its weaknesses. In
the law, one might accomplish this modification subtask by applying another past
justification as a “patch” to the inadequate portion of the justification being modified.
Figure 8 depicts the full task analysis for legal justification, including possible methods and

their subtasks.

76

Seamus. and e 5%? ea... < .m 2:3...

 

 

 

 

 

connocemi 53093.3. cowaocumam
comoaoa women—ea 383:.
£82 2.3.5 assesm
¥m<hm3m ¥m<wm3m ¥m<flmbw
soaom beau—z mango and 9 8.3. $280 he
confine—E 930 8an x85 o>u£oomm< cemtaanU conga—qua
GOEHME GOP—.92 GOP—ha: DOEm—z GOEm—z
_ _ _ _ _ m
“5385 8: 8588.3 96.53— fiofl Humane.
Mm<hm2m ¥m<hm3m ¥m<HmDm
_ fl _
.3322.
nozfimsa
cage—mung

”Mme?

77

4.2.4 A Control Strategy for Legal Justification

Given a task/method/subtask analysis of the justification task, one can develop a
control strategy for generating justifications. Table 2 displays one such strategy, which
generalizes the method proposed by Ashley [1990]. This control strategy does not give a
complete picture of the problem-solving process, since each step involves further problem-
solving. But specification of the high-level strategy provides some guidance in
understanding legal justification, as well as guidance for building a knowledge-based

system to perform the task.

 

Table 2. A Control Strategy for Legal Justification

1. Request assertion to be justified and initial data describing the
situation.

Identify relevant legal abstractions of the data.

Identify relevant past justifications (cases).

Select the best match among the available cases, and propose it as a
solution.

Evaluate the proposed solution. If it suffices to justify the assertion,
then output an appropriate citation and halt.

Critique the proposed solution to identify individual inferences
responsible for the justification’s insufficiency.

For each inference identified, modify the inference chain so that it
holds for the current situation.

Go to Step 5.

00 \l 0‘ U! #UJN

 

4.3 The Problem Solving Architecture

Following this task analysis, one can develop a problem solving architecture (PSA)
for an agent that performs the task of legal justification. The description of the PSA given
here follows the treatment of Sticklen [1990] and as such specifies three necessary features

of the agent:

78

0 the component problem solvers (sub-agents) which comprise the agent,

0 the channels of communication among the sub-agents, and

0 the problem-solving method employed by each.

These features are determined in large part by the decomposition of the analogy method
presented in the preceding task analysis. They depend directly on its subtasks and the
domain knowledge available for solving each.

Figure 9 presents a PSA for legal justification. The following sections describe the
components of this architecture, the communication channels among them, and the problem-
solving methods they employ. These task-specific agents are implemented as instances of
Generic Task problem solvers. Detailed discussion of how the Case Memory performs its
task appears in Chapters 6 and 7. For now, it is important to describe only the functional
requirements that the legal justification architecture places on the conceptual memory

component.

4.3.1 Components

The architecture consists of four sub-agents. The Justification Generator directs the
activities of all the other agents and embodies the use precedent to justify subtask. The Fact
Abstractor and Case Memory correspond to the two subtasks abstract facts and retrieve
precedents, respectively. Finally, the Situation Data Base serves as a shared blackboard for
the other agents. All facts and fact abstractions for the current situation reside in this
database, which facilitates efficient communication among system components.
Furthermore, the Situation Data Base manages all interaction between the system and the

user.2

 

2 More generally. the Situation Data Base could manage interaction with other external sources as well.
such as external data bases. See Chapter 9 for a brief discussion.

79

 

song—1:33 Emu: e8 2208302. meg—om EoEoE < .a ensure

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Baggage. 3a...— emam 8.5 5.355%
meouoahmexx
Honk
: 8532 E cabana—U
couseosbm 825385 ' 80mm
. . define?
. . 3am Sun— +
. . - aowaas
_ 5532 M 353.20 weave—331‘ 8am
coeaeogm 302882: .aomm .5982
meouogmfix 8am 232334.. San.
sown. .8883. .88"— .5982
v 3:3on v
«Seaman. oeueom E6534
and 2: ..8
59333

bees: ~30

 

 

 

8.9.2.90 553583.

80

4.3.2 Communication

Agents in the architecture communicate based on the input-output services each
provides. Communication follows a message passing model: when an agent needs a piece
of external data, it sends a message to the agent that can provide the data. Interactions
among the agents are limited to the repertoire of messages each can handle, along fixed
communication paths.

Table 3 depicts the communication channels present in the proposed PSA. Of
special interest are the paths to and from the Case Memory. The Justification Generator
sends requests to the Memory in the course of its problem solving. These requests may be
for cases similar to a particular situation, or they may be for cases that can play a certain
role in assembling a justification. In order to fill these requests, the Memory may need to
refer to the facts of the current situation. A request is then sent to the Situation Data Base,

which provides this information.

4.3.3 Problem Solving Methods

From the task analysis above, one can identify individual generic tasks
[Chandrasekaran 1983, 1987] that play a role in legal justification. Each agent in the
problem-solving architecture (Figure 9) can be viewed as composed of one or more
Generic Task problem solvers. This section describes each PSA agent in terms of the
generic method it employs and the knowledge it possesses.

Analysis of a variety of “real world” tasks and domains has led to the identification
of what Chandrasekaran has termed generic tasks (GTs). Each GT is a problem-solving
method that has been found useful for solving a particular task in a variety of domains.

GTs have proven especially beneficial as models of problem-solving types because they
explicitly delineate:

 

81

Table 3. Channels of Communication in the PSA

J ustification Generator

Input:
Requested From:

Input:
Requested From:

Output:
Requested By:

Fact Abstractor

Input:
Requested From:

Output:
Requested By:

Case Memory

Input:
Requested From:

Output:
Requested By:

Situation Data Base

Input:
Requested From:

Input:
Requested From:

Output:
Requested By:

Assertion, Facts, Fact Abstractions
Situation Data Base

Past J ustifications
Case Memory

Justification for the Assertion
< System User >

Assertion, Facts, Fact Abstractions
Situation Data Base

Fact Abstractions
Situation Data Base

Assertion, Facts, Fact Abstractions
Situation Data Base

Past J ustifications
Justification Generator

Assertion, Facts
< System User >

Fact Abstractions
Fact Abstractor

Assertion, Facts, Fact Abstractions
< All Three Problem Solvers >

 

82

0 the types of domain knowledge required to apply the method, and
0 the nature of the control strategy used to perform the task.
These elements of a GT's definition constitutes an epistemic vocabulary for describing

problem solvers independent of their implementation technology.

Justification Generator The top-level agent for generating new justifications performs
the generic task of a routine design [Brown 1987, also with Chandrasekaran 1986]. This
kind of problem solving develops when a designer works on a problem many times, each
time with different but substantially similar requirements. In routine design, the designer
decomposes the task into a number of subproblems and then solves each in turn. For each
subproblem there exists a relatively small number of well-understood operators to apply.

A routine design problem solver consists of a hierarchy of design specialists, each ‘
concerned with a particular feature of the problem solver’s output. The lowest-level
specialist is the design step, which makes one design decision (such as selecting the value
of a parameter). Aggregations of steps compose a design task, and a number of tasks
constitute a design plan. A plan specialist embodies a method for producing a design or a
part thereof. At any point in a plan or task, a constraint specialist may test the design to
insure that it satisfies a necessary condition. If the constraint specialist identifies a
problem, a failure handler is invoked to repair the design.

Viewing a justification as an artifactto be designed, the Justification Generator can
be viewed as a routine designer.3 The designer’s top-level plan corresponds to the control
strategy given in Table 2. Each operation in the strategy represents a task to be performed
in generating a new justification. The first three tasks require no design knowledge per se;

rather, they involve invocation of the Situation Data Base and Case Memory. Tasks 4

 

3 This is not to say that case-based justification is necessarily a routine process. The tasks of justification
and argumentation may be quite non-routine — as in the creation of a novel analogy. In domains with some
settled rules of law and large bodies of precedent. though, many new cases can be handled using routine
knowledge. The approach outlined here addresses this more restricted situation.

83

through 7, though, embody the Generator’s knowledge for creating new justifications from
existing ones. This process may entail further requests to the Case Memory. When a new
inference must be made, or when an existing inference in the proposed justification must be
supported, the Generator asks the memory for another precedent that can fill this role.
Methods for selecting, evaluating, critiquing, and modifying justifications employ
similar forms of knowledge across most legal domains. Chapter 3 reviews several
computational methods for these tasks, in particular the work of Ashley [1990] and
Branting [1989, 1991]. One common approach that can span the four subtasks involves
the generation of hypothetical examples. A useful hypothetical might share all of the facts
with some base situation except for one, or the value of a fact may differ slightly. Such
hypotheticals can be used to evaluate a proposed justification (e.g., by testing a “slippery
slope”) or to critique a proposed justification (e.g., by pointing to a weak inference
therein). Ashley explicates a set of rules for generating a hypothetical and another set for

using them in justification.

Fact Abstractor A set of structured matchers and hierarchical classifiers constitutes the
fact abstraction component of the architecture. The generic task of structured matching
[Bylander, Johnson, and Goel 1991] addresses the problem of selecting one item from a
small number of distinct alternatives. In this component, this task may involve selecting
one abstraction from a set of abstractions or recognizing the presence of a single
abstraction. In the course of making such a decision, the problem solver may partition its
task into simpler decisions, make the simpler decisions, and then combine these
“subdecisions” to make its assigned decision. This is the central motivation for structured

matching.

84

 

Insurance Risk Exists

 

 

l 7
Insurer is Parent Control
Indemnified of Funds
by Parent
1 l I
Parent Parent Parent Is
Guarantees Provides Obligated
Coverage Security by Contract

Figure 10. A Sample Structured Matcher

 

 

 

Figure 10 depicts a simple example from the domain of captive insurance law. This
structured matcher determines whether the abstraction of insurance risk is present in the
current situation. To make this decision, the matcher first makes decisions regarding
whether the insurer is indemnified by the parent and whether the Parent retains control of
the premium funds. It then uses these abstractions in deciding whether insurance risk
exists. Each node in this hierarchy contains an ordered set of rules mapping patterns of
data to decision alternatives. For leaf nodes, these patterns refer only to facts of the
situation; for internal nodes, patterns may also refer to the decisions made by its subnodes.
Partitioning patterns in this manner provides both computational efficiency and an explicit

identification of subdecisions that are meaningful in the domain.4

 

4 Note also that such subdecisions may themselves represent meaningful fact abstractions in the domain.

If so. their results can be made available to problem solvers other than the structured matcher (for instance, to the
routine designer).

85

Some fact abstractions can be organized as a generalization hierarchy. In these
cases, the Fact Abstractor employs the generic task hierarchical classification [Sticklen,
Chandrasekaran, and Josephson 1987]. This GT selects one or more abstractions from the
hierarchy of abstraction specialists using an establish-refine control strategy. When a
specialist determines that its abstraction accurately describes the situation (that is, the
specialist “establishes”), it attempts to refine the classification by allowing its subordinates
in the hierarchy to attempt to establish themselves. Classification typically begins at the top
of the hierarchy, with the most general abstraction, and then proceeds until one or more
leaf-level abstractions have established.

Figure 11 depicts a hierarchical classifier from the captive insurance domain. The
knowledge that each abstraction specialist uses in its attempt to establish can take a variety
of forms. In the Fact Abstractor, though, the establishment task is achieved by structured
matching over sets of pattern-mapping rules. For each specialist, there is a dedicated
structured matching problem solver to recognize the presence of the specialist’s abstraction.
Such structured matchers and hierarchical classifiers allow the Fact Abstractor to identify
abstractions along multiple dimensions of domain importance. They may be invoked either
individually or en masse by request from the Situation Data Base. On an en masse request,

the Fact Abstractor initiates the matching and classification agents in a pre-defined order.

 

Insurance Provider

 

I 1

Insurance Company Mutual Association

Figure 11. A Sample Hierarchical Classifier

 

 

 

86

Situation Data Base The Situation Data Base is an instance of the intelligent database
generic task [Mittal, Chandrasekaran, and Sticklen 1984]. An intelligent data base manages
situational, case-specific data using knowledge about data objects and relationships among
them. The behaviors that an intelligent data base can be expected to produce include:
0 the ability to obtain the value of a data object, whether from the
system user or from some other source, and
0 the ability to answer queries about data objects, retrieving either data
values or abstractions on data values.
As conceived in the context of the problem solving architecture, the Situation Data
Base provides only a subset of these capabilities. The Data Base maintains two types of
data objects, facts and fact abstractions. For each fact, it has a slot for the fact’s value and
a method for obtaining a value from the user. For each fact abstraction, it has a slot for the
abstraction’s value and a pointer to the Fact Abstractor agent capable of determining its
value. Thus, all abstraction behavior resides outside the data base, as does all

representation of data relationships.

Case Memory A detailed description of the organization and problem-solving methods

used by the Case Memory appears in Chapter 6.

4.4 Conclusion

A task analysis of legal justification leads to an appropriate problem solving
architecture for generating justifications from precedent cases. Several of the subtasks
established by this method, including data abstraction and routine design, are well
understood in the law and in Al. Consequently, one can describe problem solvers that

perform these subtasks using knowledge available in the domain.

87

Less well understood is the subtask of case retrieval. The PSA defines a task-
specific context in which to perform retrieval, placing particular functional requirements on

the case memory. Two open issues remain:

How are justifications represented in memory? This representation determines
not only the content of the case memory but also the low-level detail of the routine designer

that generates justifications.

How is the case memory organized and searched? In order to support conceptual
retrieval, the case memory must be able to retrieve cases based not only on situation

features but also on the role the case will play in problem solving.

These issues are closely related. Tax accountants and lawyers understand new
justifications based largely on their relationship to past cases. Thus, a representation of
justifications should take such relationships into account. These relationships also reflect
the roles that past cases play in generating justifications, and so they should provide
guidance for structuring case memory effectively. The next two chapters describe answers
to these open questions that take advantage of the connection between case understanding

and memory organization.

CHAPTER 5

A FUNCTIONAL REPRESENTATION

0F JUSTIFICATORY ANALYSIS

5.1 Introduction

Tax accountants organize their understanding of legal justifications in a way that
relates the justification to other justifications they already understand and to the accepted
principles of tax accounting. This method of understanding results fiom the need to justify
conclusions in terms of past decisions and existing statute. Thus, the use to which
justifications are put affects not only how cases are organized in memory but also how
cases are represented. One of the central contributions of this work is to Show the intricate
relationship between case representation and organization that follows from consideration
of how justifications are used.

This chapter proposes a representation of justificatory analyses that is based on the
Functional Representation (FR) of devices [Sembugamoorthy and Chandrasekaran 1986].
Extending the intuitions of Toulmin [1958], this representation captures the relationships
among justifications used as warrants for drawing conclusions. The rest of the chapter
includes:

0 description of the motivations underlying this work, and

0 specification of a functional representation for justificatory analyses.

The latter of these sections adapts and extends the vocabulary of the FR to the
representation of legal cases that embody justifications. In the chapter’s conclusion, this

representation is related back to the issue of case organization.

88

89

5.2 Motivations

The initial motivation for this work came from the work of Toulmin [1958], a
philosopher of science who questioned the utility of traditional logic as a tool for analyzing
scientific argument. Toulmin developed a model ofjustificatory argument that explicitly
identifies the roles that assertions can play in justifying a claim. This model is reminiscent
of the FR, which expresses device behavior in terms of the mechanisms by which behavior
is understood. The affinity between Toulmin’s model and the FR has motivated this

research aimed at understanding legal analyses as abstract devices.

5.2.1 Toulnrin’s Model of Argument

On Toulmin’s view, logic deals not with techniques of inferring but rather with
retrospective justification of Claims. Traditional work in logic has focused on the types of
proofs found in mathematics; Touhnin recognized that most arguments do not correspond
to this notion of absolute proof. Instead, he set out to extend the classical syllogism to
incorporate practical issues of justification. Abandoning mathematical proof, he adopted
jurisprudence as the foundation of his model of logical argument. Jurisprudence stresses
the persuasive nature of argument — “making the case” by providing appropriate evidence
and citing the warrant that justifies inference.

This approach led Toulmin to greatly enrich the vocabulary available for describing
arguments. Figure 12 reproduces the template for justifications he proposed. In this
model, terms such as data, backing, and warrant characterize assertions as playing
particular roles in the justification, in particular relationship to other assertions. The model
extends the classical syllogism in at least two ways. First, its vocabulary more precisely
explicates the roles filled by assertions (both explicitly stated and implicitly intended) in the
justification. Second, this form of argument permits a homogeneous representation of

different types of justifications, whether deductive, inductive, or abductive.

90

 

<DATA> so <QUALIFIER> <CLAIM>

since unless
<WARRANT> <REBU'1'I'AL>

on account of
<BACKING>

Figure 12. Toulmin's Model of Argument

 

 

 

Figure 13 depicts a typical argument in the classical syllogism and in Toulmin’s
representation. This example demonstrates how Toulmin’s vocabulary enables one to state
clearly the purpose of each assertion in the syllogism and the specific relationship among
assertions. While the classical syllogism confounds the nature of the two premises,
Toulmin’s model shows that the former serves as an assumption and the latter as the
warrant for the inference.

By representing justificatory roles unambiguously, the model facilitates
development of criteria for judging the quality of justifications in different domains. For
example, warrants express the force of a conclusion and are typically domain-independent.
The warrant in Figure 13 could be used in any domain of discourse, from law to sociology.
But the nature of the backing, which supports application of the warrant to the data, varies
across domains. In the law, the backing must usually appeal to case precedent or statute,

whereas in other domains category definition or statistical evidence are more appropriate.

91

 

Petersen is a Swede.

! S I . l . l B C l 1'
Petersen is almost certainly not a Roman Catholic.

 

Petersen is . > SO almost Petersen is not
a Swede. certainly, a Roman Catholic

SINCE
A Swede can be taken
to be almost certainly
not a Roman Catholic.

 

UNLESS
He is an immigrant.

ON ACCOUNT OF
The proportion of Roman Catholic
Swedes is less than 2%.

Figure 13. An Argument in T oulmin's Representation

 

 

 

5.2.2 Functional Device Understanding

Toulmin’s model of justificatory argument calls to mind the Functional
Representation of Sembugamoorthy and Chandrasekaran [1986]. The FR provides a
language for describing devices based on their known functions or goals. A number of

intuitions underlie the FR, including [Allemang 1990]:

92

0 Devices have functionsl

0 A device may consist of component subdevices, in which case the

device achieves its functions by coordinating the functionality of its
components.

0 How a device achieves its functions is irrelevant to understanding its

role as a component in another device.

In the FR, a device is decomposed into its components, whose own functions are
then composed to achieve the functions of the device. The causal behaviors of each device
or component are indexed according to the functions they realize. Since the functions of a
device can be expressed in terms of component functions, the FR supports abstraction of
behavioral detail across levels of the device decomposition. This feature allows the
descriptions of a device and its components to be written in languages at different levels of
absn'action.2

Figure 14 gives an example of the FR for a simple device, the ordinary household
Clothespin. The behavior Open Arm presents a sequence of causally related states that
illustrates how the function Open is achieved. The label on each link refers to the
mechanism understood to be responsible for the state transition, given that the preceding
state was reached. These labels may point to basic knowledge of the world, such as one of
Newton’s laws, which lies outside the FR itself. However, they may also point to a
function of some component of the device responsible for achieving the state change. To
see a more detailed description of how the state change is achieved, one can refer to the

behaviors implementing the component’s function.

 

1 “This is taken as the defining characteristic of a device; a device is something which has a known
function.” [Allemang 1990, page 2]

2 For example. one could describe a transistor radio in terms of transistor circuits. radio signals. or tuning
frequencies. Each level of description would be appropriate under different circumstances.

93

 

Device Clothespin
Components Arm, Pivot, Spring
Functions Open, Close, Hold

 

Function Open Of Device Clothespin
Given: Force applied at Pivot > Restoring Force
Make: Increase Distance between Arms
By: BEHAVIOR Open Arms

 

 

Behavior Open Arms

Force applied at pivot > Restoring Force

By Function Transmit Force Of Pivot

Set Force on Spring > Restoring Force

By Knowledge Of Newton's Second Law

Increase Distance between Arms

Figure 14. A Fragment of an FR for a Clothespin

 

 

94

The FR was originally intended for describing physical devices that had been
designed with particular functions in mind [Sembugamoorthy and Chandrasekaran 1986];
Later work demonstrated its utility for describing other physical systems, such as human
body physiology [Sticklen 1987], to which one can ascribe functional characteristics.
These uses of the FR are teleological in this sense: functions correspond to behaviors that
either

0 support a designer’s goals, or

' offer some analytic advantage in understanding a system.

Adopting this sense of teleology, researchers have extended the FR for use in describing
abstract devices such as computer programs [Allemang 1990] and ecological systems

[Sticklen and Tufankji 1992].

5.2.3 Viewing a Legal Case as a Device

One can view a legal case as an abstract device with the function of supporting a
claim, given an initial set of assertions. Different parts of the case play specific roles, such
as to rebut a counterargument or to propose a hypothetical situation, that in concert achieve
the case’s primary goal(s). Considered this way, a legal case can be modeled using the FR
in a way strongly reminiscent of Toulmin’s view of justification. The graph used by
Toulmin to depict a justification corresponds to a behavior in the FR, a “causal” line of
reasoning that delineates inference relationships among assertions.

Warrants conespond to link annotations in the FR. They may appeal to knowledge
of the world or to empirical data, as Toulmin’s backings do, or they may refer to the
function of another case capable of supporting the inference link. Such layering of
justifications, though not discussed explicitly by Toulmin, reflects the spirit of justification
that he sought to capture: Claims that are challenged can be supported by appeal to more

detailed arguments. This intuition of the affinity between Toulmin’s ideas and the FR

95

motivates the representation proposed here. Table 4 shows the basis for a mapping

between Toulmin’s terminology and that of the FR.

 

Table 4. Viewing a Legal as a Device in the FR

FR Terminology Toulmin’s Terminology
Device ( legal case )
Function ' ( Legal issue )
Behavior Justification
Preconditions Data
Link annotation Warrant
< None > Backing
Postconditions Claim

 

5.3 Representing Legal Analysis in the FR

In the domain of taxation law, justifications reside within the bodies of legal cases.
Each case includes not only the justifying chain of inference but also contextual information
that defines the situation to which the justification applies. Thus, an adequate portrayal of
legal justifications should account for both the case context and the chain of inference. This
section describes a language for representing legal justifications, based on the Functional

Representation, that accounts for both.

5.3.1 The Legal Case as a Device

A case corresponds to a device in the FR (Table 5). Case descriptions consist of
three elements:

0 an identifier,

0 a case context description, and

96

0 one or more issue descriptions.
The identifier serves as the name by which other devices may refer to the case. The case
context description includes information necessary for understanding and reasoning about a
case in the domain. Finally, the case is characterized by the issues that it addresses. These
issues identify the functions that the case might play in analysis. Both the case context and
issue descriptions are discussed in greater detail below. But first, several device-level

issues must be addressed.

 

Table 5. Language Grammar: A Legal Case

<CASE> ::= Case identifier
<CASE CONTEXT)
<CASE ISSUES>

 

Selection of Case Functions Since the FR is a language for describing, not defining,
' devices, every possible function of the case need not be included in its description. One
would initially include only those functions intended for the case by the justification’s
designer. In legal domains, the designer is typically the attorney who prepares an argument
for the court or the judge who writes the court’s opinion for a case. These functions will
correspond to the significant legal issues raised by the case that the court must resolve.
Later, if the body of case law evolves in such a way that the case is used for some other

purpose, then other functions can be added to the case’s description.3

 

3 This would almost certainly be done by a legally-trained system user. not by the system itself. At this
point. no provision has been made for the memory to learn new concepts and classify instances of them.

97

Such a situation occurred in the case of Helvering versus Le Gierse [l94l]4. The
Le Gierse case dealt with a claim that the proceeds from a life insurance policy should be
excluded from the decedent’s estate, as prescribed by law. On this issue, the court held
that the proceeds must be included in the estate, because the insurance policy was
accompanied by an annuity contract that essentially negated the risk inherent in insurance.
The court’s justification thus relied on an inference that, in Le Gierse’s circumstances, true
insurance did not exist. As the case law surrounding issues of insurance grew, this
argument came to be used as a precedent in many cases involving the existence of
insurance. This justification now constitutes one of the more important functions of

Helvering versus Le Gierse in tax law.

Treatment of Multiple Opinions In some situations, different judges hearing a case will
reach different conclusions regarding an issue. Similarly, two judges may reach the same
conclusion but differ significantly in their lines of reasoning. The result is multiple
opinions for the case5. In this representation, each opinion that offers a different line of
reasoning constitutes a separate function of the case. This feature reflects the different roles
a case may play as a precedent in later justifications, based on which opinion is cited. Only
the majority opinion is controlling, in the sense that future decisions are bound by its

reasoning, but the other opinions can often be used in particular contexts as persuasive

 

evidence.
4 All case references appear as a separate bibliographic listing, Appendix A.
5 Cases may contain three types of opinion. The majority opinion presents the court’s holding and its

justification for this conclusion. One or more concurring opinions agree with the majority holding but present
different lines of reasoning for reaching the conclusion. Finally, one or more dissenting opinions may disagree
with the majority’s holding and offer a line of reasoning to justify a different conclusion.

98

 

 

Table 6. Language Grammar: Case Context

<CASE CONTEXT> ::=
Context

Legal Documentation
Plaintiff plaintifiname
Defendant defendant name
Citation official citation
Date date of decision

Procedural Context
Setting court hearing the case
location of decision in appellate chain
Outcome decision
relationship to prior/later cases

Facts
<ASSERTION> {<ASSERTION> ] *

 

5.3.2 Case Context as Device Annotation

The case context description includes background information that characterizes the
situation at issue. This information consists in three parts (Table 6): legal documentation,
the procedural context, and the facts of the situation. Legal documentation essentially
defines the case for reference in the legal literature. Of greater semantic importance is the
procedural context frame. The case setting determines the pedigree of the decision, its
importance as cited evidence in different venues. Naturally, the outcome of the case and its
relationship to other cases in the appellate chain will also have a major impact on when and
how the case is useful as backing.

The facts of the case define the state of the world in which a justification is offered.
These facts may include only those held to be relevant by the court, but more generally they

will include other assertions as well. Facts stipulated by both parties and facts presented by

99

one party as relevant to the case can often provide a more complete context for creating and
evaluating a justification. Table 7 presents a sample case context frame, for the case of'
Humana, Inc., versus the Commissioner of Internal Revenue [1989]. The language for

representing facts is discussed in more detail in the following two sections.

5.3.3 Legal Issues as Function Identifiers

Just as devices in the FR are characterized by their functions, cases in the law are
characterized by the issues they address. Each case addresses at least one legal issue (Table
8). The issue can be expressed as a legal question, raised either by a party to the case or by
the court, to be resolved. For each issue, the court declares a holding, an answer to the
question in the context of relevant case facts. Such a holding will generally be afi'rrmative
(for the plaintiff) or negative (for the defendant). The question and holding define the
endpoint to which the court’s analysis must aim — a justification of the holding with
respect to the question. In this representation, this corresponds to a function of the case.

A function is specified as a precondition/postcondition pair (Table 9). One or more
preconditions serve as the assumptions upon which the justification is based. The
postcondition is the claim made by the court, composed from the legal question and holding
that define the issue. In this sense, the issue identifies a function of the case and
corresponds to the postcondition of the function. The By slot points to a justification (as
described below) that conveys the chain of inference which Concludes the claim Given the

assumptive facts of the case.6

 

6 Other expressions of the Functional Representation [Sembugamoorthy and Chandrasekaran 1986.
Allemang 1990] have also allowed an optional Provided slot. This slot would indicate a state or predicate that
must hold in order for the function to be “meaningful." Such a predicate can be thought of as a special
precondition for the function, not directly involved in the causal behavior realizing the function, that defines a
relevant context for the function. The Provided slot could fruitfully be added to the representation outlined here.
(See the discussion of this topic in Chapter 8.)

100

Table 7. A Sample Case Context Frame — Humana [1989]

Case Humana-89
Context

Legal Documentation
Plaintiff Humana, Inc.
Defendant Commissioner of Internal Revenue
Citation 89-2 USTC ‘19453
Date July 7, 1989

Procedural Context

Setting US. Court of Appeals.
Plaintiff appeals ruling of the Tax
Court that none of its intra-family
insurance premiums are tax-
deductible as business expenses.

Outcome Affirms, reverses, and remands
88 TC 197, in parts.

Facts
Plaintiff unable to obtain liability insurance at fair price.
Plaintiff incorporated wholly-owned insurance subsidiary.
Subsidiary meets all statutory and regulatory requirements.
Subsidiary provides insurance to plaintiff and its affiliates.
Subsidiary was fully capitalized at creation.
Plaintiff deducted all intra-family premiums as

ordinary and necessary business expense.

 

 

Table 8. Language Grammar: Case Issues

<CASE ISSUES> ::= Issues <lSSUE> { <lSSUE> }*
<lSSUE> ::=

Question legal question raised by case
Holding court's answer to the question

<FUNCTION>

 

101

 

 

 

Table 9. Language Grammar: Function

<FUNCTION> ::=
Function identifier
Given <ASSERTION> { <ASSERTION> }*
Conclude <ASSERTION>
B y <JUSTIFICATION>

<ASSERTION> ::= predicate from the domain

 

 

Table 10. A Sample Issues Frame — Humana [1989]

Question Are payments made by an entity to its wholly-owned
subsidiary deductible as insurance premiums?

Holding N 0.

Function Classify Parent Payments

Given Parent pays premium to a subsidiary.
Conclude Payment is not deductible.
B y Conclude No Parent Insurance

Question Are payments made by an entity to a sibling
deductible as insurance premiums?
Holding Yes.
Function Classify Subsidiary Payments
Given Sibling pays premium to a sibling.
Conclude Payment is deductible.
B y Conclude Sibling Insurance

 

102

The assertions that fill the Given and Conclude slots are specified in a state
language that is independent of the FR. Here, assertions consist of domain predicates
taken from the semantics of tax accounting. These predicates may originate in descriptions
of factual situations in the domain or as abstractions of fact patterns. As a result, the
language for expressing assertions in the issue description of a case is also used to express
the facts of the case in its context description. An example of an issue description, again

from the Humana case, appears in Table 10.

5.3.4 J ustifications as Behaviors

Each function points to the justification that supports its claim. The justification
(Table 11) embodies the chain of inference connecting case facts to the holding of the court.
That is, the function specifies a claim made by the case as an input/output relation, and the
justification specifies the reasoning that justifies the claim. In this representation, the
justification also carries a slot for additional commentary, to allow further textual
explanation of the justification. This commentary may refer to case citations not directly
part of the inference chain7 or to comments explaining the justification.

The justification definition corresponds to a behavior in the FR. A behavior is
described as an annotated directed graph with assertions serving as nodes. Each link in the
graph carries an annotation that indicates the warrant for the inference. As described in
Table 11, the graph must constitute a chain, an alternating sequence of assertions and
warrants. In general, though, each step in the inference may draw on additional

preconditions.8

 

7 Such as related or distinguished cases. See discussion of Ashley in Chapter 3.

8 Allemang [1990] has also extended the notion of a behavior to include simple cycles and multiple-link
edges between nodes.

103

 

Table 11. Language Grammar: Justification

<JUSTIFICATION> ::=

Justification identifier
Comment explanatory text
Definition
<ASSERTION> { <WARRANT> <ASSERTION> }*

 

 

 

A warrant carries one of three possible annotations (Table 12). A By Knowledge
warrant refers to knowledge outside of the Functional Representation. In the domain of
taxation, such knowledge could be of an accepted economic definition or of a statute.9 A
By Function warrant cites the function of another case as the backing for an inference.
This is the central feature of the representation for legal analyses described here: the ability
to cite particular issues and justifications from other cases in the domain as backing for
inferences in later cases. Finally, a By Justification warrant indicates that the detailed
inference between two assertions is suppressed in the current graph. The inference is
described in a separate justification. This annotation corresponds to a By Behavior link in
the FR.

Figure 15 depicts a simple example of a justification represented as a behavior. One
additional element of the representation remains, the notion of a hypothetical. A
hypothetical is an argument or justification that invokes a conjecture: what if a certain
circumstance arises? Hypotheticals comprise an important class of arguments in the law,

since they enable exploration of the reasonableness of a claim in potential scenarios. In this

 

9 Of course, such knowledge may itself be describable using the FR. This is especially true of statutes in

taxation law. One might represent a statute as a device, with the function of supporting a particular class of
claims. The BY KNOWLEDGE link could then be replaced with a BY FUNCTION OF STATUTE
link. This possibility is not pursued further here.

104

 

Table 12. Language Grammar: Warrants

<WARRANT> ::=

By Knowledge Of identifier |
By Function identifier] Of Case identifierZ I
By Justification identifier

 

 

 

Justification Conclude No Parent Insurance

Comment: This justification relies on the hypothetical
scenario proposed in Carnation to argue that no parent
can insure with its own subsidiary.

Definition

Parent pays premium to a subsidiary.

By Function Conclude Risk Not Shifted
Of Case Clougherty-87

Risk is not shifted.

By Justification Define Insurance Standard

Payment is not deductible.

Figure 15. A Sample Justification —— Humana [1989]

 

 

105

representation, hypothetical “facts” posed as preconditions in a justification are
distinguished as hypothetical. This distinction alerts users of the justification that this fact

need not hold in a situation in order for the justification as a whole to hold.10

5.3.5 A Complete Example of the Representation

Table 13 outlines the full language for representing a legal analysis. This language
enables one to depict an analysis and its justification in a way that captures domain
relationships among elements of the analysis and among different cases. Consider the case
of Humana, Inc., versus the Commissioner of Internal Revenue [1989]. In Humana, the
US. Court of Appeals had to resolve whether payments Humana and its subsidiaries made
to a wholly-owned insurance subsidiary were legally deductible as insurance premiums.
On the issue of payments made by Humana itself, embodied in the function Classify Parent ‘
Payments, the court justified its holding that the payments were not deductible by citing
both the tax code and a precedent case. These citations provided evidence for particular
steps in the court’s reasoning. Table 14 shows the case and issue descriptions for
Humana, while Figures 15 and 16 present the line of reasoning used in resolving the issue

of parent payments.

 

‘0 The next section includes an example of a hypothetical.

106

 

Table 13. A Language for Representing Legal Analysis

<CASE> ::= Case identifier <CASE CONTEXT > <CASE ISSUES>
<CASE CONTEXT > ::=
Context
Legal Documentation
Plaintiff plaintifl‘name
Defendant defendant name
Citation official citation
Date date of decision
Procedural Context
Setting court hearing the case
location in appellate chain
Outcome decision

relationship to prior/later cases
Facts <ASSERTION> [ <ASSERTION> }*

<CASE ISSUES> ::= Issues <lSSUE> [ <ISSUE>"' }

<lSSUE> ::= Question legal question raised by case
Holding court ’s answer to the question
<FUNCTION>

<FUNCTION> ::= .
Function identifier
Given <ASSERTION> { <ASSERTION> }*
Conclude <ASSERTION>
By <JUSTIFICATION>

<ASSERTION> ::= predicate fiom the domain

<JUSTIFICATION> ::=
Justification identifier
Comment explanatory text
Definition <ASSERTION> {<WARRANT><ASSERTION>}*

<WARRANT> :=
By Knowledge Of identifier |
By Function identifier] Of Case identifierz I
By Justification identifier

107

 

 

Justification Define Insurance Standard

Comment: This justification adopts the definitional test
of Helvering versus Le Gierse [1941] as the legal
standard for the existence of insurance.

Definition

Risk is shifted & Risk is distributed

By Function Define Insurance
Of Case LeGierse-4l

Insurance exists.

By Knowledge Of
Case Data (Payments in question)

Payment is an insurance premium.

By Knowledge Of
Tax Regulation §l.l62-l(a) — 1954

Payment is an ordinary and necessary business expense.

By Knowledge Of
Internal Revenue Code §162(a) — 1954

Payment is deductible.

Figure 16. The Justification Define Insurance Standard
— Humana [1989]

 

 

l ()8

Table 14. A Complete Case Description — Humana [1989]

Case Humana-89

Context

Issues

Legal Documentation

Plaintiff Humana, Inc.

Defendant Commissioner of Internal Revenue
Citation 89-2 USTC 19453

Date July 7, 1989

Procedural Context
Setting US. Court of Appeals.
Plaintiff appeals ruling of the Tax Court that
none of its intra-family insurance premiums
are tax-deductible as business expenses.
Outcome Affirms. reverses and remands
88 TC 197, in parts.

Facts
Plaintiff unable to obtain liability insurance at fair price.
Plaintiff incorporated wholly-owned insurance subsidiary.
Subsidiary meets all statutory and regulatory requirements.
Subsidiary provides insurance to plaintiff and its affiliates.
Subsidiary was fully capitalized at creation.
Plaintiff deducted all intra-family premiums as

ordinary and necessary business expense.

Question Are payments made by an entity to its wholly-owned
subsidiary deductible as insurance premiums?

Holding No.

Function Classify Parent Payments

Given Parent pays premium to a subsidiary.
C onclude Payment is not deductible.
B y Conclude No Parent Insmance

Question Are payments made by an entity to a sibling
deductible as insurance premiums?

Holding Yes.
Function Classify Subsidiary Payments
Given Sibling pays premium to a sibling.

Conclude Payment is deductible.
B y Conclude Sibling Insurance

 

109

In order to examine in greater detail why the court concluded that no insurance
existed, one can refer to the case cited in support of this conclusion, Clougherty versus the
Commissioner of [ntemal Revenue [1981]. A portion of the FR for Clougherty appears in
Figure 17. The Court applied a line of reasoning that proposed a hypothetical scenario:
suppose that the parent suffered an insured loss. Analysis of the resulting reimbursement
indicates that the risk of loss has not been shifted from the “insured” party. According to
the legal precedent of Le Gierse versus Helvering [1941], risk shifting is a necessary
condition for insurance to exist. Hence, the conclusion that true insurance does not exist is
justified by precedent.

Again, if desired, one could examine the backing for this conclusion in greater
detail by referring to the justification offered in Le Gierse. This justification “bottoms out”
in the sense that all of its backings refer to working knowledge of the business domain and
not to prior arguments. The function of Le Gierse cited in Clougherty is an example of a
function ascribed retrospectively to a case based on its use in justification. This function
would likely not have been part of an FR description of Le Gierse in 1941; only subsequent
use of the case as a precedent in insurance cases would lead one to consider defining

insurance to be one of Le Gierse’s functions.

5.4 Conclusion

This chapter presents a functional representation for justificatory analysis,
motivated by the work of Toulmin. In doing so, new emphasis is placed on particular

elements of the influencing representations.

Functional Representation The FR is applied to a new sort of abstract device, a
justificatory analysis. Each such device can be considered as a component of a top-level

device that corresponds to the body of case law as a whole. This corpus of cases consists

llO

 

 

Case Clougherty-87

Function Conclude Risk Not Shifted

Given: Parent pays premium to a subsidiary.
Conclude: No insurance exists.
By : Hypothetical Loss Scenario

Justification Hypothetical Loss Scenario
Comment: This justification adopts the definitional test of Helvering
versus Le Gierse [1941] as the legal standard for the
existence of insurance.
Definition

Parent pays premium
to a subsidiary. Parent sufi’ers a loss.

I l

* By Knowledge Of Case Data (insurance contract)
Subsidiary pays on claim.
¢ By Knowledge Of Accounting

Value of subsidiary's assets falls.

¢ By Knowledge Of Accounting

Value of subsidiary's stock falls.

$ By Knowledge Of Accounting

Value of parent's assets falls.

‘ By Knowledge Of Economics

Parent bears economic burden of loss.

¢ By Knowledge Of Theory of Insurance

Risk is not shifted.

Figure 17. A Portion of the FR for the Clougherty Case

 

 

111

of individual analyses combined to support claims in the domain. The functions of the
device comprise all the functions of its constituent devices. This type of analysis is
possible precisely because cases have roles (functions) in justifying the reasoning in other
cases.

Table 15 shows the mapping between the traditional FR and the FR for
justifications described here. In this use of the FR, a behavior (justification) does not
consist of a causal chain of device states but rather denotes an inferential chain of assertions
about the situation being described. Each level of description — device, function, and
behavior — carries additional commentary that explains the analysis and the context to

which it applies.

Toulmin ’s Model The citation of past cases as backing for a new inference elaborates
Toulmin’s notion of a warrant. Each inference can invoke a prior justification as evidence

to back its conclusion, until reaching the point at which a justification relies on assumed

 

Table 15. The FR for Legal Justifications

The FR for Legal Justifications Traditional FR
Case Device
Issue Component
Function Function
Justification Behavior
Assertion State
Assumptions Preconditions
' Postcondition
Warrant Link annotation
By Knowledge pointers By Knowledge
By Function pointers By Function

By Justification pointers By Behavior

 

112

domain knowledge. By casting the model in computational terms, the representation
described here provides a meaningful way to integrate justifications — according to the
functions they play. Use of the FR also explicates how assertions at different levels of
abstraction relate to one another. J ustifications at different levels can be related, again,
according to the functions they play (that is, according to their pre-lpost-condition
specifications). Table 16 shows the mapping of Toulmin’s terminology to elements of the
new functional representation.

By indexing justifications according to the issues they address, this representation
offers the starting point for a case indexing methodology. The use of case citation links in
justifications comprises the most direct form of indexing, from case to case. The hope
embodied in this representation, though, is that such indexing can be generalized to the

body of case law as a whole, based on relationships among issues in the domain. This

generalization constitutes the topic of the next chapter.

 

Table 16. The FR for Legal Justifications and Toulmin

The FR for Legal Justifications Toulmin’s Terminology
Case < Legal case >
Issue < Legal issue >
Function < None >
Justification Argument or justification
Assertion Assertion

Assumptions Data

Claim Claim
Warrant Warrant

By Knowledge pointers Backing

By Function pointers < None >

By Justification pointers < None >

 

 

 

CHAPTER 6

A CONCEPTUAL MEMORY or JUSTIFICATIONS

6.1 Introduction

In Chapter 4, a problem—solving architecture for justificatory legal analysis was
presented. One component of this architecture, the case memory, provides access to past
justifications for use in the course of constructing new classifications. The remaining
elements of the architecture specify the functional requirements placed on case memory.
Chapter 5 offers a functional representation for justifications that creates the possibility of
conceptual retrieval based on the roles that justifications can play in problem solving. By
applying the principles behind the Functional Representation, one can conceive of an
organization of case memory that promotes focused access to cases based on the inferences
that they make and support.

This chapter proposes a model of conceptual retrieval that is motivated by the
functional representation of justifications. The model capitalizes on the idea of a functional
decomposition to organize the memory of cases according to the issues for which they are
relevant in justifying future cases. The remainder of the chapter describes:

- the index vocabulary for the model,

0 a particular index organization, called an issue composition

hierarchy, that relates functional indices and partitions case memory,
and

0 the retrieval algorithm by which the memory is searched for relevant

C8868.

113

114

 

Table 17. A Language for Representing Legal Analysis

<CASE> ::= Case identifier <CASE CONTEXT> <CASE ISSUES>
<CASE CONTEXT > ::=
Context
Legal Documentation
Plaintiff plaintijfname
Defendant defendantname
Citation official citation
Date date of decision
Procedural Context
Setting court hearing the case
location in appellate chain
Outcome decision

relationship to prior/later cases
Facts <ASSERTION> { <ASSERTION> }*

<CASEISSUES> ::= Issues <lSSUE> { <lSSUE> }*

<lSSUE> ::= Question legal question raised by case

Holding court’s answer to the question
<FUNCTION>

<FUNCTION> ::=
Function identifier

Given <ASSERTION> { <ASSERTION> }*
Conclude <ASSERTION>
By <JUSTIFICATION>

<ASSERTION> ::= predicate from the domain

<JUSTIFICATION> ::=
Justification identifier
Comment explanatory text
Definition <ASSERTION> { <WARRANT> <ASSERTION> }*

<WARRANT> := By Knowledge Of identifier I

By Function identifier] Of Case identifierz I
By Justification identifier

 

 

115

The last of these sections also addresses the notion of how the case retrieval algorithm can
be adapted to provide automatic indexing of cases into the issue composition hierarchy as

cases enter the conceptual memory.

6.2 Index Vocabulary

An index vocabulary consists of those terms from which the user may construct
queries to the memory. In traditional key-word systems, this vocabulary is defined as the
set of all significant words appearing in a stored document. These terms are then combined
using Boolean and adjacency operators to identify syntactic patterns of words in
documents. Thus, the nature of the index terms, and the queries formed from them, is
based directly on how items in the data base are stored — as full-text documents in natural
language.

In one sense, the approach to conceptual retrieval proposed herein applies a similar
notion: the index vocabulary arises from case representation. Cases in memory are
represented using the functional notation described in Chapter 5 (Table 17). This notation
serves as the source of index terms by explicating particular conceptual roles for
justification statements to fill — most notably, warrants and claims. In capturing these
basic elements of justificatory reasoning, such a memory can provide responses to queries
generated in the course of problem-solving activity, queries that deal with particular steps in
the process of justification.

In this respect, however, the index vocabulary proposed here differs from the
general-purpose lexicons such as KWIC. The problem-solving architecture in which the
conceptual memory resides places particular functional requirements on the content of its
performance (Figure 18). Queries are sent to the memory in the course of problem solving.
These requests may be for specific cases, or they may be for any cases that can play a

certain role in assembling a justification. In order to fill these requests, the memory may

conga—mama: 2&0..— .8 2808392. wEZom EuBEm 2:. .E 9:59.—

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

..Suahma< Bum 3.3 3.5 gun—Em
macuoebmne.
Sam
: 5:032 E Loafing—U
oceanogm 302853: ' flown—
. . .5303
. . 3am «30
. . a sewage
_ 8:822 _ 3:830 8385“? 8am
3.82:5 302832: .aoam .8982
m
mcouombmn< 8a”— mcoaogmn< 8am
a8". SE83 .803 .coEomm/x
i Bewfioo '
gonna—mums». 26:3— Bahama
“an a: a8
Sumo—human
3.3.2—ow :euaouumsa

5.8592 3.50

 

117

refer to the facts of the current situation. This architecture, together with the functional
representation for justifications, provides strong guidance in identifying an appropriate
index vocabulary for the conceptual memory. The memory’s task is focused, which
focuses the types of queries to which it must respond.

This chapter describes the domain knowledge and problem-solving method that
comprise the case memory black box in Figure 18. This description begins with a
specification of the kinds of query that the memory can answer. Queries to the case
memory can be of two typesl:

° a request for a particular case and function, by citation, or

0 a request for any cases that addresses a particular issue.

Together, these can be viewed as requests based on a specific sort of abstract feature: the
ability to help justify a particular assertion. Each of these two types of query is
characterized by a particular set of index terms that shapes the search space of cases.
Furthermore, each type of index can be traced back to a particular element of the functional
case representation. This synergy between representation and vocabulary is addressed

further below.

6.2.] Index Terms for Case Citation Queries
The simplest form of query requests a particular case. The case representation
supports a direct form of case-to-case retrieval through the explicit use of case citations as

warrants. In the course of problem solving, the justification generator may wish to obtain

 

1 Chapter 4 also mentions a third type of query to conceptual memory — a request for cases that share a set
of surface features with the current fact situation. Such “surface features” consist of the facts and fact abstractions
that characterize the situation. Research on retrieval of this type has been conducted in several disciplines (see
Chapter 2). Meldman [1977]. N1ch [1977], Hafner [1981, 1987], and Ashley [1990] have all offered potential
approaches to this problem. Retrieval based on surface features is a general-purpose retrieval strategy, not tied in
any direct way to a particular problem-solving task or method. Since this thesis is concerned with conceptual
retrieval for a task-specific problem solver, retrieval based on surface features will be deferred until it can be
discussed in the context of taskodirected retrieval.

118

more detail about the justification for an assertion in an already—retrieved case. This need
corresponds to the expansion of a By Function of Case... warrant. Queries of this
type involve requesting the case cited in the warrant, and in particular the function
responsible for the inference. The index vocabulary for such queries consists of the set of

all case identifiers in the memory.2

6.2.2 Index Terms for Justification Queries

The second kind of query focuses on a specific assertion to be supported. In order
to justify the assertion, the justification generator can request precedent cases that have
justified the same assertion (or a similar one). This sort of retrieval comprises the main
contribution of representing cases in the FR — the ability to retrieve cases based on the
justifications they support. Assertions of this type appear as postconditions of functions,
in the By slot of a case’s Function frame. The index vocabulary for issue justification
queries consists of assertions to be justified, in the form of domain predicates.

Additionally, the justification generator may desire precedents that justify the
assertion from one or more specific facts as assumptions. In this situation, the generator
may also include in its query the facts of the current case (or a subset of them). These facts
may appear as preconditions of a function, in the Given slot of a case’s Function frame,
or as Facts in the Context frame. Like assertions to be justified, these facts are also
predicates from the domain, either facts input to the system or fact abstractions derived in
problem solving. Thus, the full index vocabulary for this class of queries consists of

domain-predicate assertions.

 

2 Note that this is not the same as the full legal citation stored in the Citation slot of the Legal
Documentation frame of the case, but rather is a system-dependent name assigned to the case for internal
reference. In traditional KWIC systems, one can request individual cases by specifying the legal citation in a
query. This is not because such citations are used as case identifiers but because the legal citation appears in the
text of the stored document. Thus, KWIC systems maintain a consistency of index vocabulary by including in the
free-text data base all information that a user is likely to request. However, care must be taken in specifying the
query so as not to retrieve all cases in which the target case is merely mentioned. for instance, as a citation or
sidebar.

119

6.2.3 Summary of Index Vocabulary

These query classes yield two distinct components of the index vocabulary: case
identifiers and assertions about a fact situation. The case memory must be able to retrieve
cases characterized by each of these classes of index terms. Furthermore, retrieval of cases
based on assertions must handle two circumstances: assertions to be justified in the context
of a given fact situation, and assertions that characterize the situation in which an assertion

is to be justified.3

6.3 Index Organization

An index organization specifies the relationship among terms in an index
vocabulary. Such an organization is closely tied to the development of efficient algorithms
for searching the index space. This section describes index organizations of the vocabulary
for case citation and justification queries. The research contribution made here is the
introduction of the issue composition hierarchy. This hierarchy relates assertions based on
their roles as preconditions for justifying other assertions in the domain, thus providing a

mechanism for organizing indices for justification queries.

6.3.1 Index Organization for Case Citation Indices

The index vocabulary for case citation queries consists of case identifiers,
specifically those that appear in the By Function Of Case slot of any
justification warrant. These indices provide for a “direct look-up” of cases and functions.
Cases can be organized in any suitable data base format (for example, a flat table or a B-

tree), and the identifier provided in the query can be used as the key in a standard data base

 

3 One can also conceive of retrieval based on other attributes of the representation. For example, one may
wish to retrieve all cases that cite a particular case. This type of query could be supported by the representation,
through indexing of By Function warrants according the case cited. KWIC systems support such retrieval by
searching for text references to the target case. This and other such queries are not considered further in this work.

120

search. In this situation, the case organization and the retrieval algorithm are trivially
defined: given a case citation index, retrieve the case by look-up.4

Such a direct look-up is one of the central features of the functional representation
for legal justifications: whenever a case is cited directly in support of an assertion, then that
case may be relevant in justifying a similar assertion in a later situation. The FR for cases
makes the retrieval of such a case a trivial matter of table access. Once the case has been
retrieved, the target function can also be retrieved directly, using the function identifier
appearing in the warrant. This kind of direct indexing of behaviors (here, justifications)
according to the functions they denote was one of the central features of the original

Functional Representation for devices.

6.3.2 Index Organization for Justification Indices

Viewing a Body of Case Law as a Whole The cases that make up a body of law — for
example, the body of case law surrounding captive insurance arrangements — comprise a
set of justifications. These cases classify fact patterns as positive or negative instances of
important domain concepts, such as insurance. Each classification, an assertion regarding a
particular fact situation, is supported by a justificatory line of reasoning. In justifying a
classification, a case decision may refer to other issues in the domain as part of its
justification. For example, the courts have found that the presence of risk shifting and risk
distribution can be determinative of the presence of insurance. One can thus support a

claim that insurance is present by showing that risk was both shifted and distributed. In

 

4 One can also conceive of a case citation query that requests a particular case and all cases cited therein.
Such a query might be useful in a situation where the problem solver wishes to adapt all phases of the requested
case‘s justifications to the situation at hand. The case memory would then have to recursively expand all case
citation warrants in the requested case. More generally, though, the problem solver can request the specific case
first and then request each individual expansion as desired. This approach leaves the specification of expansion
knowledge to the external problem solver, making the case memory’s interface with such problem solvers simpler
and more general.

121

this sense, the issue of insurance has been “decomposed” into two subissues, risk shifting

and risk distribution (Figure 19).

 

Insurance exists.

Risk was shifted. Risk was distributed.

Figure 19. Issue Decomposition: Insurance

 

 

 

Cases are characterized by the issues they address. For each such issue, the case is
assigned a function that denotes the line of reasoning that supports the case’s conclusion
regarding the issue. Thus, the issues of the domain can be characterized by the various
case functions that address them. This connection between issues in the domain and
functions in the case representation indicates an extra utility in considering issues as the
defining features of the case law. Not only do issues capture the important classifications
made in the domain, but they also provide direct links to the functions of individual cases.

The body of case law is characterized by the issues addressed in its cases. These

issues are further related along the dimension of composition. If an assertion about issuez
can be used as a precondition for justifying an assertion about issue 1, then a justification of

the former assertion can be used as a component in justifying the latter assertion. In this

case, issuel is composed of issue;. The composition relation is not a necessary one; it
denotes only that issue; may be used in support of issue;. This way of relating assertions

and the issues to which they refer facilitates an organization of the case law that directly

supports justification of assertions, via reference to case functions.

122

The Issue Composition Hierarchy The index vocabulary of justification queries consists

of domain assertions: the memory is given a query in the form of an assertion to be
justified. These indices are related in a directed acyclic graph called an issue
composition hierarchy (ICH). This hierarchy consists of two elements:

0 Nodes correspond to issues from the domain. In the domain of

captive insurance, these might include insurance, risk shifting, and

risk distribution.

0 Edges reflect the composition relation. For nodes N1 and N2, there

exists an edge (N1,N2) if and only if justification of an assertion

regarding N2 can be used as a precondition for justifying an

assertion regarding N1.
The graph is directed and acyclic because legal issue composition is both antisymmetric and
acyclic.5 Such a hierarchy is rooted at the central concept of the domain. In the case of
captive insurance arrangements, this is the concept of insurance. All other issues in the
domain relate back to the concept of insurance via composition links. Figure 20 presents a
partial issue composition hierarchy for the domain of captive insurance.

The ICH may not be a true hierarchy. For instance, in Figure 20, an assertion that
the alleged insurer is a legitimate provider of insurance can be used as data in justifying that
risk was both shifted and distributed.6 Consequently, the node for legitimate provider of
insurance is linked to two parent nodes, risk sharing and risk distribution. This feature of

the ICH reflects the interrelated nature of legal concepts. The definition of the ICH’s edges

 

5 That is, if classification as an instance of one issue is potentially a precondition for classification as an
instance of another, the latter issue cannot also be a precondition of the former. This is true for a single edge
(antisymmetric) or for multiple edges (acyclic).

6 See Sears, Roebuck, and Company versus Commissioner of Internal Revenue [1991].

123

 

Insurance exists.

/\

Risk is shifted. Risk is distributed.

Payment is made
to related company.

Insurer is a legitimate
insurance provider.

 

Payment is made Payment is made
to a subsidiary. to a sibling.

Figure 20. A Sample Issue Composition Hierarchy

 

 

 

also permits the notion of a negative correlation. A justification for the assertion that the
payment is made to a related company will be useful for justifying the assertion that risk is
not shifted. Conversely, a justification for the assertion that the payment is not made to a
related company will be useful for justifying the assertion that risk is shifted. The critical
relationship that defines the ICH is that assertions regarding the child issue may be useful

in justifying assertions about the parent issue.

The ICH as an Index into Case Memory The connection between issues and case

functions makes the ICH an effective and efficient index into case memory. Each issue

124

node contains a pointer to the cases that deal with the issue, in particular, to the case
functions that justify the corresponding assertions. Thus, cases are organized, at the level
of the body of case law, according to their defining features — the functions they play in
justifying domain assertions. Just as individual cases are defined by their functions, the
body of case law is organized according to these functions. In one sense, then, one can
view the body of case law as a device, with its individual cases as component subdevices.
The top-level device has no independent functions except those of its components. The
issue composition hierarchy comprises a device-level index of these constituent functions
(and the corresponding justifications).

However, this single stage of indexing will be insufficient for providing adequate
retrieval of cases for justificatory reasoning. At most nodes in the hierarchy, there will be
several cases that deal with the issue. When constructing a new justification, the problem
solver will certainly desire a finer grain of distinction among the cases. In each case, the
issue will have been decided positively or negatively.7 Further, in each case, the decision
favored either the related group of companies or the Internal Revenue Service. These
features of the case play an important part in the problem solver determining whether or not
the case will be useful for the purpose at hand. For example, a case in which the issue was
decided negatively may be of little use if the problem solver needs to justify a positive
classification.

At this point in the indexing process, the use of case features to guide retrieval also
becomes critical. Retrieval of cases based solely on their feature similarity provides little
direct support for the task of justificatory reasoning. However, in the context of
justification roles, case features themselves can become useful in choosing the most
relevant items from a set of cases judged to be useful for the task at hand. If two cases can

provide assistance in justifying the claim that risk shifting has occurred, but the facts of one

 

7 That is, classification of the case’s fact situation was either justified or unjustified.

125

of the cases more closely matches the current situation’s facts than the other, then the more
similar case will likely be more persuasive in a justification. Thus the salience of a case’s
features is determined only in the context of a particular task and its goals.

For these reasons, the conceptual memory requires a second stage of indexing. At
each node, there exists a fact similarity matcher that maps the facts of the situation against
those of the cases indexed by the node. This matcher segregates cases first according to
issue resolution and case outcome and then according to facts and critical fact abstractions.
The fact similarity matcher at each node provides a function similar to that of the compound
pointer introduced by Hafner [1987].8 This compound pointer consisted of four separate
pointers, based on the cross product of two dimensions, issue resolution and case
outcome. Figure 21 depicts such a pointer, for the issue of risk sharing. In effect, the
structured matcher proposed here incorporates these two dimensions but also takes into ‘
account important fact patterns that characterize the cases. This second stage permits the
case memory to provide even more focused retrieval, within the context of justifying a
particular assertion.

Thus, the two stages of index organization provided by the ICH are:

0 At the level of the body of case law, case functions are segregated

by the issues they address. A case can be retrieved at any node
corresponding to an issue with which the case deals. Nodes are
linked by edges that reflect the composition relation among issues.

0 At the level of the issue node, a fact similarity matcher segregates

cases according (1) to issue and case outcome and (2) characteristic
fact patterns. Each case indexed by the node will appear in exactly

one of the equivalence classes defined by the similarity matcher.

 

8 See Chapter 2 for a discussion of Hafner and her model of legal information retrieval.

126

 

The issue was decided

 

 

Positively Negatively

Risk was shifted, Risk was not
and the case was shifted, yet the case

Related group decided for was decided
Case was the related group. for the related group.

decided
for

Risk was shifted, Risk was not
IRS yet the case was shifted, and the case

decided for was decided

the IRS. for the IRS.

 

 

 

 

Figure 21. The Four Kinds of Issue Node Pointer into Case Memory

 

 

 

6.4 The Case Retrieval Algorithm

Case memory will be searched using one of two algorithms, depending on the
' nature of the query. Case citation queries provide ease identifiers as indices into memory,
and (as introduced above) a standard table look-up or B-tree traversal suffices as the
retrieval algorithm. Justification queries provide assertions as indices into the memory, and
thus they require search of the issue composition hierarchy. This algorithm is outlined in
Table 18.

This “match and decompose” algorithm is a variation of the establishcand-refine
method of hierarchical classification [Bylander and Mittal 1986; Sticklen, Chandrasekaran,
and Josephson 1987]. At each node in the ICH, the algorithm attempts to’match the
assertion against the node’s issue. Three possible values can be returned by the match

process: exact match, subissue match, and no match. Any exact match indicates that the

127

 

Table 18. A Case Retrieval Algorithm for Searching the [CH

GWEN: An assertion to justify

An issue node (initially, the root of the ICH)
1. Match the assertion against the issue at the current node.
2. If there is an exact match: Determine which of the cases at the

cunent node is most similar to the situation at hand. Return the set
of cases so indicated. Halt.

3. If there is no match: Return in failure.
{Otherwise, the match signifies a potential subissue match}

4. For each subnode of the current node, call the case retrieval
algorithm with the subnode.

5. If there are no subnodes, or if all subnodes return in failure: Return
in failure.

 

assertion deals directly with issue at hand. A subissue match indicates that the assertion
deals with a subissue of the current issue. Finally, no match indicates that the assertion has
no semantic connection at all to the current issue and thus has no connection to any of its
subissues.

Any time an assertion exactly matches an issue node, the set of cases pointed by the
node are directly relevant to justifying the assertion. In this situation, the algorithm invokes
the structured matcher that selects among the node’s cases and returns the resulting case
set. Any time there is no match at all, the subgraph of the ICH rooted at the cunent node
can be ignored; none of the issues in that subgraph are related to the assertion. In the
situation of a “partial” subissue match, the node has recognized that the assertion is not
directly related to the issue at the current node but may be related to one of its subnodes.
So the current issue is decomposed into its subissues, and the “match and decompose”
algorithm is called recursively for each subissue node. This recursive decomposition
continues until one of two conditions is met: If an exact match is found, the relevant cases

pointed to by the matching node are returned as the result, and processing halts. If the

128

whole ICH has been explored and no exact matches have been found, then the algorithm
returns no cases and halts.

The idea behind this algorithm is a simple one: the issue composition hierarchy
reflects relationships among issues in the domain. By traversing the ICH in a match-and-
decompose fashion, only those parts of the hierarchy that are potentially relevant to the
target issue are explored. As soon as a node recognizes that the assertion is not relevant to
its issue, exploration of that portion of the hierarchy can be terminated. Eventually, the
assertion is matched against the issue with which it deals, and the cases pointed to by that
issue’s node can be sent to the problem solver.

Only two situations exist in which an assertion will find no match in the hierarchy.
First, the assertion may be a new one in the domain, never having been addressed in a prior
case. This situation, which calls for an extension of the hierarchy, is considered in the
following section. Second, the knowledge necessary for recognizing an exact match is
missing in the appropriate node. This situation is primarily a problem of language. If the
input to the case memory were an assertion in natural language, it is quite plausible that a
query having an exact'match in the ICH could find no match, since the assertion could be
phrased in a novel way. However, in the context of the problem-solving architecture
presented in Figure 18, queries to the case memory consists of formalized domain
predicates handled by the justification generator and situation data base. This limited
representation enables a more complete enumeration of the (kinds of) assertions that are

relevant to any given issue.

6.4.1 Match Knowledge in the ICH
The case retrieval algorithm does not specify the nature of the match knowledge
possessed by the nodes in the hierarchy for determining the issue’s relevance. In

hierarchical classification, establishment knowledge typically consists of compiled fact

129

patterns that characterize the classification category. A structured matcher compares
situation data with these fact patterns, returning a confidence value for the hypothesis that
the situation is an instance of the category. A similar strategy suffices for the match of an
assertion to an issue node in the ICH. Indeed, the ICH node matching problem will often
be simpler than the hierarchical classification scenario. Match knowledge must identify an
assertion as an exact or partial match. Since an assertion consists of a formal domain
predicate, the match process can directly compare the predicate for an exact match. A
subissue match will be indicated whenever the assertion matches an assertion known to
relate to a particular subissue, or when the assertion is a specific instance of a more general
predicate related to a subissue. This latter form of partial match can often be accomplished

by sending a request to the situation data base regarding the predicate in question.

6.4.2 Use of the Retrieval Algorithm for Automatic Indexing

In order for the issue composition hierarchy to grow as new cases are decided,
some mechanism must exist for indexing cases as they are added to the conceptual
memory. The case retrieval algorithm offers a potential mechanism for automatic indexing
of new cases. Table 19 offers an adaptation of the retrieval algorithm aimed at this goal.
Given a case represented in the FR, the indexing algorithm adds the case to the case citation
index (Step 2) and then attempts to add each function of the case to the issue composition
hierarchy (Step 3). To achieve the last step, the match-and-decompose technique of the
case retrieval algorithm is called in an attempt to find the proper node for indexing each
function of the case.

However, this search may fail, for one of two reasons: the proper node does not
have adequate match knowledge to recognize that the assertion justified by the function
should be indexed there, or there is no matching node. In the former situation, some

outside agent — perhaps a human user — must instruct the memory to index the case at the

130

Table 19. A Case Indexing Algorithm

GIVEN: A case represented in the FR

1. Add the case to memory.

2. Add the case identifier, and a pointer to the case, to the case
citation index.

3. For each function of the case:

a. Apply the match-and-decompose algorithm to find an
exact match in the issue composition hierarchy.

b. If an exact match is found, then add the case and function
'to the set of cases indicated by the matching node’s fact
similarity matcher. ,

c. If no exact match is found, then ask the user whether

(1) the function should be stored at an existing node or
(2) thefunction represents a new issue.

1. If the function should be stored at an existing node,
store the case and function as in Step 3b.
ii. If the function represents a new issue, create a new

node as directed by the user, and store the case and
function as in Step 3b.

 

appropriate node. (The agent should then, if possible, also augment that node’s match
knowledge so that the node will recognize the assertion on later attempts.) The latter
situation signifies that the ICH is incomplete, that a new issue decomposition should be
added to the hierarchy. At the node or nodes which were unable to decompose themselves
further, the algorithm can add an appropriate subissue node and index the new case at that

point9

 

9 Another possibility exists: no exact or subissue match succeeds at any node. In this scenario, the
algorithm requires outside assistance in indexing the new case. Furthermore, the outside agent should examine
this case, the ICH, and the match knowledge in the hierarchy to determine the ultimate source of the algorithm‘s
failure.

131

6.5 Conclusion

The functional representation for legal analysis makes possible a memory indexing
strategy that reflects the utility of cases in justificatory reasoning. This chapter presents a
model of conceptual retrieval, based on the FR for cases, for use in a problem solving
architecture for this task. This model — which specifies how cases are requested, how
they are indexed and organized, and how they are retrieved — employs the advantages
inherent in a functional representation for indexing knowledge about abstract devices. In
essence, the model consists of two interrelated indices:

' the case citation index, for when the problem solver knows of a
particular case and function that can help to justify a particular
assertion, and

0 the issue composition hierarchy, for when the problem solver needs
cases that can help to justify an assertion but does not already have a

direct reference to a particular case.

This two-part model relies directly on features of the functional representation of
'cases for its indexing and organization scheme. The use of case citations as warrants for
justifications and the indexing of justifications according to the role they play in the case are
fundamental attributes of this FR, and they capture basic elements of legal reasoning. In
this way, the representation of cases is closely tied to the organization of cases in memory,
and both follow from domain-specific and task-specific understanding of how justifications

are used.

CHAPTER 7
CRISTA: A COMPUTER PROGRAM

FOR CONCEPTUAL RETRIEVAL

7.1 Introduction

The previous three chapters outline the elements Of a legal analysis problem solver.
Chapter 4 presents a problem solving architecture (PSA) for justificatory legal analysis in
terms Of Generic Task problem solvers. In Chapter 5, a functional representation for legal
cases is developed. Finally, Chapter 6 proposes a model of conceptual memory that is
based on the functional representation Of cases. This model is designed to serve as the case
memory component in the problem solving architecture of Chapter 4. These chapters
provide an abstract description Of a case-based justificatory reasoner, with examples taken
from the tax law.

This chapter describes CRISTA, a computer program that implements a portion of
the problem solving architecture for the domain Of captive insurance taxation. In particular,
CRISTA embodies the model Of conceptual memory described in Chapter 6. Cases in
CRISTA’s data base are represented using the FR for cases and are retrieved via an issue
composition hierarchy for the domain. The chapter consists Of two main sections:

0 discussion Of the CRISTA program and its implementation, and

- presentation of two samples problems solved by CRISTA, focusing

on its use of case memory.

132

 

133

7.2 The Implementation of CRISTA

The computer program CRISTA implements a portion Of the problem solving '
architecture outlined in Chapter 4. In order to describe this implementation in more detail,
three issues must be addressed:

° the software environment in which CRISTA was constructed,

0 the top—level organization and algorithm Of CRISTA, and

0 the sub-agents of CRISTA and their implementation.

7.2.1 The Software Environment

CRISTA is specified as an architecture Of Generic Task problem solvers. In
addition to providing an analytic theory Of problem-solving types, the Generic Task
approach [Chandrasekaran 1983, 1987] also Offers a basis for constructing programming
languages that embody these types. Each generic task specifies a particular problem-
solving method. Such a method delineates the types of domain knowledge it requires and
the control strategy for applying this knowledge. As a result, one can construct for each
generic task a programming language that incorporates the abstract method and its control
strategy. This kind of language can be thought of as a programming “shell,” in which the
user enumerates domain knowledge of a particular type, for a particular task.

In the AI/KBS Laboratory at Michigan State University, a software environment
consisting of the Generic Task languages has been developed. The foundation for this
environment is ParcPlace’s Objectworksm, a full-featured implementation of the Object-
oriented language Smalltalk. Smalltalk was selected as the infrastructure for the AI/KBS
environment in large part because of its pure message-passing model Of Object interaction.
Use Of such a model follows natrually from the Generic Task notion Of an intelligent agent
as a community of cooperating knowledge specialists. Furthermore, ObjectworksTM Offers

a number of other features important to a research software environment, including

134

complete compatibility of source code and compiled images across hardware platforms as
well as compatible graphics support
The basis of the AI/KBS Lab’s Generic Task environment is SNOl, a set of
Smalltalk classes that provides basic support for named Objects and Object relations. The
SNO language was implemented largely by Dr. Jon Sticklen, with assistance from the
author and other graduate students in the Lab. SNO ftunishes primitives for the assembly
Of individual Generic Task languages and for the combination Of multiple agents (built from
these languages) into a problem-solving architecture Of the type described in Chapter 4. In
essence, SNO extends ObjectworksTM by adding primitives that support the Generic Task
languages.
The Generic Task software environment consists, at this time, of four generic task
languagesZ:
° HC, for the GT of hierarchical classification [Bylander and Mittal
1986; Sticklen, Chandrasekaran, and Josephson 1987],
0 SM, for the GT Of structured matching [Bylander, Johnson, and
Goel 1991; Sticklen, Chandrasekaran, and Josephson 1987],
0 DS PL, for the GT Of routine design [Brown 1987, Brown and
Chandrasekaran 1986], and
' FM, for the functional representation of devices [Sticklen 1987,
Sembugamoorthy and Chandrasekaran 1986].
The author of this thesis wrote HC and SM. Other graduate students in the Lab, Ahmed
Karnel and Kurt Patzer, were the principal builders Of DSPL and FM, respectively. These

languages now comprise a “tool bench” of GT shells from which one can construct Generic

 

1 For “Spartan Named Object."

2 Each of these languages was originally implemented in Lisp environments at the Laboratory for AI
Research at the Ohio State University. The corresponding languages — CSRL, HYPER, DSPL, and FR — are
described in the cited works.

135

3.3.534‘ 8a...—

aaoaaoé 328 8285 efimau .mu 2.5:.

 

: 8532
33835

_ .8532
3562.5

 

E command—U
52282:

2 Lacuna
382852:

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

E252 080

8am Sun 55255
8385?.
8mm
1' $8”.
.=oEomm<
23% «:5 Al
5 _
I a :85
meouogmne. 8mm
.fioam .coEomme.
3338?. 8am
flown £39.83.
5?: 395.85
5:35 82.8qu0 '
ammo 25$ meouaocugm
5am

136

Task problem solvers in an integrated environment. It is from this tool bench that CRISTA

was built.3 Any extensions to the basic languages are described when appropriate below.

7.2.2 CRISTA: The Top Level

CRISTA was designed and implemented to test the idea Of functional case
representation and the model of conceptual retrieval based on this representation. As such,
it embodies a large portion of the problem-solving architecture for justificatory analysis
outlined (Figure 22). CRIST A’s architecture includes three Of the agents from the legal
analysis architecture: the Fact Abstractor, the Situation Data Base, and the Case Memory.
These agents constitute the basic requirements for conceptual retrieval in the model detailed
in Chapter 6. The Data Base manages all data and data acquisition for the Case Memory.
The Fact Abstractor generates fact abstractions for use by the Case Memory, on request
from the Data Base. Finally, the Case Memory retrieves cases based on their ability to

assist in justifying assertions provided by the user.

 

Table 20. CRISTA’s Control Strategy — Abstract Level

1. Request an assertion to be justified and the initial data that describe
the legal situation.

 

 

2. Identify relevant legal abstractions Of the data.
3. Identify relevant past justifications (cases).
4. GO to Step 1.
3 SNO also provides a primitive form of the inferencing data base described by Mittal. Chandrasekaran.

and Sticklen [1984]. This incomplete version of the intelligent data base generic task is sufficient to implement
the Situation Data Base of the architecture discussed below.

137

The omission of the Justification Generator agent from the full justification
architecture means that CRISTA must have a top-level control strategy for directing its
components in the task Of case retrieval. Indeed, CRISTA’s control strategy (Table 20)
follows that Of the proposed Justification Generator with respect to the retrieving of past
cases. This control strategy is implemented as a sequence of message exchanges among
the agents and system user. Table 21 expands the control strategy to demonstrate the
messages that are exchanged. The reader will note that, while the architecture itself is
fixed4, the number and sequence Of requests made in retrieving cases is dynamic.
Requests are made only when an agent requires particular facts or fact abstractions in the
course of addressing a given legal situation. Thus, the control strategy for CRISTA as a

whole is driven solely by interactions among its individual agents.

7.2.3 CRISTA: The Subagents

CRISTA consists of three subagents. Each Of these subagents embodies a specific
problem-solving method and possesses specific forms Of domain knowledge. This section
describes the domain knowledge and particular implementation details that characterize each

of these subagents: the Situation Data Base, the Fact Abstractor, and the Case Memory.

Situation Data Base The Situation Data Base provides a subset of the capabilities
found in an instance Of the intelligent database generic task (Footnote 2). Its two primary
responsibilities are to manage the situational data that describe a case and to manage all
interaction with the system user. In CRISTA, the Data Base is an instance of the class
AbstractionDataBase, which is a subclass Of the standard SNO data base. This subclass

extends the SNO data base to allow direction of data variables to fact abstraction agents.

 

4 That is, the paths of communication among the agents are fixed at program design time. See Table 3.

138

 

Table 21. CRISTA’s Control Strategy — Message-Passing Level

System user initiates a session with CRISTA by requesting a legal
justification from the Case Memory.

Case Memory requests an assertion from the Situation Data Base,
which forwards the request to the user.

On the first such request, Data Base also asks the user for the initial
data that describe the legal situation.

The assertion is returned to the Case Memory.

In the course of retrieving relevant justifications, the Case Memory
may request relevant legal abstractions Of the case data from the
Situation Data Base.

In the course of filling this request, the Situation Data Base may
request (a) additional case data from the user and (b) new fact
abstractions from the Fact Abstractor.

In the course of identifying relevant legal abstractions, the Fact
Abstractor may request additional case data from the Situation Data
Base.

The Case Memory answers the initial request (Step 1) by sending
relevant past justifications to the system user.

GO to Step 1.

 

 

139

Table 22. Channels of Communication in CRISTA

Fact AbstractOr

Input:

Requested From:

Output:
Requested By:

Case Memory

Input:

Requested From:

Output:
Requested By:

Situation Data Base

Input:

Requested From:

Input:

Requested From:

Output:
Requested By:

 

Assertion, Facts, Fact Abstractions
Situation Data Base

Fact Abstractions
Situation Data Base

Assertion, Facts, Fact Abstractions
Situation Data Base

Past J ustifications
< System User >

Assertion, Facts
< System User >

Fact Abstractions
Fact Abstractor

Assertion, Facts, Fact Abstractions
Fact Abstractor, Case Memory

140

The Data Base maintains two types of data variables, facts and fact abstractions.
For each fact, it has a slot for the fact’s value and a method for Obtaining a value from the
user. For each fact abstraction, it has a slot for the abstraction’s value and a pointer to the
Fact Abstractor agent capable Of determining its value. When the Data Base receives a
request for a fact or abstraction, it first checks to see if the requested item has a value in the
current case. If so, the Data Base returns that value to the requestor. If not, it sends a
request to the appropriate source agent, either a specific fact ahstractor or the system user,
to determine a value. It then stores the value and sends it to the requestor.

The Situation Data Base has one variable for each fact and one variable for each fact
abstraction. The fact abstraction agents are described below. For the domain of captive
insurance taxation, CRISTA maintains approximately thirty fact variables. These are
variables are Of several types:

0 numeric variables, such as PercentOfRevenuesFromOutsideRisks

and NumberOflndependentBuyers,

0 binary variables, such as IsInsurerFullyLicensed, and

0 qualitative variables, such as NatureOfl’remiumTerms.
Several of these fact variables are listed in Table 23. The number of qualitative variables is
smaller than that for the other two types. Typically, qualitative variables require some
domain knowledge in order to abstract a qualitative value from a non-qualitative value.
CRISTA’s few qualitative variables require little judgment (leaving abstraction for the Fact

Abstraction agents), instead calling for “multiple choice” answers from observables.

Fact Abstractor A set Of structured matchers and hierarchical classifiers constitutes the
fact abstraction agent of CRISTA. Each structured matcher generates an individual
abstraction, and each hierarchical classifier generates an abstraction within a generalization

hierarchy. In order to generate an abstraction, each fact ahstractor may request the values

141

Table 23. A Sample Of CRISTA’s Fact Variables

Numeric Variables

PercentOtRevenuesFromOutsideRisks
NumberOflndependentBuyers
PercentOwanershipInterest
PercentParentRisks

PercentSiblingRisks
PercentOfIndependentBuyers
PercentOfRiskReinsuredThroughThirdParty
NumberOfJurisdictionsLicensed
AmountOflnsurersCapital
AmountOfInsurersExposure

Binary Variables

IsInsurerFullyLicensed
IsInsurerRegulated
IsThereACapitalizationAgreement
AreThereOtherRelatedContracts
IsPaymentAFixedExpense
DoesParentDirectlyOwnSubsidiary
DoesSubsidiaryInsureSiblings
IsRiskReinsuredThroughThirdParty

Qualitative Variables

NatureOfPremiumTerms
J urisdictionOfInsurer

 

142

for data variables (either for facts or for other abstractions) from the Situation Data Base.
The structured matchers are written in the language SM, and the hierarchical classifiers are
written in the language HC. These languages directly implement the generic tasks of
structured matching and hierarchical classification as described in Chapter 4.

For the domain of captive insurance taxation, CRISTA’s fact abstractors supply
two kinds of knowledge. First, matchers demonstrate hierarchical relationships among fact
patterns for evaluating the value of a specific abstraction. Second, classifiers exhibit
hierarchical relationships among abstractions in the domain. Each classification specialist
possesses a structured matcher whose task is to determine the value of that specialist’s
abstraction. If this abstraction is present in the case data, then the specialist’s children in
the hierarchy — which represent more specific types of the abstraction — are explored.

CRISTA’s fact ahstractor consists of approximately five classifiers and twelve
matchers. Together, these agents represent approximately fifteen classifier abstractions and
thirty-five matcher abstractions. In total, these agents provide all the necessary fact
abstractions for use in selecting relevant cases from the Case Memory. Figures 23 through
25 depict three of CRISTA’s classifiers, and Figures 26 through 31 present six of
CRISTA’s matchers. These problem solvers are typical of the fact abstraction agents in
CRISTA.

One may note that each of these agents is relatively small in comparison to
classifiers and matchers in other domains (for example, in diagnostic medicine or biological
taxonomy). In such domains, a single classification hierarchy might have dozens of
classification specialists, with each specialist holding a structured matcher consisting of
several simple matchers. The modest size of CRISTA’s agents can be traced to two

causes. Some agents are small because not all of the available domain knowledge has been

143

 

 

Insure Related Party

 

| |
Insure Parent Company Insure Sibling Company

Figure 23. The Related Party Classifier

 

 

 

Captive Insurer

 

| f l
Wholly-Owned Captive Partially-Owned Captive

Figure 24. The Captive Insurer Classifier

 

 

 

 

Insurance Provider

 

Insurance Company Mutual Association

Figure 25. The Insurance Provider Classifier

 

 

 

144

 

 

 

Standard
Insurance
Contract
| l | I |
Standard Standard Standard Standard Standard
Contract Contract Contract Contract Contract

Execution Authorization Modification Performance Renewal

‘ Figure 26. The Standard Contract Matcher

 

 

Diverse Pool of Risks

 

I - 1

Diverse Customer Base Diverse Coverage Base

Figure 27. The Risk Diversity Matcher

 

 

 

 

145

 

 

Insurance Risk Exists

 

| I l

 

Insurer Indemnified Parent Control
by Parent of Funds
I l I
Parent Parent Parent
Guarantee Security Contract
Obligation

Figure 28. The Insurance Risk Matcher

 

 

Arm's Length Transaction

Figure 29.
The Arm's length

Transaction Matcher

 

 

 

 

146

 

Unavailable Commercial Coverage

Figure 30.
The Unavailable
Commercial

Coverage Matcher

 

 

 

 

 

Legitimate Business Purpose

 

l I
Viability Bona F ide Insurer

Capitalization

Figure 31. The Business Purpose Matcher

 

 

147

supplied to them. The Business Purpose matcher (Figure 31) could be decomposed to
include other constituent abstractions from the domain, such as the role of current financial
position in determining the insurer’s viability. This kind of knowledge has been left out of
CRISTA because it was not necessary to demonstrate the case memory facility of the
program. More importantly, though, abstractions in this domain do not seem to consist of
several levels Of subdecisions. Adding more abstraction knowledge to CRISTA would

more likely involve adding new, relatively independent matching and classification agents.

Case Memory CRISTA’s Case Memory embodies the model of conceptual retrieval
described in Chapter 6. The Memory accepts requests of two types. One can request the
case indicated by a particular case identifier, or one can request cases that may help in
justifying a particular assertion. The implemented Case Memory consists of three basic
elements: the body Of cases, the case citation index, and the issue composition hierarchy
for the domain of captive insurance.

The cases themselves are implemented as devices in the language FM. This
language was extended to allow two features of the FR for cases. First, the case
representation supports legal documentation at the level of the case (device), issue
(function), justification (behavior), and warrant (behavior links). Consequently, a subclass
of each of these FM objects was created and augmented with slots to hold the appropriate
documentation notes. Second, any precondition to a justification can be tagged as a
“hypothetical” fact, which indicates that the fact need not be present in order for the
justification to be applicable to a fact situation. An example of a case stored in CRISTA’s
memory appears in Table 24. In total, CRISTA contains 25 cases from the domain of
captive insurance taxation.

The case citation index provides a direct mapping from case identifiers to cases.

This index is implemented as a dictionary in Smalltalk, a data structure that consisting of

148

 

Table 24. A Sample Case from CRISTA — Humana [1989]

Case Humana-89
Context

Legal Documentation

Plaintiff Humana, Inc.

Defendant Commissioner of [ntemal Revenue
Citation 89-2 USTC ‘19453

Date July 7, 1989

Procedural Context
Setting US. Court of Appeals.
Plaintiff appeals ruling of the Tax Court that
none of its intra-family insurance premiums
are tax-deductible as business expenses.
Outcome Affirms, reverses, and remands
88 TC 197, in parts.

Facts
Plaintiff unable to obtain liability insurance at fair price.
Plaintiff incorporated wholly-owned insurance subsidiary.
Subsidiary meets all statutory and regulatory requirements.
Subsidiary provides insurance to plaintiff and its affiliates.
Subsidiary was fully capitalized at creation.
Plaintiff deducted all intra-family premiums as

ordinary and necessary business expense.

Issues

Question Are payments made by an entity to its wholly-owned
subsidiary deductible as insurance premiums?

Holding No.

Function Classify Parent Payments
Given Parent pays premium to a subsidiary.
Conclude Payment is not deductible.
B y Conclude No Parent Insurance

Question Are payments made by an entity to a sibling
deductible as insurance premiums?

Holding Yes.
Function Classify Subsidiary Payments
Given Sibling pays premium to a sibling.

Conclude Payment is deductible.
B y Conclude Sibling Insurance

 

149

ordered key-value pairs. The case identifier serves as a key into the dictionary, which
retrieves the associated value, which in CRISTA is the case itself. All the cases that appear
in CRISTA’s library are listed in Appendix A, along with the case identifiers that act as the
keys in the case citation index.

The issue composition hierarchy organizes cases according to the roles they play in
justifying assertions. In CRISTA, this hierarchy is implemented as a specialization of the
hierarchical classifier provided by the language HC. This specialization embodies two
special features Of the' ICH. First, HC is specialized to incorporate the match-and-
decompose case retrieval method (in lieu of the standard establish-and-refme classification
method) and the adapted case insertion method. Second, each node in the hierarchy
requires two types of match knowledge. The issue specialist must have knowledge for
determining whether an input assertion pertains to the issue. This knowledge corresponds
to the “establish” knowledge of the standard classification specialist. Once the issue
specialist has matched the assertion, it then needs additional knowledge for selecting
relevant cases from the cases to which it points. This knowledge corresponds to the fact
similarity matcher described in Chapter 6 — it selects cases based on the resolution Of the
issue and outcome of the case. Both Of these forms of match knowledge are implemented
in CRISTA as structured matchers, in the language SM.

Since CRISTA operates in the domain of captive insurance, its issue composition
hierarchy captures relationships among assertions in the domain of insurance taxation. In
Figures 32 and 33, the ICH that indexes CRISTA’s Case Memory is given.5 This index
consists of eighteen nodes, which collectively point to the thirty-five functions that
comprise the cases in CRISTA’s memory. The nodes at the top of the hierarchy denote

more general assertions about the domain, and so they point to a larger number of cases

 

5 The issue of risk shifting has been investigated in much greater detail by the courts. and as a result the
portion of the ICH rooted at the node, “Risk is shifted.” is considerably more developed than other parts of the
hierarchy.

150

2 5.: 2282: 82838 8E $555 .S 25:

.3255
3Q 3:3 5
we .0an was .eoufioemma 322:
a 3: 5:55 2:. a m_ .285 2:.

.33 Eozo 883w

 

 

 

ESE—ta a
we: .83.: 2:. .338 can 85:55
. 52.53 85.5.0.5
SaEEwo— a
a: 3:55 2:.
2 use".
5 32550
mug—5:56 60335 .5565 uuabcoo
mm 35E 5: M55— 682”. <

.3538 005.355

151

a 5.: 2282: 8:838 2...: 25min .8 2.8...

Q .fiOflQ—Oh .OMOSQ “Gunman—D .wfln—D—m a Cu .S—Uumnzm a 8
28256 on 28: 2mg: 2 a 238 .2 808.8: 238. 2 808.8:
352: Ea :23: 2:. 22:8 .882: 2:. . . .283
8:88: :8 :o
.888 2:32
:28: 2:.
82853 no omega 2: $82.88 @228 8 .82 amp—Ema
8.: 8228 2838 8a 288 m: 8288: 3:888:

   

.882: :5 :23: 2:. m: 3.82: 2:.

.582: 2:
3:888 8
333:5 a:
382: 2:.

 

.828
a 8.8

 

.326 288sz

152

and case functions. At this point, the ICH is relatively stable; addition of the four cases
most recently decided in the domain to the Case Memory resulted in the addition of only
one new node to the hierarchy. Unless the courts adopt a radically new course in deciding
captive insurance issues, the existing ICH will likely remain as an accurate model of issues
in the domain.

The division between assertions that need to be justified and assertions which can
be abstracted using “compiled” domain knowledge is not always distinct. Several of the
fact abstraction agents in CRISTA — for instance, the matchers for the existence of
insurance risk and legitimate business purpose — are also still the subject of debate in die
domain of captive insurance. For each of these issues, there exists some compiled
knowledge for identifying the particular instances of the abstraction in case data. This
knowledge is incomplete, in the sense that it is inconclusive in some fact situations and still
unsettled for other classes of situations. Consequently, some cases justify assertions
regarding these abstractions as if they were any other issue in the domain. CRISTA
accounts for this duality by including fact abstractors for these concepts as well as nodes
corresponding to the concepts in the ICH. Thus CRISTA is able to rely on compiled
knowledge for recognizing the abstractions when appropriate while retaining recourse to

case-based justifications when they are more appropriate.

7.3 Samples of CRISTA’s Problem Solving

Given this summary description of CRISTA, its knowledge content, and the
pertinent details of its implementation, one can now consider how the program functions in
the context of specific problems. This section follows two traces of CRISTA in action. In
the first, CRISTA interacts with a user who presents a case similar to Humana versus the

Commissioner of Internal Revenue [1989]. This trace demonstrates CRISTA’s behavior in

153

a relatively simple interaction. In the second, CRISTA adds a new case to its case
memory, The Harper Group versus the Commissioner of Internal Revenue [1991].

Note that these samples are not intended to provide an all-inclusive view of
CRISTA’s capabilities. The full specification of these appears in Chapters 4 through 6, in
terms of the Generic Task problem solvers from which CRISTA is built. Additionally,
implementation-specific details for CRISTA appear in the preceding section. Rather, these
samples aim to illustrate CRISTA’s behavior in two typical circumstances involving

conceptual retrieval.

7.3.1 Sample Problem #1

The first sample problem involves a case very similar to Humana (Table 24). It
differs from Humana in the presence of two additional facts: the captive insurer is
incorporated in Bermuda, and it is licensed as an insurer in only three states. These facts
may weaken the plaintiff’s case in that they shed doubt on the belief that the captive insurer
has a legitimate business purpose outside of tax avoidance. The user seeks cases that will
support the claim that insurance exists between the captive subsidiary and its sibling
companies, the other subsidiaries of the captive’s parent. This trace follows the steps of
the control strategy outlined in Table 21. ‘

The user initiates a session with CRISTA (Step 0) by requesting cases that will
support an assertion. The Case Memory forwards this request to the Situation Data Base
(Step 1), which is responsible for interaction with the user. The Data Base requests the
specific assertion, InsuranceExists(Sibling, Captive), and the initial case data. The user
enters the data listed in Table 25 by selecting the appropriate questions from a menu and

then selecting or typing the answers. At this point, the desired assertion is returned to the

Case Memory.

154

 

Table 25. The Facts of Sample Problem 1

ClientRole Plaintiff
DoesParentDirectlyOwnSubsidiary Yes
PercentOwanershipInterest 100
DoesSubsidiaryInsureSiblings Yes
PercentOfRevenuesFromOutsideRisks O
IsInsurerFullyLicensed Yes
IsInsurerRegulated Yes
IsThereACapitalizationAgreement No

N umberOfl urisdictionsLicensed 3

J urisdictionOflnsurer Bermuda

 

Since the query is for a justification, the Memory invokes its match-and-decompose
method for the ICH (Step 2). The assertion, InsuranceExists(Sibling, Captive), matches at
the root node of the hierarchy. The node’s similarity matcher identifies six cases in the _ O
memory in which the decision favored the plaintiff and in which insurance was found to
exist (that is, in which the issue was decided positively). These six cases, listed in Table
26, are made available to the user in a browser window, from which any of the cases’

functional representation may be viewed (Step 3).

 

Table 26. Cases Returned by CRISTA on the First Pass of Sample 1

AMERCO versus the Commissioner of [ntemal Revenue [1991]
Crawford Fitting Company versus the United States [1985]

The Harper Group versus the Commissioner of Internal Revenue [1991]
Humana versus the Commissioner of Internal Revenue [1989]

Sears and Roebuck versus the Commissioner of [ntemal Revenue [1991]
Weber Paper Company versus the United States [1962]

However, in the course of retrieval, the node’s fact similarity matcher further
distinguishes Crawford from the current case, because it does not involve insurance among

siblings. Likewise, it distinguishes Weber, because it involves a mutual insurance

155

arrangement, and AMERCO, Harper, and Sears on the basis of the degree of ownership
and outside risks present in those cases. Each of these distinctions relied on a fact
abstraction. When the similarity matcher came to such an abstraction, it requested a value
for the abstraction from the Situation Data Base. Since these abstractions had not yet been
computed in the current case, the Data Base sent requests to the appropriate fact abstraction
agents seeking their values.

The user sees that only Humana was not significantly distinguished from the
current situation. The selection of Humana as the most similar relevant case is denoted
explicitly in the browser’s presentation of the cases. The differences between the Humana
case and the current case are deemed by the similarity matcher to be less serious than the
differences from the other five cases. The user examines the top-level justification of
Humana with respect to sibling insurance. In the Humana case, the target assertion is
justified by reliance on the claim that the risk of loss was both shifted and distributed in the
Humana fact situation. Thus, the user decides to find cases that will help show the
presence of risk shifting and risk distribution in the current case.

The user asks CRISTA to retrieve cases that will help to justify the assertion, that
risk is shifted in the current situation (Step 1). Invocation of the match-and-decompose
method on this iteration finds that the root node of the ICH returns only a subissue match.
The node Insurance exists cannot exactly match the query RiskShifted(Sibling, Captive)
because the query does not deal directly with the existence of insurance. However, the
node’s match knowledge recognizes the predicate RiskShifted as being relevant to one of
its subissues. So the node decomposes itself and asks its subnodes to try to match the
query. Working from left to right in the hierarchy, the node A related contract exists first
tries to match the query but fails. Then the node Risk is shifted tries and achieves” an exact

match (Step 2).

156

CRISTA reports that five of the previously-retrieved cases (all but Weber) are again
relevant to the target assertion. And, once, again, the Humana case is judged to be the
most similar of the relevant cases. The five cases are displayed in a browser window, with
Humana highlighted as the most similar case (Step 3). Examination of the Humana case’s
justification for the claim of risk shifting reveals that all its preconditions are either facts of
the current case (for example, DoesSubsidiaryInsureSiblings = Yes) or abstractions of
those facts (for example, OwnershipRelation(Insurer,Insured) = None). As a result, this
justification can be applied to the situation at hand. A similar sequence of requests and
retrievals follows the user’s request for cases to help justify the assertion that risk is

distributed in the current situation.

 

Table 27. CRISTA’s Case Indexing Algorithm

GIVEN: A case represented in the FR

1. Add the case to memory.

2. Add the case identifier, and a pointer to the case, to the case citation
index.

3. For each function of the case:

a. Apply the match-and-decompose algorithm to find an exact
match in the issue composition hierarchy.

b. If an exact match is found, then add the case and function
to the set of cases indicated by the matching node’s fact
similarity matcher.

c. If no exact match is found, then ask the user whether (1)

the function should be stored at an existing node or (2)
the function represents a new issue.

1. If the function should be stored at an existing node,
store the case and function as in Step 3b.
ii. If the function represents a new issue, create a new

node as directed by the user, and store the case and
function as in Step 3b.

 

157

7.3.2 Sample Problem #2

In the second sample problem, CRISTA adds a new case to its case memory,
Harper Group versus the Commissioner of [ntemal Revenue [1991]. This trace follows
CRISTA’s case indexing algorithm (Table 27), first introduced in Chapter 6. This
algorithm takes as input a case represented in the FR for cases. Table 28 presents the FR
case description for Harper.

The first two steps of the algorithm are trivial. In CRISTA, whenever a functional
representation of a case is created, this FR object becomes an item in the CRISTA’s
Smalltalk image. Thus, having created the FR for Harper means that the case is now in
memory (Step 1). Adding Harper to the case citation index (Step 2) merely requires that a
new key-value pair be added to the Smalltalk dictionary that implements the index. The key
is Harper-91, the case identifier, and the value is a pointer to the FR object that represents
Harper.

Step 3 of the algorithm involves the addition of Harper’s three functions to the issue
composition hierarchy. For each function, indexing proceeds in a similar fashion. First,
apply the match-and-decompose algorithm to find the appropriate issue node. Each of
Harper’s functions will find an exact match in the ICH — at nodes Insurance exists, Risk
is shifted, and Risk is distributed, respectively. Second, invoke the node’s fact similarity
matcher to determine the category of cases in which Harper most appropriately belongs. If
the pattern matcher finds a match at a high enough level of confidence, index Harper in that
group of cases; otherwise create a new case group and a new fact pattern in the matcher.6
The indexing of Harper at this second stage depends on the current state of CRISTA’s case
memory. Ifeither of two cases — AMERCO [1991] or Sears [1991] — is already a part of

the case memory, then Harper’s functions are placed in the same group of cases. If

 

6 The creation of an appropriate fact pattern and the corresponding update to the fact similarity matcher
requires the assistance of a knowledgeable user. The system has no facility for identifying the most predictive or
relevant facts from which to form a “good” patt

158

Table 28. The Case Description for Harper [1991]

Case Harper-91
Context

Legal Documentation
Plaintiff The Harper Group and includible subsidiaries

Defendant Commissioner of Internal Revenue
Citation 96 TC No.4
Date January 24, 1991

Procedural Context
Setting U.S. Tax Court
Plaintiff challenges deficiencies assessed on
payments to a wholly-owned captive insurer.
Outcome In favor of plaintiff.

Facts
Plaintiff owns insurer indirectly through subsidiaries.
Subsidiary meets all licensing and regulatory requirements.
Subsidiary provides insurance to plaintiff and its affiliates.
Subsidiary provides insurance to tmrelated parties.
Subsidiary provides variety of insurance coverages.
Subsidiary was fully capitalized at creation.
Premium rates determined by industry-standard pricing.
Percentage of revenue from unrelated companies = 30%.
Percentage of revenue from sibling companies = 25%.

Issues

Question Do payments made by an entity and its subsidiaries
to a wholly-owned subsidiary constitute deductible

insurance premiums?

Holding Yes.
Function Determine Insurance

Given Risk is shifted 8r. Risk is distributed.

C onclude Payment constitutes insurance.

B y Determine Insurance Justification
Function Determine Risk Shifting

Given Subsidiary was fully capitalized at creation...

Conclude Risk is shifted.

By Determine Risk Shifting Justification
Function Determine Risk Distributing

Given Percentage of unrelated revenue = 30%.

Conclude Risk is distributed.
By Determine Risk Distributing Justification

159

neither of these cases is already in memory, then a new group of cases is created for
Harper. In the latter case, subsequent addition of AMERCO or Sears to CRISTA will

result in their being indexed into Harper’s case set.

7.4 Conclusion

This chapter describes CRISTA, a computer program that implements the ideas
proposed in the preceding chapters. CRISTA embodies a model of conceptual retrieval
based on the roles that prior cases can play in justifying assertions. The functionality and
content of the case memory are guided by two central ideas: the problem solving context in
which the case memory will be used, and the functional understanding of past justifications
as abstract devices. These ideas have, in turn, greatly affected design decisions made

during the development of CRISTA.

mm
COMPARISONS TO RELATED WORK:

EXTENSIONS AND ELABORATIONS

8.1 Introduction

The preceding four chapters introduce an approach to conceptual retrieval based on
the idea that justifications can be understood as abstract devices. Those chapters propose:

- a problem solving architecture, in the technical sense of Sticklen

[1990], for the task of case-based legal justification,
0 a representation for justifications based on the Functional
Representation of devices [Sembugamoorthy and Chandrasekaran
1986], and
- a model of conceptual retrieval for the task of justification that is
based on this representation.
These proposals are made in the context of two analyses — an investigation of various
approaches to the problem of conceptual retrieval (Chapter 2), and an examination of the
task of justificatory legal analysis in taxation (Chapter 3).

This chapter evaluates the ideas proposed in this thesis with respect to related
research in AI and other disciplines. Particular attention is paid to how this work extends
and elaborates ideas found in existing approaches. The remaining sections of the chapter
are organized according to the two central contributions proposed in this work: a functional

representation of legal justifications, and a task-directed model of conceptual retrieval.

160

161

8.2 A Functional Representation of Legal Justifications

In large part, the use of the Functional Representation to model legal justifications
was motivated by two ideas. First, specialists in the law seem to organize their
understanding of legal cases in a way that relates new justifications to past cases, to the
main uses of cases in precedent-based reasoning, and to the concepts and statutes of a
particular domain. This idea was one of the original intuitions underlying the development
of the Functional Representation, in the context of understanding engineering devices
[Sembugamoorthy and Chandrasekaran 1986]. Second, the work of Toulmin [1958] on
the characterization of justificatory arguments bears a strong resemblance to the graphical
representation of behaviors in the FR. This idea provided an explicit bridge to the use of
the FR for understanding legal cases as abstract devices.

The case representation proposed in Chapter 5 raises several issues of interest.
Two such issues are discussed here: the representation of justificatory arguments, and the

issue of the appropriate level of granularity in case representation.
8.2.1 Representation of J ustificatory Arguments

Warrants With the syntactic extensions to the FR outlined in Chapter 7, the
representation of legal cases follows closely the principles that underlie the FR. The
distinguishing feature of the case representation is its use of behavior links to embody
justificatory warrants.

The idea of warrants is also central to many other approaches for representing
arguments. Toulmin [1958] first sought to explicitly identify warrants and the other roles
that an assertion can play in justification. Other researchers adopted his terminology and
sought to identify classes of warrants that operate in various forms of argument. Birnbaum

[1982] catalogued a small number of “argument molecules” that incorporated warrants for

162

particular beliefs in an adversarial argument. Branting [1989, 1991] has used his notion of
complementary warrantsl to gain leverage in automating justificatory legal analysis. Other
researchers who have focused on the explicit representation of warrants in the law include
Marshall [1989], Lutomski [1989], and Dick [1987].

Toward this end, the functional representation of cases offers a principled way to
classify justificatory warrants in the legal domains. The three types of link pointer in the
case representation are:

0 By Knowledge of <some world knowledge>,

0 By Justification <justification identifier), and

0 By Function <function identifier) of Case <case identifier>.

The first type allows reference to various forms of knowledge about the world. In the law,
this might refer to a statute or regulation, a generalized legal rule, or a piece of common-
Sense knowledge outside the law. The second type allows reference to another justification
in the case. Such indirection makes it possible to hide the technicalities of a justification
until such time as one wishes to examine the warrant in greater detail. Finally, the third
type allows direct citation of another case, and in particular direct citation of a specific line
of reasoning adopted in the case.

These three types of warrant encompass those proposed by previous researchers.
Birnbaum’s argument molecules corresponded to By Knowledge pointers in the FR.
While Branting recognized the need for explicit case citation in the law, his theory of
complementary warrants does not differentiate among the various justificatory roles that a
single case might play. Thus there is no precise sense of a case’s different issues or lines

of reasoning, and thus no sense of case’s function.

 

1 Legal rules and legal cases are complementary in the sense that they can support each other in such a way
that assertions not justifiable using only one type of warrant are justifiable using both in concert. More
generally. Branting’s complementary warrants denote compiled classification knowledge and episodic
recognition knowledge.

163

Indeed, the richness of the FR’S warrants affords further latitude in identifying
meaningful citation types. One can conceive of decomposing the class of By Knowledge
using other forms of legal knowledge. For instance, pointers to statutes could be
segregated into a class of By Statute links. This link type would allow the user of the
represented case to make direct use of the case citation in retrieving the statute from a data
base. Furthermore, such a link would also facilitate access to a structured representation of '
the statute (perhaps using an FR-like formalism), thus enabling the representation and use

of a case’s reasoning in even greater detail.

Function Specification In the FR for cases, functions are specified as a ternary relation:

Given assumptions
Conclude assertion
B y justification

But in the original Functional Representation [Sembugamoorthy and Chandrasekaran
1986], functions were defined as a four-tuple:

If preconditions

To Make goals

Provided meaningful context (optional)

B y causal behavior
The Given/Conclude/By triple corresponds to IflToMake/By in the standard FR. But
what of the Provided clause? Roughly, this clause specifies conditions on the state of the
device that must be true in order for performance of the function to be meaningful. That is,
it specifies a set of “semantic” preconditions. These conditions are not strictly necessary
for the causal behavior to be performed, but they are necessary for the achieved state to be
meaningful in the domain.

This clause could be used to great effect in representing justifications in the law and

many other domains. Typically, a justification refers only to those assumptions that are

salient to the issue in dispute. Yet any given justification will be meaningful only in a

164

particular context. Amertions so specified would serve as super-preconditions, in that they
would have to hold in order for the justification to be persuasive. For example, in the case
Humana versus the Commissioner of [ntemal Revenue [1989], the function Conclude
Insurance might be defined as follows:

Function Conclude Insurance

Given Risk is shifted AND Risk is distributed
Conclude Insurance exists
B y Conclude Insurance Behavior

There are, however, a number of other assertions that may have to hold in order for this
justification to apply. The recipient of the risk may, in a tax domain, have to be a licensed
and regulated insurer.

Whether a necessary assertion is classified as an assumption or as part of the
semantic context of the case depends largely on the domain. In the law, this classification
will also depend on the particular state of the domain, that is, the existing body of case law.
As case law evolves, different case features become more or less salient and more or less
the subject of adversarial debate. This shift is a necessary property of the law. Features
that were once considered contextual are challenged, thus becoming part of the argument
discourse as explicit assumptions (e.g., sibling companies cannot purchase insurance from
one another). Conversely, features once considered part of the argument discourse as
explicit assumptions become settled matters of law and thus part of the legal context (e. g.,
the features of risk shifting). An FR for justifications must ultimately account for this
interplay, perhaps treating preconditions and contextual states identically, except for
providing notational distinctions that benefit human users. This remains an interesting

problem for future research.

165

8.2.2 Relationship to the Issue of Case Granularity

An important open question in the field of case-based reasoning involves '
determining the appropriate granularity for representing cases in memory. Past case-based
systems have demonstrated the difficulties to be found in representing complete problem-
solving episodes, from the initial problem to the computed solution, as individual cases.
Often, only a small fragment of the problem solving done on the case will be relevant in a
later situation. Representing the whole case as a single item in memory means that the
whole case will have to be retrieved in the later situation, with the related computational
cost of fmding the appropriate fragment for application in the new situation. This approach
also leads to difficulties in generalizing from comparable experiences that constitute
elements of individual cases.

In order to address this problem of large case granularity, several researchers have
investigated ways of decomposing cases into more appropriately-sized fragments. Sycara
and Navinchandra [1989, 1991] approached this problem as one generating indices that
point directly to a particular fragment within a unitary case. Redmond [1989, I990]
attacked the problem more directly, proposing a theory of case decomposition based on
problem-solving goals for case-based diagnosis. Each resulting case fragment, termed a
“snippet,” carries knowledge of its problem-solving context, its goal, and the way the goal
was achieved. These snippets are then indexed individually for retrieval when they will be
most useful.

The functional representation of cases proposed here offers many of the same
advantages of Redmond’s snippets. Like Redmond’s approach, the case FR allows direct
access to case fragments — in the FR, functions and behaviors — based on the goals they
achieve. Furthermore, the FR also facilitates generalization based on the goals that each
case function achieves. Cases are accessible as a whole, but they can also be thought of as

distributed throughout memory and accessible as functional fragments. The FR provides

166

some distinct advantages over Redmond’s approach, though. Explicit use of the known
functions of a device focuses the selection of appropriate case fragments. One would not
wish to create a fragment for each assertion in the justification, but the functional approach
eliminates this possibility, instead focusing on fragments that achieve particular design
goals for the device. In physical domains such as Redmond’s automotive trouble shooting,
the FR provides the additional benefit of allowing model—based simulation of the device.
This capability enables the use of mechanisms such as Sticklen’s [1987] compilation of

classificatory knowledge through model-based consequence finding.

8.2.3 Potential Impact on Legal Practice

Finally, the functional representation for cases could have a positive impact on legal
practice. The FR offers a systematic means to critique a justification. By implementing
Toulmin’s role-based model of argument in a computational manner, the representation
makes explicit commitments to the roles that assertions play in a justification. This
facilitates annotation of the case’s reasoning at the level of the case itself, or at the level of
individual issues, justifications, and warrants. In the situatiOn of a court deciding a case,
the majority opinion of the court would be issued as the court’s holding. Judges wishing
to author concurring or dissenting opinions would have two options available. They could
issue a functional representation of their reasoning, but they could also generate specific
criticisms of the majority opinion by annotating the court’s FR of that opinion. Indeed, a
similar notion has been described by Lowe [1985] in the context of a user-organized

information system.

8.3 Conceptual Retrieval based on Problem-Solving Roles

In Chapter 2, the problem of conceptual retrieval was analyzed, and past research

aimed at understanding and solving the problem was evaluated. One of the central goals of

167

the research described in this dissertation is to propose and evaluate a model of conceptual
retrieval based on problem-solving roles. Chapter 6 proposes such a model, in particular
focusing on the issue composition hierarchy (ICH). In this section, the model is appraised

with respect to the important issues that it addresses.

8.3.1 The Use of Abstract Indices

Underlying most work on conceptual retrieval is the notion of an abstract index.
Such an index corresponds to some abstract feature that characterizes a case. This feature
is predictive of the case’s future utility; that is, the case will likely be useful in future
situations characterized by the same feature. Typically, in case-based reasoning, abstract
indices have consisted of “task-related goals and features of the world causally relevant to
the status of these goals” [Domeshek-1991a]. Much of the research on abstract indexing of
case memory has been aimed at specifying more concretely the nature and content of these
two index types.

The model of conceptual retrieval implemented by CRISTA offers one approach:
indices consist of assertions to be justified. Cases are organized across these indices based
on their functions, their abilities to justify assertions. This approach supports Hammond’s
[1989] claim that abstract indices — though often objected to because they require
computational resources to generate them —— may already be present in the vocabulary of
the problem-solving agent CRISTA is designed as a component of a larger problem-
solving architecture for legal justification. This problem solver places particular functional
demands on the performance of case memory, including the need to access cases by their
functions. No additional computation is expended in generating these indices, since they

correspond directly to necessary elements of the problem solver’s vocabulary.

168

8.3.2 Relationship to Goel and Hafner
The two primary motivations for the model of conceptual retrieval proposed here
were Hafner’s [1987] issue/case discrimination tree and Goel’s [1989] use of functions in

the FR as indices into a memory of design cases.

Goel and Functional Indices Goel indexed memory by the functions that each design
case could deliver. In this way, cases could be retrieved when they would be most useful
in generating a new design — when they were capable of fulfilling a design need.
Retrieval was accomplished by matching the functional specifications of a desired design
against the functional specifications of all cases in memory. The case that matched most
closely was then adapted to the current need. If multiple cases matched partially, then the
problem solver applied deliberative design knowledge in ordering the cases for adaptation.
The matching process itself was what Goel called “intuitive”, an associative, qualitative
match against every case in the memory.

While this approach outlines an index vocabulary, it did not specify an index
organization or an associated retrieval algorithm sufficient for handling large case
memories. The computational cost of matching a structured design specification against
that of every case in a large memory would soon become unacceptable. Goel suggested
two possible approaches to the organization/retrieval issue. A component hierarchy of
designs would capture relationships among devices and their components. A generalization
hierarchy of designs would organize cases in terms of the specificity of case features along
multiple dimensions of domain interest.

CRISTA’S index vocabulary was motivated by Goel’s use of design knowledge
about devices. Once the notion of representing justifications as abstract devices developed,
through consideration of Toulmin’s work, the notion of using device functions — the

ability to justify assertions — developed naturally. However, the tax law domain in which

169

CRISTA was to operate was far too large to consider the use'of a total-memory retrieval
algorithm. This fact led to the integration of an existing idea to make conceptual retrieval a

reasonable goal.

Hafner and Issue Discrimination CRISTA’s index organization was influenced most
directly by Hafner’s [I987] issue/case discrimination tree (Section 2.3.1). This mechanism
related important issues in a legal domain (broker liability) according to the influence that
one issue had on another in the outcome of a case. Cases in memory were organized
according to the issues that characterized them and according to how these issue related to
one another. Hafner’s model was intended as an information retrieval mechanism, not as a
component in a larger problem-solving system. As such, there was no retrieval algorithm

to speak of, since retrieval was driven by a human user.

A Synthesis The model of conceptual retrieval proposed here merges these two lines of
research. Given an index vocabulary based on case functions, the model required a means
for organizing case memory in order to provide efficient and task-directed case retrieval.
The idea of the issue/case discrimination tree was adapted to take into account the functional
representation of cases. Analysis of the domain was now guided by the consideration of
issues as the targets for justification. Thus, issues in the index organization — now called
an issue composition hierarchy — were related by their specific justificatory relationships
in past cases.

Because this hierarchy also resembled Goel’s design component hierarchy, his
proposed retrieval method — match and refine — was adapted to the ICH, in the form of
the match-and-decompose algorithm of Chapters 6 and 7. This adaptation also incorporates
elements of Hafner’s issue discrimination method, in which issue resolution and case

outcome segregate the cases stored at any issue node.

170

One important difference between the ICH model and those of Hafner and Goel lies
in the details of case retrieval. For Hafner, discrimination between cases at each node'
involves attention only to the contextual factors of issue resolution and case outcome; there
is no consideration of the similarities and differences between fact situations. In CRISTA’s
case retrieval algorithm, these contextual factors are used as the last filter on the retrieved
cases, but only after an intermediate step in which cases are filtered based on their factual
similarity to the situation at hand. This step is necessary to insure that factual similarity —
a crucial element in the persuasiveness of legal arguments — is provided by the case
memory. However, unlike approaches that emphasize factual similarity to the exclusion of
functional utility in argument, this algorithm concentrates on factual similarity only in the
context of task requirements.

For Goel, the retrieval process consists in matching the functional specification of
the desired component with that of each case in memory. In CRISTA, cases are first
matched on their goal assertion, not on their preconditions, and are then matched according
to factual similarity and contextual factors. This omission of the preconditions reflects the
nature of legal argument. As discussed in Section 8.2.1, there is often a large degree of
ambiguity in which facts and assertions constitute the assumptions of a justification and
which constitute domain context. Rather than rely on uncertain assumptions, search of the
[CH considers only the target assertion and the facts of the case. Distinctions based on the
preconditions of an argument can be made by the problem solver.

Ultimately, the proposed model of conceptual retrieval elaborates the ideas of Goel
and Hafner. The model investigates some of the issues left open by Goel with regard to
memory organization and retrieval. Moreover, the model provides a principled means for
generating Hafner’s issue/case discrimination tree. Though the notion of case functionality
is not made explicit, Hafner’s approach implicitly reflects compositional relationships

among domain issues. These relationships deal with the ability of a case to assist in

 

171

justifying similar assertions in a new circumstance. The issue/case discrimination tree does
incorporate more general semantic relationships among its issues nodes than does the ICH,
but this is mostly a result of developing the model outside the context of a specific problem-

solving task.

8.4 Conclusion

This chapter compares the research results reported in Chapters 4 through 7 to
related work in Al and information retrieval. The functional representation of justifications
provides its greatest contribution in elaborating the idea of a warrant for inference. By
extending the notion of warrants to explicit citation of device functions and behaviors, this
representation establishes the basis for a case memory that is organized according to roles
in justification. This FR also offers some answers to questions regarding the most
appropriate granularity of case representation. Finally, one shortcoming in the
representation is discussed: the lack of a function definition slot to hold assertions that are
necessary in order for the justification to be meaningful. This shortcoming points to a
potential avenue for extending this work.

Second, the model of case memory is shown to be a synthesis of ideas from
Hafner’s work on issue/case discrimination trees and Goel’s work on the representation
and indexing of designs in a case memory using a Functional Representation. This
synthesis extends both lines of research toward the idea of conceptual retrieval by
specifying an index vocabulary, index organization, and retrieval algorithm based on the

functional representation of cases.

CHAPTER 9

CONCLUSION

9.1 Introduction

This chapter concludes the dissertation with a discussion of the contributions made
by this work and avenues for future research. The final section of the dissertation offers a

few remarks regarding the ultimate impact of this research.

9.2 Contributions of this Research

In short, the primary goal of this research was to develop a theory of conceptual
retrieval from a problem-solving perspective —— how do capable problem solvers, with their
knowledge of a task and a domain, retrieve cases relevant to a problem-solving situation?
Efforts aimed at answering this question have led to contributions in three broad areas:

0 the modeling of legal justifications as abstract devices,

° the development of a model of conceptual retrieval based on this

view of justifications, and
0 the extension of Generic Task theory.

Contributions in each of these areas are discussed further below.

9.2.] Modeling J ustifications as Abstract Devices
Viewing a justification as an abstract device offers insight into the understanding of
both justifications and the Functional Representation. Toulmin recognized the benefit of

explicating the roles that assertions can play in justificatory argument, and the functional

172

173

representation of cases makes these roles explicit. The FR provides greater benefit,
though, in its treatment of case citation and warrants. Through case citation, justifications
can be represented at multiple levels of detail. This sort of layering is a natural part of
argumentation in many domains, but its nature is made transparent in the functional
representation of arguments. By providing a well-developed set of link types, the FR
offers a rich vocabulary for expressing and manipulating warrants.

Furthermore, the application of the FR to the modeling of legal justifications can
also contribute to the understanding of the Functional Representation itself. Considering
the demands of new types of devices on the FR forces re-exarnination of representational
commitments. Concepts such as hypothetical preconditions can be incorporated into the

language only after determining their meaning within the framework of the FR.

9.2.2 Modeling Conceptual Retrieval Based on Problem-Solving Roles

This research contributes three new ideas in the area of conceptual retrieval. First,
the representation of justifications as devices facilitates the development of a methodology
for indexing and organizing case memory. This methodology relies both on the case
representation and on an understanding of how cases in the domain relate to each other, in
terms of their functionality in justifications. Second, this model of conceptual retrieval is
integrated with a problem solving architecture for generating legal justifications. Indeed,
the model owes much of its analytic power to the fact that it was developed with the
problem solver in mind, allowing the functional demands of the task to shape and guide the
development of the index organization and retrieval algorithm. Third, this theory integrates
indexing on the basis of surface features with indexing based on abstract functional
features. This integration benefits from the utility of feature-based retrieval but only within

the context of providing retrieval in support of a particular problem-solving goal.

174

9.2.3 Extending Generic Task Theory

This research extends Generic Task theory in two ways. The integration of a case
memory with a problem solving architecture of Generic Task agents demonstrates the
natural relationship between the deliberative problem solving of generic tasks, on the one
hand, and the recognition and retrieval of past cases, on the other. By incorporating case
memory with other forms of “compiled” problem-solving knowledge, systems built of
Generic Task agents are capable of providing a wider range of problem-solving behaviors.

The application of Generic Task analysis to the general method of case-based
reasoning begins to illustrate the utility of GT constructs in explaining and building case-
based systems for routine tasks. This research attempts to understand how Generic Task
problem solvers can be used to model the subtasks of case characterization and adaptation.
More to the point, the proposed model of conceptual retrieval relies directly on the problem-
solving roles established by a Generic Task problem solver. The abstract vocabulary of
task-specific problem solving offers clear advantage over ad hoc methods for explaining the

relationship among problem-solving type and memory organization.

9.3 Avenues for Future Research

Even with its diverse contributions, the research reported here raises as many new
questions as it answers. These new questions form the basis of a diverse program of
research aimed at extending contributions in various directions and overcoming some of the
difficulties that arise in applying current results. These avenues for future research are
outlined with respect to the results from which they derive: the functional representation of
cases, the model of conceptual retrieval based on problem-solving roles, and the problem

solving architecture for legal justification.

175

9.3.1 The Functional Representation of Cases

Research into the use of the Functional Representation for modeling justifications ‘
has certainly not been completed here. At least three important issues with regard to the FR
language remain to be addressed. First, this version of FR for cases treats several types of
By Knowledge link in the same manner, when in fact the indicated knowledge differs
significantly across type. In particular, one could imagine representing statutes and tax
regulations using the FR, thus enabling a more meaningful use of such citations. A new
kind of link, By Statute, could be given independent status and used in a way analogous
to the use of the By Function of Case link. It is an open question whether the FR
would provide the same benefits for representing legal statutes as it does for representing
legal justifications. Intuitively, though, similar benefits seem possible.

Second, the idea of negation is poorly developed in this version of the FR. Under
what conditions can one use the negative of an assertion (say, “Risk is not shifted”) as
evidence for justifying the negation of another assertion (say, “Insurance does not exist”)?
The issue composition hierarchy relates assertion types based on their utility for justifying
other types, but this formalism makes no commitments regarding negative states or
negative assertions. Hafner [1981] proposed a particularly well-developed sense of
negation in her Legal lnforrnation Retrieval System. A similar approach applied to the FR
might greatly increase its robustness across different scenarios.

Third, as noted in Chapter 8, the FR for cases does not include a slot for semantic
preconditions of the meaningfulness of functions. The Provided slot available in the
original FR may or may not be useful in a legal domain, since the fluidity of domain
concepts might mean a shift over time in what is considered contextual information and
what is considered assumptive information. An attempt to apply the Provided clause in

such a domain could offer great insight into the nature of legal concepts and their evolution.

176

9.3.2 The Model of Conceptual Retrieval

The methodology for indexing, organizing, and retrieving cases based on the roles
they can play in problem solving is a new one and has only been applied to the abstract
device of justification. One can argue for the generality of such a methodology across tasks
and domains, but only by applying it to various tasks and domains can one gain a true
sense of the power or ineffecriveness of the methodology. Thus, one potential avenue of
research available is the one that has been applied to all the generic tasks: apply the model
under a variety of conditions. These conditions might include application to other domains
in which the FR has proven analytically useful, such as engineering devices, human body
physiology, and ecological biosystems. One particularly important question involves the
scalability of the approach to larger case law domains. In Chapter 7, it was argued that the
issue composition hierarchy of approximately twenty nodes would likely be sufficient to
account for many more cases in the domain of captive insurance. This claim was based on
analysis of twenty-five cases from the domain. Empirical evidence for the scalability of the

ICH methodology will produce a more persuasive claim.

9.3.3 The Problem Solving Architecture

The CRISTA program implemented only three of the four modules in the proposed
Generic Task architecture for legal justification. Many interesting research issues involve
the routine designer for justification assembly. The nature of the knowledge required for
evaluating, critiquing, and modifying a candidate justification was specified in Chapter 4,
but the exact content of this knowledge remains open. Furthermore, testing CRISTA’S
case retrieval performance more carefully would be facilitated by having a predictable and
capable problem solver available for experimentation.

One experiment of particular interest would be to test CRISTA’s case memory

against alternative methodologies. One can conceive of building a copy of Ashley’s HYPO

177

program and measuring its retrieval performance on the cases in CRISTA’s memory.
These measurements could then be compared to CRISTA’s performance on the same case
set. Such an experiment could be run under varying conditions (the density of cases at
each node in the ICH, the number of dimensions available to HYPO, etc.) that would

permit some judgment regarding the programs’ respective strengths and weaknesses.

9.3.4 Practical Matters

In addition to the theoretical issues that surround various elements of this research,
several practical issues also arise. To the extent that the model of conceptual retrieval
proposed here provides access to relevant cases in memory based on problem-solving
goals, the memory should also prove useful as a research tool for human problem solvers.
One practical issue of importance is the question of how best to provide human users with
access to the system. The issue composition hierarchy itself offers a starting point, being a
graphic representation of the issues of the domain and one relevant relation among them.
However, as Kolodner [1991] argues, many ergonomic and interface factors play a role in
determining how useful a computer-aided memory will be for expert problem solvers as
they perform their tasks. Research aimed at discovering the best way to provide such
assistance would add greatly to the practical utility of this model of conceptual retrieval.

Of graver concern is the problem of creating a realistic case memory. For a small
legal domain, such as that of captive insurance, historic cases can be encoded by hand into
the functional representation. And a relatively small number of cases (say, one hundred)
would suffice for many knowledge-based systems applications in narrow domains. But
for more complete coverage of the law — if one wished to extend CRISTA from one
hundred cases to thousands of cases — the approach of hand-encoding cases becomes
untenable. The same problem arises in trying to encode new cases as they are generated by

the courts.

178

Research addressing this problem could take one of two forms. First, one could
provide assistance to the human encoder, whose job is one of abstracting a case into the FR
notation. This form of assistance is already a part of many legal reporting services, such as
Westlaw, in which cases are presented with “headnotes” or other external documentation.
Support for the encoding task might take the form of an interactive but constrained dialogue
with the case memory, which guides the user in filling a“template” for FR notation. The
functional representation is considerably more complex than typical headnotes, though, and
so providing encoding support may require more sophisticated techniques than those
currently available.

Second, and more speculative, is the idea of providing a “sketchy parser” [DeJong
1979] capable of parsing legal cases from full-text into the FR format. Even for restrictive
domains such as captive insurance taxation, this approach is far removed from current
natural-language processing capabilities. A sketchy parser would still require some human
involvement in the encoding process, but it could perform some of the more tedious and
routine parsing tasks. Ultimately, if approaches to conceptual retrieval requiring special-
purpose knowledge representations are ever to become practical, this sort of research must

first be explored.

9.4 Final Discussion

The anival of the Information Age has indeed changed how we approach many
knowledge-intensive tasks. Yet the problem of conceptual retrieval — the problem of
locating in a large data base only those items relevant to a particular semantic topic of
interest — stands as an impediment to the efficient use of large computer memories. This
research has sought to address the conceptual retrieval problem from a task-specific
perspective. The result has been the identification of an indexing methodology that is

intimately tied to a particular problem-solving task and a particular case representation. In

179

essence, this work marries a task-specific theory of problem solving with case-based
reasoning ideas about indexing to achieve a more complete picture of conceptual memory.
The two major results of this work are a functional representation of justifications
and a model of conceptual retrieval based on this representation. With respect to other
related disciplines, the central contribution of this research may lie in its description of how
one can employ knowledge of a device and its teleology in constructing more effective and

efficient case memories.

APPENDIX

APPENDIX A

LEGAL CASE REFERENCES

AMERCO-91. AMERCO and Subsidiaries, Republic Western Insurance Company,
versus Commissioner of [ntemal Revenue (1991), 96 TC 18.

ASMG-85. Anesthesia Service Medical Group, Inc., versus Commissioner of Internal
Revenue (1985), 85 TC 1031

Beach-86. Beach Aircraft Corporation versus United States (1986), 86-2 USTC 9601,
797 F.2d 920 (10th Circuit)

Carnation-81. Carnation Company versus Commissioner of [ntemal Revenue (1981),
81-1 USTC 9263, 640 F.2d 1010 (9th Circuit)

Clougherty-87. Clougherty Packing Company versus Commissioner of Internal
Revenue (1987), 87-1 USTC 9204, 811 F.2d 1297 (9th Circuit)

Consumers-60. Consumers Oil Corporation of Trenton, New Jersey, versus United
States (1960), 188 F.Supp. 796, 61-1 USTC 9124 (DC. New Jersey, 1960)

Crawford-85. Crawford Fitting Company versus United States (1985), 85-1 USTC
9189, 606 F.Supp. 136 (ND. Ohio 1985)

Gulf-87. Gulf Oil Corporation versus Commissioner of Internal Revenue (1987), 89 TC
1010 '

Gulf-90. Gulf Oil Corporation versus Commissioner of [ntemal Revenue (1990), 90-2
USTC 50,496, 914 F.2d 396 (3rd Circuit)

Harper-91. The Harper Group and Includible Subsidiaries versus Commissioner of
Internal Revenue (1991), 96 TC 45

Humana-87. Humana, Inc., versus Commissioner of Internal Revenue (1987), 88 TC
197

Humana-89. Humana, Inc., versus Commissioner of Internal Revenue (1989), 89-2
USTC 9453, 881 F.2d 276 (6th Circuit) ‘

LeGierse-41. Helvering versus Le Gierse (1941), 312 US 531, 41-1 USTC 10,029

Mobil-85. Mobil Oil Corporation versus United States (1985), 85-2 USTC 9585, 8
Cl.Ct. 555

180

181

Moline-43. Moline Properties, Inc. versus Commissioner of [ntemal Revenue (1943),
319 US. 436, 43-1 USTC 9464

Natharbide-48. Commissioner of [ntemal Revenue versus National Carbide
Corporation (1948), 167 F.2d 304, (2nd Circuit)

Natharbide-49. National Carbide Corporation versus Commissioner of [ntemal
Revenue (1949), 336 US. 422, 49-1 USTC 9223

Ocean-91. Ocean Drilling & Exploration Company versus United States (1991), 90-
80T.

Sears-91. Sears, Roebuck, and Co., and Affiliated Corporations versus Commissioner
of [ntemal Revenue (1991), 96 TC 61.

Spring-30. Spring Canyon Coal Co. versus Commissioner of Internal Revenue (1930),
. 43 F.2d 78, 2 USTC 574 (10th Circuit)

StearnS-84. Stearns-Roger Corporation versus United States (1984), 84-1 USTC 9165,
557 F.Supp. 833 (10th Circuit)

Stearns-85. Steams-Roger Corporation versus United States (1985), 85-2 USTC
9712, 774 F.2d 414 (10th Circuit)

Steere-78. Steere Tank Lines, Inc. versus United States (1978), 78-2 USTC 9605, 577
F.2d 279-280 (5th Circuit)

Treganowan-50. Commissioner of Internal Revenue versus Treganowan (1950), 183
F.2d 288, 50-1 USTC 10,770 (2nd Circuit)

Weber-62. Weber Paper Company versus United States (1962), 204 F.Supp. 394, 62-1
USTC 9423 (W.D. Missouri)

BIBLIOGRAPHY

 

 

 

BIBLIOGRAPHY

Allemang, D. (1990) W5. Ph. D., Ohio State University.

Ashley, K. D. (1985). Reasoning by Analogy: A Survey of Selected AI Research with
Implications for Legal Expert Systems. In C. Walter (Eds.,)
LegaLReasQnmgmp. 105-127). St. Paul, Minnesota: West Publishing Co.

Ashley, K. D. (1989). Toward a Computational Theory of Arguing with Precedents:
Aceomodating Multiple Interpretations of Cases. In
,1 (pp. 93- 102). Vancouver, British Columbia:
ACM Press.

Ashley, K. D. (1990).
Hypotheticals. Cambridge, Massachussetts: MIT Press.

Bareiss, R. (1989). 5W. San Diego: Academic Press.

Belew, R. K. (1987). A Connectionist Approach to Conceptual Information Retrieval. In
WW 1 (pp 116-126)

Wm
Boston, Massachussetts: ACM Press.

Bickford, H. C. (1956). MW (3rd ed.). Englewood Cliffs, NJ:
Prentice-Hall Inc.

Bing, J. (1978). Legal Information Retrieval Systems: The Need for and Design of
Extremely Simple Retrieval Strategies. W 1(2), 379-401.

Bing, J. (1987). Designing Text Retrieval Systems for Conceptual Searching. In E1151
1 (pp. 43- 51). Boston,
Massachussetts: ACM Press.

Birnbaum, L. (1982). Argument Molecules. A Functional Representation of Argument
Structure In WWW 1 (PP 83- 85)
Pittsburgh, Pennsylvania: Morgan Kaufmann.

Birnbaum, L. (1989). Panel Discussion on Indexing Vocabulary. In W
Wiring, 1 (pp. 46). Pensacola Beach, Florida: Morgan Kaufmann.

Birnbaum, L. C,ollins, G. Brand, M. ,,Freed M. ,Krulwich, B., & Pryor, L. (1991). A
Model- Based Approach to the Construction of Adaptive Case- Based Planning

Systems InWasnmngflcrkshnn 1 (pp 215- 224) Washington. DC:

Morgan Kaufmann.

182

183

Birnbaum, L. ,,Flowers M, & McGuire, R. (1980). Towards an AI Model of

Argumentafion mamammamw 1(PP 313- 315)
Stanford University. Morgan Kaufmann. '

Blair, D. C., & Maron, M. E. (1985). An Evaluation of Retrieval Effectiveness for a Full-
Text Document Retrieval System. WW 28(3), 289-299.

Branting, L. K. (1989). Re presenting and Reusing Explanations of Legal Precedents. In
' ,1 (PP 103-

 

110). Vancouver, British Columbia: ACM Press.

Branting, L. K. (1991). Exploiting the Complementarity of Rules and Precedents with

Reciprocity and Fairness. In W, 1 (pp. 39- 50).
Washington, DC. Morgan Kaufmann.

Brown, D. C. (1987). Routine Design Problem Solving. In J. Gero (Eds.), 1835.111
WWW Reading, Massachusetts. Addison-Wesley.

Brown, D. C., & Chandrasekaran, B. (1986, July 1986). Knowledge and Control for a
Mechanical Design Expert System. W p.

Buchanan, B. G., & Headrick, T. E. (1970). Some Speculation about Artificial
Intelligence and Legal Reasoning. WM 23(1), 40-62.

Burstein, M. H. (1989). Analogy versus CBR: The Purpose of Mapping. Inflorkshmn
Wm 1 (pp. 133-136). Pensacola Beach, Florida. Morgan
Kaufmann.

Bylander, T.,Johnson, T. R., & Goel, A. (1991). Structured Matching: A Task-Specific
Technique for Making Decisions. WM, 3, 1-20.

Bylander, T., & Mittal, S. (1986, CSRL: A Language for Classificatory Problem Solving
and Uncertainty Handling. Magazine, p. 66-77.

Chandrasekaran, B. (1983). Towards a Taxonomy of Problem Solving Types. A1
Manama“). 9-17.

Chandrasekaran, B. (1987). Towards a Functional Architecture for Intelligence Based on
Generic Information Processing Tasks. In Won
,2 (pp. 1183- 1192). Milan, Italy: Morgan Kaufmann

Publishers.

Chandrasekaran, B. (1990, Design Problem Solving: A Task Analysis. ALMagazing, p.
59-71.

Cross, G. R., & deBessonet, C. G. (1985). Representation of Legal Knowledge for
(30110er81 Retrieval WHO) 35-44

deBessonet, C. (1983). An Automated Intelligent System Based on a Model of a Legal
System. WW 10(1). 31-58.

 

184

DeJong, G. F. (1979). A New Approach to Natural Language Processing. Cognifixe
35212113.. 3(3). 155-170.

Dick, J. P. (1987). Conceptual Retrieval and Case Law. In
911mm

Watering:
W, 1(pp. 106- 115). Boston, Massachussetts: ACM
Press.

Domeshek, E. A. (1991a). Indexing Storiesas Social Advice. In Wan
WW, 1 (pp. 16- 21). Philadelphia: Morgan Kaufmann.

Domeshek, E. A. (1991b). What Abby Cares About. In W
Reasoning, 1 (pp. 13-24). Washington, DC: Morgan Kaufmann.

Duer, W. M.,Horvitz, J. S., & Coberly, J. W. (1988, Captive Insurance Companies:
Deductions for Premium Expenses. W p. 218-232.

Flowers, M. ,McGuire, R., & Birnbaum, L. (1982). Adversary Arguments and the Logic
of Personal Attacks. In W. G. Lehnert & M. H. Ringle (Eds.), W

MW (pp. 275-294). Hillsdale, New Jersey: Lawrence

Erlbaum.

Forbus, K. 1)., & Gentner, D. (1991). Similarity-Based Cognitive Architecture. 5mm '
Bulletin. 2(4). 66-69. '

Gardner, A. v. d. L. (1985). Overview of an Artificial Intelligence Approach to Legal
Reasoning In C Walter (EdS) MW (PP 247-
274). St. Paul, Minnesota: West Publishing Company.

Gardner A v d L (1987) ALWW:
Cambridge: NflT Press.

 

Gentner, D. (1983). Structure-Mapping: ATheoretical Framework for Analogy. Cognitixe
m, 1(2), 155-170.

Gentner, D. (1989). Finding the Needle: Accessing and Reasoning from Prior Cases. In
W ,1 (pp. 137- 143). Pensacola Beach, Florida:
Morgan Kaufmann Publishers.

Gentner, D, & Toupin, C. (1986). Systematicity and Surface Similarity in the
Development of Analogy. CognimSQienQe, LQ, 277- 300.

Gick, M. L., & Holyoak, K. J. (1980). Analogical Problem Solving. Cognitiy;
W12, 306-355.

Goel, A. (1939) ..-° °

 

. A ..Ph D. Ohio State University, Deflpartment of
Computer and Information Science.

Goel, A., & Chandrasekaran, B. (1989). Using Device Models 1n Adaptation of Design
Cases In W 1 (PP 100- 109) Pensacola
Beach, Florida: Morgan Kaufmann.

185

Goel, A. ,Kolodner, J. L. ,,Pearce M. ,Billington, R. & Zimring, C. (1991). Towards a
Case- Based Tool for Aiding Conceptual Design Problem Solving. In

W 1 (pp. 109- 120). Washington, D..:C Morgan Kaufmann.

Golding, A. R, & Rosenbloom, P. S. (1991). Improving Rule- Based Systems through
Case-Based Reasoning InNanpnalLonferenmaAmfieraUmelhgenee 1(PP 22-
27). Anaheim, California: Morgan Kaufmann.

Grady, G., & Patil, R. S. (1987). An Expert System for Screening Employee Pension
Plans for the Internal Revenue Service. In W

Amflsziauntemgeneeanim. (PP 137- 144) Boston: ACM Press.

Greeno, J. G. (1978). Natures of Problem- S-olving Abilities. In W. K. Estes (Eds.,)
(pp. 239- 270). Hillsdale, New
Jersey: Lawrence Erlbaum.

Hafner, C. D. (1981). ' '
. Ann Arbor, Michigan: UMI Research Press.

Hafner, C. D. (1987). Conceptual Organization of Case Law Knowledge Bases. In Eirst
WWW

MW 1 (PP 35- 42) Boston
Massachussetts: ACM Press.

Hammond, K. J. (1986) °
.Ph. D., Yale University, Department of Computer Science.

Hammond, K. J. (1989). On Functionally Motivated Indexing Vocabularies: An Apologia.
In - ' 1 (pp. 52-56). Pensacola Beach, Florida:
Morgan Kaufmann Publishers.

Hellawell, R. (1980). A Computer Program for Legal Planning and Analysis: Taxation of
Stock Redemptions. W 80(7), 1363-1398.

Horty, J. F., Jr. (1962, March 1962). The 'Key Word In Combination' Approach.
WWP 54-64

Kehl, W. B. ,Horty, J. F., Jr. ,,Bacon C., R. T., &Mitchell, D. S. (1961, September
1961). An Information Retrieval Language for Legal Studies. Cnmmnnigatinnm
WP 380- 389

Kolodner, J. L. (1991). Improving Human Decision Making through Case-Based Decision
Aiding- We. 12(2). 52-68.

Koton, P. (1988). Reasoning about Evidence in Causal Explanations. In National
ConfemomAnifieialeemgense 1 (pp 256- 261). St. Paul MN: Morgan

Kaufmann.

Koton, P. A. (1990). A Method for Improving the Efficiency of Model-Based Reasoning
Systems. In W. Horn (Eds.), MW (pp. 273-282). New York:
Hemisphere Publishing Corporation.

186

Krovetz, R. (1985). The Use of Knowledge Representation Formalisms m the Modeling of

Legal Concepts In C Walter (Eds) WM
275- 317). St Paul, MN: West Publishing Company.

Lehnert, W. (1987). Case- Based Problem Solving with a Large Knowledge Base of

learned Cases In W. 1 (PP 256- 261)
Seattle, Washington: Morgan Kaufmann.

Levi, E. H. (1949). WWW. Chicago: University of Chicago

Press.

Lowe, D. G. (1985). Co-operative Structuring of Information: The Representation of
Reasoning and Debate W 23. 97-111

Lutomski, L. S. (1989). The Design of an Attorney's Statistical Consultant. In W
WWW. 1 (PP 224-233)

Vancouver, British Columbia: ACM Press.

MacRae, C. D. (1985). Tax Problem Solving with an If-Then System. In C. Walter

(EdS). WWW (PP 595- 620) St Paul.
Minnesota: West Publishing Company.

Marshall, C. C. (1989). Representing the Structure of a Legal Argument. In mm
1 (PP 121-127).
Vancouver, British Columbia: ACM Press.

Martin, C. (1989). Complex Indices: A Metaphorical Example. MW
,1 (pp. 295- 299). Pensacola Beach, Florida: Morgan Kaufmann
Publishers.

McCarty, L. T. (1977). Reflections on TAXMAN: An Experiment in Artificial Intelligence
and Ilegal Reasoning. BMW. 20. 837-893.

McCarty, L. T. (1983). Intelligent Legal Information Systems: Problems and Prospects.
W. 2(2). 265-294.

McCarty, L. T. (1985). Permissions and Obligations. In C. Walter (Eds.), Computing
WW (pp. 573-594). St. Paul, Minnesota: West Publishing
Company.

McCarty. L. T. (1987). WWW (Technical
Report No. LRP-TR-20). Department of Computer Science, Rutgers University.

McCarty, L. T. ,Sridharan, N. S“: & Sangster, B. C. (1979).]he_imn1;men1flnn_nf

 

Report No. LRP- TR- 2). Laboratory for Computer Science Research, Rutgers
University.

McDougal, T.,Hammond, K., & Seifert, C. (1991). A Functional Perspective on

Reminding. In WW 1 (pp. 63-76). Washington.
DC: Morgan Kaufmann.

187

Meldman, J. A. (1977). A Structural Model for Computer-Aided Legal Analysis. ngm
Warnings 6. 27- 71 ~

Mittal, S.,Chandrasekaran, B., & Sticklen, J. (1984, Patrec: A Knowledge-Directed
Database for a Diagnostic Expert System. W p. 51-58.

O'Brien, & Tung (1973). Captive Off-Shore Insurance Corporations. Won
Wanna. 11. 665.

Owens, C. (1989). Plan Transformations as Abstract Indices. In
Roasoning, 1 (pp. 62- 65). Pensacola Beach, Florida: Morgan Kaufmann
Publishers.

Peterson, M. A, &Waterman, D. A. (1985). An Expert Systems Approach to Evaluating
Product Liability Cases. In C. Walter (Eds),
Roasoning(pp 627- 659). St. Paul, Minnesota: West Publishing Company.

Popp, W. G., & Schlink, B. (1975). Judith, A Computer Program to Advise Lawyers in
Reasoning a Case. lnnmotrios, 15(4), 303-314.

Raphael, B. (1968). SIR: A Computer Program for Semantic Information Retrieval. In M.

Minsky (EdS). Semannflntonnanonlmssmg (PP 33-145) Cambridge
Massachussetts: The MIT Press.

Redmond, M. (1989). Learning from Others' Experience: Creating Cases from Examples.
In W 1 (PP 309- 312) Pensacola Beach.
Florida: Morgan Kaufmann.

Redmond, M. (1990). Distributed Cases for Case-Based Reasoning: Facilitating Use of
Multiple Cases In W. 1 (PP 304-309)

Boston. Morgan Kaufmann.

Riesbeck, C., & Schank, R. (1989). W. Hillsdale, NJ:
Lawrence Erlbaum Associates.

Rissland, E. L. (1985). Argument Moves and Hypotheticals. In C. Walter (Eds.),
Com ' LlegaLReasoning (pp. 129-143). St. Paul, Minnesota: West
Publishing Co.

Rissland, E. L. (1990). Artificial Intelligence and Law: Stepping Stones to a Model of

15881 Reasoning Will. 22(8). 1957-1981.

Rosch, E. H. (1975). Cognitive Representations of Semantic Categories. .[ournaLof
WM 199.. 192- 233.

Rose, D. E., & Belew, R. K. (1989). Legal Information Retrieval: A Hybrid Approach. In
1 (PP 138-

 

 

146). Vancouver, British Columbia: ACM Press.

Schank, R. C. (1982). Remindings and Memory. In W Cambridge,
England: Cambridge University Press.

188

Schlobohm, D. A. (1985). TA— A Prolog Program which Analyzes Income Tax Issues
under Section 318(a) of the Internal Revenue Code. In C. Walter (Eds),
(pp. 765- 815). St. Paul, Minnesota: West-

Publishing Company.

Schlobohm, D. A., & McCarty, L. T. (1989). EPS 11: Estate Planning with Prototypes. In
W

W. (PP 1-
10). Vancouver, British Columbia: ACM Press.

Schlobohm, D. A., & Waterman, D. A. (1987). Explanation for an Expert System that
Performs Estate Planning In WWW

W, (pp. 18- 27). Boston: ACM Press.

Seifert, C. M. (1988). Goals in Reminding. MW, 1
(pp. 357-369). Clearwater Beach, Florida: Morgan Kaufmann.

Seifert, C. M., & Hammond, K. J. (1989). Why There's No Analogical Transfer. In
moon - ' , 1 (pp. 148-152). Pensacola Beach, Florida:
Morgan Kaufmann Publishers.

Seifert, C. M. ,,McKoon G. ,Abelson, R. P., & Ratcliff, R. (1986). Memory Connections
Between Thematically Similar Episodes. Emhologx;

W. 12. 220-231

Sembugamoorthy, V., & Chandrasekaran, B. (1986). Functional Representation of
Devices and Compilation of Diagnostic Problem-Solving Systems. In J. Kolodner

& C Riesbeck (Eds.) WOP 47- 73)
Hillsdale, New Jersey: Lawrence Erlbaum Associates.

Shpilberg, D.,Graham, L. E., & Schatz, H. (1986, July 1986). ExperTax: An Expert
System for Corporate Tax Planning. W p. 136-151.

Simoudis, E., & Miller, J. S. (1991). The Application of CBR to Help Desk Applications.
In WW: 1 (PP 25-36) Washington D C:
Morgan Kaufmann.

 

Sprowl, J. A. (1979). Automating the Legal Reasoning Process: A Computer that Uses
Regulations and Statutes to Draft Legal Documents. WW
mammal. 1912(1). 1- 81

Stanfill, C. (1988). Learning to Read: A Memory-Based Model. In W
W, 1 (pp. 402-413). Clearwater Beach, Florida: Morgan Kaufmann.

Stemberg, R. J. (1977). '
Hillsdale, NJ: Lawrence Erlbaum Associates.

Sticklen. J. (1987) WWW Ph.D.. The Ohio
State University.

 

Sticklen, J. (1990). Problem Solving Architecture at the Knowledge Level. LonmaLof
WWW. 1(3). 1- 52

 

189

Sticklen, J. ,,Chandrasekaran B., & Josephson, J. R. (1987). Modularity of Domain
Knowledge W. I.

Sticklen, J., & Tufankji, R. (1992). Utilizing a Functional Approach for Modeling
Biological Systems W. 1.6. 145-160

Sycara, K. P., & Navinchandra, D. (1989). Index Transformation and Generation for

Case Retrieval. In W 1 (PP- 324-328).

Pensacola Beach, Florida: Morgan Kaufmann.

Sycara, K. P., & Navinchandra, D. (1991). Influences: A Thematic Abstraction for

Creative Multiple Use of Cases. In WW 1 (PP-
133-144). Washington, DC: Morgan Kaufmann.

Thagard, P., & Holyoak, K. J. (1989). Why Indexing 1s the Wrong Way to Think about
Analog Retrieval In MW 1 (PP 36-40)
Pensacola Beach, Florida: Morgan Kaufmann.

Tong, R. M. ,Reid, C. A. ,,Crowe G. J., & Douglas, P. R. (1987). Conceptual Legal
Document Retrieval Using the RUBRIC System. In W

mAmfiQialeclligcnctLandm 1 (PP 28- 34) Boston. Massachussetts: ACM

Press.

Tong, R. M., & Shapiro, D. G. (1985). Experimental Investigations of Uncertainty m a
Rule- Based System for Information Retrieval. MW
Machmefimdres. 22.. 265- 282

Toulmin, S. E. (1958). MW Cambridge: Cambridge University Press.

Walter, C. (Ed.). (1985). W. St. Paul, MN: West
Publishing Company.

Waltz, D. L. (1989). Is Indexing Used for Retrieval? In -
Reasoning, 1 (pp. 41-44). Pensacola Beach, Florida: Morgan Kaufmann.

Waterman, D. A.,Paul, J., & Peterson, M. (1986, October 1986). Expert Systems for
Legal Decision Making. W p. 212-226.

Zarri, G. (P. (1985). Inference Techniques for Intelligent Information Retrieval. In C.

Walter (Eds.). WW (PP. 215-246). St Paul.
MN: West Publishing Company.

Zhang, X., & Waltz, D. L. (1989). Protein Structure Prediction Using Memory- -Based
Reasoning: A Case Study of Data Exploration. In
Reasoning, 1 (pp. 351- 355). Pensacola Beach, Florida: Morgan Kaufmann.