BUCALEGIS.ccj.ufsc.br
CHIRON: Planning in an open-textured domain
KATHRYN E. SANDERS
Department of Mathematics and Computer Science, Rhode Island College, 600 Mount Pleasant
Ave., Providence, RI 02908 USA
E-mail: ksanders@ric.edu
Abstract. Planning problems arise in law when an individual (or corporation) wants to perform a
sequence of actions that raises legal issues. Many lawyers make their living planning transactions,
and a system that helped them to solve these problems would be in demand. The designer of such a
system in a common-law domain must address several difficult issues, including the open-textured
nature of legal rules, the relationship between legal rules and cases, the adversarial nature of the
domain, and the role of argument. In addition, the system’s design is constrained by the fact that the
intended users are lawyers, and its operation and output must be convenient for lawyers to use. In
this article, I describe a system called CHIRON that I have developed to explore solutions to these
issues. This system develops simple plans from representations of statutes and cases in the domain
of United States personal income tax planning.
1. Introduction
The popular view of lawyers is the trial lawyer – Perry Mason, Clarence Darrow,
or the lawyers on LA Law. Many lawyers, however, make their living planning
transactions, such as the sale of a piece of property, the establishment of a trust, or
the reorganization of a corporation.
Planning problems arise in law when an individual (or corporation) wants to
perform a sequence of actions that raises legal issues. For example, suppose you
want to sell your house. That transaction has both real estate and tax aspects. If you
give a lawyer information about your goals – do you want to reinvest the proceeds
in another house? – and facts relevant to the situation – how long have you owned
the house? do you live in it? do you have any other residence? – the lawyer will
suggest a plan or plans for achieving your goals, taking into account the legal issues
raised by the transaction.
In constructing plans, lawyers generally have two types of information to work
with: statutes and cases. In some domains, there are statutory provisions that are
so recent that no cases have been decided interpreting them. In general, however,
both statutes and cases must be taken into account.
In practice, lawyers often take advantage of a third source of information, plans
based on past experience of similar transactions. Often, however, no such plan is
available, perhaps because the lawyer has never performed this particular type of
226 KATHRYN E. SANDERS
§121. Exclusion of gain from sale of principal residence.
(a) Gross income shall not include gain from the sale or exchange of property if,
during the 5-year period ending on the date of the sale or exchange, such property
has been owned and used by the taxpayer as the taxpayer’s principal residence for
periods aggregating 2 years or more.
Figure 1. §121(a) of the Internal Revenue Code.
transaction before, perhaps because the law has changed so recently that no plans
are available. And even where there is such a plan, if it is challenged in court, it
must be justified in terms of the statutes and case law.
In this project, therefore, I examine the way in which plans are developed from
statutes and cases. Specifically, I focus on tax planning. Almost every transaction
has some tax aspect, so tax planning forms part of almost all legal planning; and
the way in which the statutes and cases are used in this area is typical of general
legal planning.
Statutes are rules that have been created formally, by legislation. They are published
by the government and often by private companies as well. For example,
consider §121 of the Internal Revenue Code, which now governs the tax treatment
of the income from the sale of a personal residence, given in Figure 1. The Internal
Revenue Code is the most important statute for United States tax planning.
It contains approximately 7000 sections.
Detailed as the Internal Revenue Code is, it still contains phrases that are not
defined within the statute, for example, the phrase “principal residence” in §121.
To qualify for the benefit of §121, a taxpayer must show, among other things, that
he bought and sold properties which belong to this category. These phrases are
defined partly by commonsense knowledge about the meaning of the words used
and partly by example, not by statute. In determining whether a transaction satisfies
a given statutory predicate, or planning a transaction that will satisfy it, a lawyer
must use both rules and cases.
What makes reasoning about these statutory predicates difficult – and interesting
– is that defining them is not just a matter of inferring defining characteristics
from a set of examples. Generally speaking, there is no set of essential characteristics
shared by all positive instances of the statutory predicate. Some examples are
typical, and others are more or less similar to them along various dimensions. As a
result, classifying a particular object as an instance or noninstance of one of these
categories is not always a simple task.
This issue arises throughout legal reasoning, not just in tax. Indeed, it is part of
a general natural language problem. Many ordinary categories, such as “tiger” or
“cup”, are surprisingly difficult to define. This indeterminacy has been studied in
linguistics and philosophy, where it is labeled open texture (Waismann 1945; Hart
1961; Lakoff 1987). Any planning rule expressed in natural language, such as “be
careful”, “never get involved in a land war in Asia”, or “buy low, sell high”, suffers
from the same problem.
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 227
In any domain, open-textured rules can be partially defined by examples. The
legal domain has the advantage that examples are recorded and published. Each
court case is an example – an application of the law to a particular set of facts. Facts
and results are recorded by the courts in “opinions” and published, both by the
government and by private companies. Thousands of examples are readily available
in any law library.
The facts of these examples can be used as the basis for new plans. Since the
courts are bound by precedent, similar cases must be decided similarly. Thus,
planners attempt to construct plans that are similar (or identical) to examples of
previous successful plans.
In constructing a plan based on previous cases, a planner can make use of the
court’s reasoning. This must be included in the case report along with the facts
and result. In a case interpreting a statute, for example, the court will suggest
intermediate rules connecting the facts of the case to the open-textured statutory
predicates.
These rules have some predictive value. The courts are likely to follow them in
later cases. They are not required to do so, however. They are free to adopt new
rules, as long as the new rules are consistent with the results in all previous cases.
Thus, the courts’ reasoning in previous cases is useful in constructing plans, but
does not make the success of a plan certain.
For example, suppose you want to construct a plan for an academic who wants
to sell a house he has owned for two and a half years, but has just spent a year on
sabbatical away from home. Suppose you have §121 and one case to work with.
Suppose that the earlier case involved the following facts:
John and Jane Smith bought a house in Providence, Rhode Island in June, 1990,
and a second house in Chatham,Massachusetts in April, 1995. On September 3,
1998, they sold the house in Providence and bought a third house in Barrington,
Rhode Island two days later. They are bankers at Fleet. Before the sale, they and
their children lived in the Providence house most of the year; now they live in
the Barrington house most of the time. They stay in the Chatham house (on
Cape Cod) for two weeks during the summer and occasional weekends during
the rest of the year.
Suppose the holding in that case was that the Providence house was their “principal
residence” before the sale, so the income from the sale of the Providence
house qualifies for exclusion under §121, and the court’s reasoning was that if a
taxpayer has more than one residence, his or her “principal residence” is the one
where he or she spends the most time.1
By the reasoning in the Smiths’ case, your client’s house is not his principal
residence, since he has not spent any time there for the past year. On the other hand,
if your case comes to court, you could argue that “principal residence” should be
1 Like the other examples in this article, this one is strictly hypothetical, and not to be relied upon
as tax advice.
228 KATHRYN E. SANDERS
defined as, for example, the residence closest to the taxpayer’s job, where he or she
is registered to vote, and from which his or her tax returns are filed. Moreover, this
rule would be consistent with the facts of the previous case, since the taxpayers
there worked in Providence, not on the Cape. Therefore, a court would be free to
adopt this rule and hold in favor of your client.
The use of open-textured rules and examples has a pervasive effect on legal
reasoning. It makes the legal system flexible. The fact that terms like “principal
residence” are underspecified means that the courts can respond to changing
circumstances. For example, they can interpret “principal residence” to cover cooperatives
and condominiums, even if those forms of ownership did not exist at the
time the statute was passed. Similarly, the First Amendment protection of freedom
of the press can be extended to cover television and radio, as well as newspapers.
On the other hand, because the system is flexible, it is also uncertain. In law, unlike
domains such as chess, it is impossible to prove a plan correct. This uncertainty
is not due to lack of factual information (we can assume complete knowledge of
the facts); but to the underspecified nature of the rules. Given complete information
about the client’s situation and the relevant law, an experienced lawyer can give
an opinion about the probability of success of a given plan, but even the most
conservative plan is not certain to succeed.
Because there is uncertainty, there is room for argument. Lawyers are trained
to find support for different conclusions in a given set of cases. A large part of a
lawyer’s training involves learning to make arguments for and against the application
of some statutory predicate, and for and against the similarity of a previous
case to the current one. This is one of the key legal skills, particularly in the Anglo-
American legal system (Ashley and Rissland 1987; Levi 1949; Llewellyn 1930).
Anglo-American law “requires the presentation of competing examples” (Levi
1949, p. 5). Although one result may be more likely than another, it is generally
possible to make these arguments in both directions.
Ideally, experts in other domains would reach the same conclusion: all doctors
would give the same diagnosis, all engineers would agree on how to build a building
that would be safe in an earthquake, and so forth. In fact, experts in medicine,
engineering, and even mathematics disagree, and find it useful to be able to argue
for and against particular conclusions (see, e.g., Lakatos 1976). The ability to argue
for and against a particular conclusion is useful in many domains. For lawyers it is
central; and as a result, the process is particularly well-illustrated in law.
Tax planning is no exception. Here, the adversaries are the taxpayers and the
government. Taxpayers seek to exclude or defer items of income and deduct expenses,
and the government seeks to include income and disallow deductions. Each
precedent can be viewed as the execution of a plan by the taxpayer in that case,
some successful (the favorable precedents) and some unsuccessful (the unfavorable
ones). Taxpayers must construct plans that are similar to favorable precedents and
different from unfavorable ones in some relevant way. If the similarity is too distant,
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 229
or the differences are too small, the plans will be vulnerable to challenge by the
government.
The open-textured rules and examples interact in interesting ways. Being reminded
of a similar case directs a lawyer’s attention to the rules applied in that
case; being reminded of a potentially applicable rule directs his or her attention
to the cases interpreting that rule. In computational terms, cases limit the search
through the statutory rules. If a lawyer knows of a case involving a plan similar
to his or her client’s, the case report will indicate which rules were applied to that
plan. Similarly, rules limit the search through the case base. If the lawyer knows
what statutory rule he or she is interested in, it is easy to retrieve exactly the cases
interpreting that particular rule.
In this article, I describe CHIRON,2 an open-textured planning system in the domain
of United States individual income tax. CHIRON solves a cluster of problems
having to do with buying, selling, renting, and owning residential housing.
CHIRON is an exploratory system. By this I mean the type of system described
in Hendler (1995):
I discovered that, much to my amazement, many members of the biological
research community spend their time doing what many AI researchers do –
building systems! In fact, I was even more amazed to discover something else
– the systems built by the biochemist are heavily used in experimentation during
the exploratory stages of theory formation! That is, these systems are not
created for running statistical tests to confirm what theorists already believe
(the Baconian ideal of experimentation as it is usually presented to us) but are
used to help develop those theories.
These systems are designed to have certain properties and then tests are run, not
to prove a theory correct, but to explore what happens. Only after exploration are
more precise theories formed and rigorously tested. The success of an exploratory
system is determined, not by whether its output is correct, but by the properties it
incorporates and whether exploration of those properties using this system suggests
interesting directions for future research.
CHIRON was designed to have the following properties:
- it reasons with both cases and rules;
- the rules are extended, limited, and partially defined by the concrete facts of
particular cases;
- it has a detailed representation of case facts, corresponding closely to the facts
given in the case report;
- it distinguishes between conservative and aggressive plans, reflecting the
adversarial nature of the domain; and
- it gives a substantial amount of control to the user, reflecting the expertise of
the system’s hypothetical users.
2 Named for the centaur in Greek mythology, known for giving good advice.
230 KATHRYN E. SANDERS
This list can be interpreted as a partial theory of legal planning, embodied in
CHIRON. CHIRON’s tests were not run to prove or disprove this theory, however, but
rather, as in the biological systems described above, to explore what would happen.
In a young discipline like artificial intelligence and law, open-ended exploration is
particularly important, and many AI and law programs, such as HYPO (Ashley
1991), GREBE (Branting 1990b), and Gardner’s program (Gardner 1987), can also
be considered exploratory programs in this sense.
In Section 2, I discuss CHIRON’s design, implementation, and knowledge representation;
in Section 3, I discuss examples of CHIRON’s execution; in Section 4,
I review the related work; and in Section 5, I summarize the lessons learned from
this project.
2. Design and implementation
2.1. OVERVIEW
An overview of CHIRON’s architecture is given in Figure 2. At the top level, the
system is made up of several modules: the Classifier, the Strategy-Retriever, the
Strategy-Processor, the Combination-Retriever, and the Combiner. The Strategy-
Processor, the heart of the system, is shown in detail.
CHIRON’s input generally includes both goals – the action or actions a client
intends to perform – and some background information. The goals may be missing
if the action has already been performed, for example if the client has already sold
his house. Given this input, the Classifier applies rules to determine what types of
transactions are involved. Transactions are classified according to the type of rights
and property transferred. Transfers can be gifts, sales, rentals or loans; property
can be money, tangible or intangible personal property, real estate, or services. For
more details of the transfer types and property types, see Sanders (1994).
The Strategy-Retriever takes these transactions, retrieves the past cases indexed
under those transactions, and returns a list of the tax-reduction strategies used in
those cases. These strategies are expressed in the form of a path through the space
of possible plans, specifically, a sequence of decomposition rules that can be used
by the hierarchical planner to transform the user’s goals into one of the candidate
partial plans in the hierarchical planner’s repertoire.
The Strategy-Processor then calls the hierarchical planner, which attempts to
construct a plan for each of the strategies suggested by previous cases, backtracking
between plans. For each strategy, the hierarchical planner first constructs an
abstract plan, including the taxpayer’s goals, if any, and the general system goal of
reducing the taxpayer’s income tax. There are a number of ways this top-level plan
can be refined: the system can attempt to show that no taxable income results from
the transaction, or if that fails, exclude income, defer it, attribute it to someone
other than the client, or find deductions or credits. In addition, there are combined
plans, since several of these approaches could apply simultaneously to the same
transaction. Each of these plan refinements, in turn, can be refined in several ways.
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 231
.
Figure 2. CHIRON’s architecture, with details of the Strategy-Processor
Some plans can be repeated with different variable bindings; some can be repeated,
but not during a given tax year; and others can’t be repeated at all. The plan space
is not deep, but it is potentially very wide. The hierarchical planner refines the toplevel
plan as far as it can, using suggestions from the Strategy-Retriever, until all
the remaining tasks are either primitive or open-textured.
The case-based planner then attempts to refine the open-textured partial plan. It
instantiates the prototype plan, adding facts and changing parameters as necessary
to fit the current situation. Next, it retrieves all the previous cases where the taxpayer
executed, or attempted to execute the current strategy. For each such case,
if it involves any facts comparable to those in the current situation, the case-based
planner computes a mapping between the facts of the current situation and those of
the previous case. The result of the mapping is a list of pairs of facts that match, a
list of pairs of facts that don’t match but are comparable, a list of facts that appear
only in the previous case, and a list of facts that appear only in the current situation.
The case-based planner then calls the Argument-Builder module to construct a
subset lattice, storing the current situation at the root, and the prototype and previous
cases at nodes determined by the number of facts they share with the current
232 KATHRYN E. SANDERS
situation. This use of a subset lattice to capture the relationship between a set of
cases and the current situation is based on HYPO’s, with the difference that CHIRON
uses the case facts, while HYPO uses dimensions inferred from the underlying facts.
Unlike CHIRON’s case facts, HYPO’s dimensions incorporate substantial domain
knowledge. They “identify the features that are the bases of important similarities
and differences among cases” (Ashley 1988, p. 377).
The case-based planner then considers each of the dimensions on which the
current situation is weaker than the prototype, if any. If the plan is stronger than at
least one successful case on each of these dimensions, then the case-based planner
returns to the hierarchical planner a set of steps that can be substituted for the
open-textured plan step. The hierarchical planner continues refining the plan until
all the steps are primitive, and Report-Results prints out the plan and relevant
citations. If it is weaker on some dimension than any previous case, but there is
a “trend” towards weakening cases along that dimension (defined as two previous
cases where the taxpayer won with increasingly weak positions), then the new set
of steps is also returned to the hierarchical planner. Otherwise the plan is rejected.
After the Strategy-Processor has processed each plan, Report-Results prints out
the results for the user. If a plan is rejected by the case-based planner, Report-
Results prints out the weaknesses that caused the plan to be rejected and relevant
case citations. Otherwise, Report-Results prints out the plan and citations to relevant
provisions of the Internal Revenue Code and the most-on-point previous cases.
It then gives HYPO-style arguments for and against the success of the plan and
indicates the ways in which the plan is weaker than the prototype (if any) and the
support provided by previous cases for each weakness.
Next the Combination-Retriever takes a list of tax-reduction strategies, retrieves
the previous cases that executed any of those strategies, and returns a list of subsets
of those strategies that have been used successfully in combination in past cases.
This module looks only at the strategies and whether they succeeded in combination,
without considering the facts of the current situation or comparing them to
the facts of previous cases.
Finally, the user selects a combination of strategies for the Combiner to examine.
The user may choose one of the combinations suggested by the Combination-
Retriever or some other combination, and may continue to request different
combinations until he or she is satisfied. After each request, the Combiner calls the
hierarchical and case-based planners to construct, if possible, a plan and supporting
arguments for the desired combination, and calls Report-Results to print out the
plan and supporting arguments, if the plan is being recommended, or if not, the
reasons why it is being rejected.
When called by the Combiner, instead of backtracking, it constructs a plan for
the first strategy and then attempts to add steps to achieve each additional strategy.
Whether a combination of plans can be recommended depends on the facts. The
hierarchical planner rejects combinations that involve a direct conflict, for example
if one plan requires that the taxpayer sell a house and the second requires that he
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 233
or she give it away. The case-based planner determines whether the combination
has been tried before, and if so, whether it succeeded in previous cases. A lack
of case support does not cause a plan to be rejected, however. As long as there
is no obvious conflict, CHIRON constructs the combined plan and reports the case
support for the combination, if any; the final decision whether to risk executing the
plan is left to the user.
2.2. KNOWLEDGE REPRESENTATION
CHIRON’s knowledge base includes representations of statutory rules, cases, safe
harbor plans (prototypes), and the relationship between the rules, prototypes, and
cases.
Since CHIRON is a tax planner, its knowledge base must include part of the tax
law. The rules governing the United States income tax are contained in Subtitle
A of the United States Internal Revenue Code. Subtitle A in turn is broken down
into more than fifteen hundred sections, each of which can be anywhere from a
paragraph to several pages long.
These rules fall into a few broad categories. First, there are rules concerning
the items that should be included in the taxpayer’s income, for purposes of the tax.
Gifts received by the taxpayer, for example, are not counted as income, but salary
and interest on bank accounts generally are. Second, there are rules providing for
deductions – items that may be subtracted from the taxpayer’s income before the
tax is calculated. There are rules for calculating the amount of income or loss
generated by selling a piece of property. There are rules for computing the amount
of tax, once the net income has been calculated. And there are rules specifying
credits, amounts that can be subtracted from the tax itself before it is paid.
These statutory rules form the basis for the hierarchical planner’s decomposition
rules. Following the same basic structure as the tax rules, the hierarchical planner
starts with a top-level system goal of reducing the taxpayer’s taxes. This can be
achieved by finding credits or reducing taxable income. Reducing taxable income,
in turn, can be achieved by finding exclusions, showing that income should be
attributed to someone else, finding deductions, or deferring income. The plan of
finding a deduction can be achieved by establishing that one has satisfied Section
121, and so forth. This basic framework is shown in Figure 3.
Examples of the hierarchical planner’s rules are shown in Figure 4. The first of
these rules states that one way to reduce a taxpayer’s taxable income for the tax
year starting on ?start-year and ending on ?end-year is for the taxpayer to deduct
expenses during that time period. This corresponds roughly to Section 63 of the
Internal Revenue Code, which provides, “[T]he term ‘taxable income’ means gross
income minus the deductions allowed by this Chapter . . . ” The second rule states
that one way for the taxpayer to deduct expenses during a given year is to establish
that he or she has satisfied Section 1034 of the Internal Revenue Code during that
year.
234 KATHRYN E. SANDERS
.
Figure 3. Part of the plan hierarchy
(todo ?tsk (occurs (reduce-taxable-income ?taxpayer)
?start-year ?end-year)
(plan
:id 14
:steps
((step1 :action (occurs (deduct-expenses ?taxpayer)
?start-year ?end-year)))))
(todo ?tsk (occurs (deduct-expenses ?taxpayer)
?start-year ?end-year)
(plan
:id 141
:steps
((step1 :open-textured
(establish (satisfied Section-1034 ?taxpayer)
?start-year ?end-year)))
:citation 1034))
Figure 4. Examples of CHIRON’s rule schemas.
Each rule gives the hierarchical planner a task and a plan for achieving that
task. Each plan includes a list of steps, which may be of type “primitive”, type
“action”, or type “open-textured”. Action steps and primitive steps are types that
can be found in many planners, the former being steps that the planner can reduce
into a more concrete plan, and the latter, steps that can be part of a final executable
plan. The open-textured steps, also called “strategies”, are unique to CHIRON.
These strategies represent the open-textured rules. The hierarchical planner, whose
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 235
knowledge is derived only from statutory rules, is unable to reason about opentextured
steps, but they are not primitive, so they cannot be included in the final
plan as they stand.When the hierarchical planner encounters an open-textured step,
it must call the case-based planner for help.
A single strategy may correspond to several pages of text: the pre-1997 §1034,
for example, was about four pages long, but it corresponds to a single strategy, with
the token “Section-1034”. Other strategies include deduction of moving expenses,
deduction of mortgage interest, exclusion of the value of housing provided for the
convenience of the employer, the one-time exclusion of gain on sale of a residence
by the elderly provided by the pre-1997 §121, like-kind exchange of property, and
deduction of business travel expenses. Since taxpayers construct plans for each
strategy as a whole, and previous cases illustrate the entire strategy, CHIRON also
represents and reasons about each strategy as a whole.
From the hierarchical planner’s point of view, strategies are a single token.
More accurately, considering CHIRON as a whole, each strategy is represented by
a prototype; a set of deformations, or ways in which the prototype can be adapted;
and a set of past cases where the taxpayer attempted (successfully or unsuccessfully)
to execute the strategy. All of these knowledge sources are indexed under
the strategy-token and retrieved when the case-based planner reasons about the
strategy. Another way to look at this is that each strategy corresponds to a region
partially defined by a set of cases, with the prototype at the origin. The boundary
of the region is extended by positive cases and limited by negative ones.
2.2.1. Cases
CHIRON’s knowledge base also includes representations of twenty-four United
States tax cases. These cases are used in a variety of ways: to suggest partial plans
for the planner to consider, to outline the limits of the space of possible plans,
and as the basis for arguments for and against the success of a particular plan. In
order to support these tasks, each case representation includes the case’s name (i.e.,
its official legal citation), short-name (an abbreviated form of the citation for use
in text), court, date, facts, strategies involved, and holdings of the case, as well as
various indices. Information about which features strengthen or weaken a case with
regard to a particular strategy is stored separately.
Part of the representation of Hughston v. Commissioner, T.C.M. (P-H) 50,188
(1950), one of the cases in CHIRON’s case-base, is given in Figure 5; for the full
representation and the original text of the case, see Sanders (1994).
As shown in Figure 5, the representation of case facts includes both detailed
facts and abstractions from those facts. The abstractions are those that seemed
useful, based on the reasoning in the particular case, other cases, or general domain
knowledge. The detailed fact representation corresponds closely to the facts as
given in the case report. It is not exact – for example, the taxpayer in this case
was employed by Shell “during 1947 and for some time prior thereto”. The exact
date is not given, nor is it relevant to the taxpayer’s sale and purchase of two houses
236 KATHRYN E. SANDERS
(make-case
:name "Hughston v. Commissioner, T.C.M.(P-H) 50,188 (1950)"
:short-name "Hughston"
:court "Tax Court"
:date 1950
:facts ’((occurs (individual-return return2) (1947) (1947))
(occurs (taxpayer return2 Hughston) (1947) (1947))
(occurs (house house1) (October 1945) (December 1947))
(occurs (real-property house1)
(October 1945)(December 1947))
(occurs (spatial-part room1 house1) (1947) (1947))
(occurs (bathroom room1) (1947) (1947))
(occurs (room room1) (1947) (1947))
(occurs (real-property room1) (1947) (1947))
(occurs (floortype room1 tile) (1947) (1947))
(occurs (house house2) (October 1945) (December 1947))
(occurs (real-property house2)
(October 1945)(December 1947))
(occurs (physically-occupy Hughston house1)
(October 1945) (February 27 1947))
(occurs (physically-occupy Hughston house2)
(February 28 1947) (1951))
...
(occurs (selling selling1)
(February 27 1947)(February 27 1947))
(occurs (seller selling1 Hughston)
(February 27 1947) (February 27 1947))
(occurs (object selling1 house1)
(February 27 1947) (February 27 1947))
... )
:action-types ’(:sale)
:property-transferred ’(:real-property :cash)
:property-received ’(:cash :real-property)
:transfer-types
’((:sale :real-property :cash)
(:sale :cash :real-property))
:strategies ’(:like-kind-exchange)
:holdings ’((:like-kind-exchange :government)))
Figure 5. Part of CHIRON’s representation of Hughston v. Commissioner.
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 237
in 1947, so I simply approximate the starting date of his employment as 1946. In
general, however, the goal was to stay as close to the facts given in the case report
as possible.
CHIRON’s case structures are much like HYPO’s legal case frames (Ashley
1991); both include the official legal citation, a shortened form of the citation, the
open-textured provisions interpreted by the court (“claims” for HYPO, “strategies”
for CHIRON), holdings, and facts. Both reflect the way that lawyers analyze cases.
Law students are taught to “brief” cases, that is, to prepare short abstracts of
cases including their important elements: the name, court, and date, the facts, the
holdings, and the ratio decidendi, or reasoning in the case.
Neither system’s case representation includes the court’s reasoning. The Tax
Court cases that CHIRON uses often have little explicit reasoning. Like HYPO,
CHIRON captures some of the courts’ reasoning along with general domain knowledge,
without associating it with specific cases, in the form of knowledge about
which features strengthen or weaken a case and the ways in which cases are
compared and contrasted.
CHIRON’s representation also includes the name of the court deciding the case.
That information is not used in the current implementation. In future work, however,
CHIRON’s arguments could be refined by making use of this information. To
support more subtle arguments based on the identity of the court, it would also be
useful to include the procedural context of each decision, as suggested in Berman
and Hafner (1991).
The facts given in the case report and useful abstractions from those facts
are represented as a list of propositions in a modal temporal logic with a formal
syntax and semantics, an extension of the temporal logic developed by Shoham
(1988), modified to incorporate the modal operators “know”, “believe”, “want” (or
“goal”), “obligated”, “permitted”, and “prohibited”. This language is influenced
by the work of McCarty in representing legal concepts, especially (McCarty 1977,
1989b). It is flexible enough to represent basic concepts such as property and time,
plus the unpredictable idiosyncratic details that occur in each new case, such as the
fact that a kitchen in a house has a tile floor, or the taxpayer is a war veteran, or
the taxpayer has two children, one of whom is ten years old. The language used for
case facts is the same as the language used for input.
Timing is critical in tax problems, so it is necessary to represent the time of
actions and events. The token “*now*” is used to indicate the time at which the
plan is being constructed, some point in time later than all the cases in the case
base and earlier than the performance of the client’s intended actions. The tokens
“*beginning-of-time*” and “*end-of-time*” are arbitrarily assigned the dates of
January 1 in the year 1 and December 31, 99999, respectively. A fact that is always
true is said to hold over the interval from *beginning-of-time* to *end-of-time*.
The endpoints of these intervals correspond to points in time. Time constants can
be the year (e.g., (1776)), the month and year (e.g., (July 1776)), or the month, day,
238 KATHRYN E. SANDERS
and year (e.g., (July 4 1776)). So far, this has been sufficiently precise; it would be
simple to add the time of day, if that became necessary.
Certain intervals of time are also important in this domain: the length of time the
taxpayer has owned a particular piece of property, the length of time during which
the taxpayer occupied a piece of property, the amount of time the taxpayer had
been away from a piece of property before selling it, the time elapsed between one
sale and another, and so forth. Intervals are represented as triples (years, months,
and days).
Note that actions and obligations may have distinct times. Generally, they will
coincide, because as a rule, the system needs to reason about the relationship
between an agent’s actions and the obligations or prohibitions that affected those
actions at the time they were performed. But sometimes a legislature will pass a
law now obligating or prohibiting the performance of a certain action next year,
and the fact that that prohibition is in effect now will affect plans made for the
future, so it can be useful to express the distinction.
It is also necessary to represent various types of property, property rights such
as ownership and possession, and the ways in which those rights can be transferred.
The United States’ income tax is transfer-based; that is, tax is triggered by
the transfer of money or property from one legal entity (individual, corporation,
trust, etc.) to another (Chirelstein 1988). Concepts of property, property rights, and
property transfers are central to this domain. Some of these concepts were already
formalized in Sanders (1989a, b); others, I added as necessary.
Property is divided into categories which are often treated differently by the tax
system: real property, tangible personal property, and intangible personal property.
Intuitively, real property includes land and buildings: houses, condominiums, cooperatives,
and so on. Tangible personal property includes tangible property other
than real property, such as cars or books. And intangible personal property includes
cash, stocks, and the rights of a tenant to his or her rent-controlled apartment.
“Own” and “possess” are primitives in this language. Intuitively, ownership is
a collection of legal rights, including the right to possess, use, give, rent, or sell an
object. By contrast, one possesses an object if one physically controls it. Thus, if a
taxpayer rents a house, he or she possesses the house, even though it is still owned
by the landlord. Anyone who borrows a book, possesses it, even though they do
not own it.
Transactions are classified according to the type of rights and the type of property
transferred. A transfer of property can be a sale, gift, rental, or loan. For
this purpose, I define a sale as a transfer of permanent ownership of property in
exchange for a price; a gift, as a transfer of permanent ownership of property
in exchange for which the transferor receives nothing; a rental, as a temporary
transfer of possession in exchange for a price; and a loan, as a temporary transfer
of possession in exchange for which the transferor receives nothing. The transferors
are individuals who are legally capable of owning property, that is, human beings,
partnerships, corporations, and so forth. Services are generally performed by in-
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 239
dividuals, though a corporation or partnership may have its employees perform
services on its behalf. The property transferred is anything that can be owned,
tangible or intangible. Property can be transferred in exchange for cash, other
property, or services; services can be transferred in exchange for cash, property,
or services.
These transactions are reified. For example, in Hughston, the taxpayer sold a
house. Part of the facts of that case are represented as follows:
(occurs (selling selling1)
(February 27 1947) (February 27 1947))
(occurs (seller selling1 Hughston)
(February 27 1947) (February 27 1947))
(occurs (object selling1 house1)
(February 27 1947) (February 27 1947))
In other words, there was a sale on February 27, 1947, the seller was Hughston,
and the object of the sale was a certain house, house1.
In this representation, the seller, the object, the payment, and the buyer are all
expressed as separate relations to the reified “selling”. Rentals, gifts, loans, employment
transactions, and volunteering are represented similarly. This approach
makes it possible to represent transactions where, as here, part of the information
is unknown. Because there was a sale, there must have been a buyer, but no information
about the buyer is given in the case, and it is not needed to reason about
the tax consequences of the sale to the seller. The system needs a representation
that can handle this incompleteness. The representation also allows us to represent
transactions where there is more than one individual with the same role, where the
taxpayer received both real estate and cash in payment, where there are two buyers,
two pieces of property are sold, two things are received as part of the same gift, and
so forth.
Note that in Hughston, the facts that house2 is a house and is also a piece of
real property, are both included in the case representation. For each object in a case
representation, both its types and abstractions from those types are included.
In reasoning about property, I have also found it necessary to represent certain
basic spatial concepts: location, nearness, distance, area, and the idea that one piece
of property is part of another, which I call “spatial-part”.
Other concepts related to property that I have found useful include physical descriptions,
such as the type of floor in a room and the fact that one piece of property
contributes part of the value of a whole package. For example, in comparing the
two houses bought and sold by the taxpayer in Hughston, the court noted that one
had a tiled kitchen and the other did not:
(occurs (house house1) (October 1945) (December 1947))
(occurs (spatial-part room5 house1) (1947) (1947))
(occurs (kitchen room5) (1947) (1947))
240 KATHRYN E. SANDERS
(occurs (floortype room1 tile) (1947) (1947))
(occurs (house house2) (October 1945) (December 1947))
(occurs (spatial-part room7 house2) (1947) (1947))
(occurs (kitchen room7) (1947) (1947))
(occurs (not (floortype room7 tile)) (1947) (1947))
The court also noted that one of the houses had three bathrooms, plus a half-bath,
and the other had only one:
(occurs (number-of-bathrooms house1 3.5) (1947) (1947))
(occurs (number-of-bathrooms house2 1) (1947) (1947))
I have also found it useful, particularly in the context of employment transactions,
to express knowledge about types of organizations, occupations, and work.
Employers include corporations, universities, and the government, for-profit and
tax-exempt organizations. Types of services performed all fall into the general
category of “labor”; they include, for example, research, teaching, and the practice
of law. Employees also fall into certain categories. Hughston was a lawyer; he
was also an individual and a taxpayer. Individuals in other cases fall into other
categories, such as “war veteran”. They are students at particular institutions and
parents or children of other individuals.
Some tax provisions relate to age, so I define predicates to express a person’s
(or thing)’s age and the fact that one person (or thing) is older than another:
(occurs (age child1 10) (1947) (1947))
(occurs (older child1 child2) (1947) (1947))
Within the tax domain, this project focused on a cluster of provisions having
to do with residential housing. Actions that are important in this domain include
maintaining and occupying the property in question. In Trisko v. Commissioner,
29 T.C. 515 (1957), the taxpayer occupied his house for several years, then moved
abroad temporarily and rented the house to a tenant who took care of it in his
absence, and then found when he returned that a strict rent-control law had been
passed that had the effect of prohibiting him from returning home. A portion of
these facts is represented as follows:
(occurs (physically-occupy Trisko house1)
(October 1944) (February 1948))
(occurs (maintain lessee2 house1)
(February 1948) (October 9 1951))
(prohibited (occurs (physically-occupy Trisko house1)
(1951) (1951)) Law (1951) (1951))
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 241
In some cases, the taxpayer is occupying more than one residence at a particular
time (for example, the date of sale); if the number of the taxpayer’s residences is
given in the case, I represent that as well.
Finally, there are certain primitive states and events that are specifically related
to the tax domain: filing and signing of forms, being a taxpayer on a particular
return, and types of returns.
In addition to working with the facts that are explicitly stated in the cases,
CHIRON can also make certain limited inferences from those facts. First, inferences
about property types, types of work, types of employer, and types of transaction
are implemented as forward-chaining rules that fire when the input facts are first
asserted. Thus, for example, if the input facts include the fact that a given piece of
property is a house, then the rule that houses are real property will fire, and the fact
that the given house is also real property will be added to the knowledge base.
Second, the case-based planner makes inferences that are implemented as Lisp
functions called by the case-based planner, rather than as forward or backwardchaining
rules. Most of these inferences involve reasoning about time: for example,
determining how long the taxpayer has occupied a given piece of property, how
much time passed between the sale of one piece of property and the purchase of
another, and what the taxpayer’s age is today, based on his age at some earlier date.
In addition to temporal inferences, the case-based planner counts the instances
of certain propositions in the input. For example, with respect to the rollover
strategy, it computes the number of residences the taxpayer has at the date of sale
by retrieving all the instances of the taxpayer occupying any residence on that
date, and counting them. This kind of inference might be difficult to formalize in
general, but since the number of possible instances of each type of fact is bounded,
the problem becomes relatively simple. With a closed-world assumption, that all
the relevant facts having to do with the taxpayer’s residences, tax forms filed, and
so forth, are included in the input, and a unique-names assumption, that any two
distinct terms can be assumed to be unequal, axioms can be written to cover all
possible cases.
2.2.2. Prototypes
In addition to representations of actual legal cases, the casebase includes prototype
cases. Prototypes are represented using the same structure as cases, but indexed
separately.
CHIRON’s prototypes are not real cases. In Protos, actual previous cases are
used as prototypes (Bareiss 1988). Both approaches are influenced by work in
psychology suggesting that concepts are best represented by a set of more or less
typical cases, interrelated by “family resemblances” (Rosch and Mervis 1975). In
law, the reported cases are not typical, almost by definition: the standard, typical
cases do not go to court. None of the court cases in CHIRON’s case base is suitable
for use as a prototype.
242 KATHRYN E. SANDERS
Instead, CHIRON’s prototypes are based on general domain knowledge. This
knowledge includes the commonsense meaning of statutory phrases. Terms such
as “principal residence”, although they are underspecified, do have some commonsense
meaning. Additional information is provided by the cases. In any law case,
there are some easy questions that are not at issue. For example, the taxpayer’s
old house may be clearly his principal residence, while the new one is at issue, or
vice versa. The easy questions give some information about the prototypical case.
And hard questions can also provide information. For example, if the issue in a
case is whether a house can qualify as a principal residence if the owner is not
living there, one can infer that actually living in the house is part of the prototype.
Consistent with this approach, it would also be possible to base prototypes on lawyers’
experiences of more typical cases, as suggested in (Schlobohm and McCarty
1989).
CHIRON’s prototypes are templates, or general fact patterns, instantiated for
each case. For example, the prototype rollover plan stored in CHIRON’s knowledge
base refers to a taxpayer selling a house, but the prototype CHIRON uses
in constructing and reasoning about a particular plan involves a specific person and
property. Similarly, in a treatise or regulation, a lawyer would find a generalized
plan, with the details of particular cases to be filled in by the lawyer.
2.2.3. Facts
One of the questions I wanted to explore in CHIRON is what would happen if the
representation of case facts corresponded as closely as possible to the facts in the
case report. Accordingly, CHIRON’s vocabulary includes 75 different predicates.
My goal was to choose an economical set of predicates that would enable the
case representation to capture as much as possible of the statement of facts in the
original case reports, from obviously significant facts such as the purchase or sale
of a house to facts whose importance is not immediately obvious, such as the fact
that the taxpayer was a war veteran, the number of bathrooms in a house, or the
type of floor in a bathroom. Rather than choosing a set of predicates in advance,
I added new predicates as necessary when representing each case. Most facts are
represented explicitly, but a few, such as the fact that a building that is a house is
also real property, are inferred by the system. An example of CHIRON’s predicate
representation is given in Figure 6, and a complete list of predicates is given in
Appendix A.
For purposes of exploration, CHIRON initially included relatively little domain
knowledge about the relative significance of facts. CHIRON has just one simple
assumption: any fact mentioned in a case report must be important with respect to
the strategies involved in the case. This is of course an over-simplification, but does
represent what might happen if a lawyer read a set of cases in a new area, keeping
an open mind at first about the facts. As such, it seemed a useful starting point for
the system.
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 243
(make-fact
:name ’apartment
:user-name "Apartment"
:supports nil
:compare-fn #’(lambda (f1 f2 strategy)
(cond ((and (equal (modality f1)
(modality f2))
(equal (fact-negations f1)
(fact-negations f2))
(equal (predicate f1)
(predicate f2)))
:match)
(t :unknown))))
Figure 6. CHIRON’s representation of the predicate “apartment”.
Some domain knowledge about the facts is incorporated in the system, however,
and the representation is designed so that more domain knowledge could be added
easily. First, as shown in Figure 6, the predicate representation allows for encoding
domain knowledge about individual facts. If it is known that the presence of a fact
in a case supports the taxpayer’s position with regard to a particular strategy, that
can be encoded in the “:supports” field. In Figure 6, the value of the :supports field
is nil, indicating that nothing is known about the predicate’s significance to any of
the strategies.
Second, the representation allows for encoding domain knowledge about the
relative strengths of two given facts. Every predicate in CHIRON’s vocabulary has
an associated comparison function (“compare-fn”) that is called when comparing
two facts from different cases or prototypes. In Figure 6, the compare-fn takes
two facts and the strategy being considered as arguments and returns either :match
or :unknown. The system has no knowledge about the relative strength of two
facts involving this predicate and makes no distinction based on the strategy being
considered.3 In general, however, the possible return values include :stronger
and :weaker, in addition to those used here. By using the strategy parameter and
the return values :stronger and :weaker, the compare-fn can incorporate domain
knowledge about the relative strengths of the facts with respect to a particular
strategy.
Currently, only one of the predicates uses the :supports field, and only a
few incorporate knowledge about the relative strengths and weaknesses of facts.
The :supports field is used by the predicate “war-veteran”, which supports the
taxpayer’s case for a rollover under the old §1034. The compare-fn for “absence-
3 The compare-fn here indicates that two facts involving the “apartment” predicate are considered
to match if they involve the same predicate, are both positive or both negative, and have the same
“modality” (“occurs”, “prohibited”, or “goal”.).
244 KATHRYN E. SANDERS
before-sale”, for example, captures the knowledge that a taxpayer who has not
occupied a house for a year has a stronger case with respect to §121 than someone
who’s been away for five years.
More knowledge of this type could be added without modifying CHIRON’s
design in any way. An interesting direction, though more challenging, would be to
give the system the ability to reason about why a fact might strengthen or weaken
a case with respect to a particular strategy. This would require more fundamental
modifications, however.
2.3. DESIGN DECISIONS
2.3.1. Rules and cases
In order to reason about open-textured rules, cases, and the relationship between
them, CHIRON combines hierarchical and case-based planners in a hybrid system.
Both rules and cases exist in the domain. The rules are not, as in many expert
systems, an artifact created by the system designer. We could translate all the cases
into rules; but there are many rules for which each case could stand. We could
translate all the rules into cases, but each rule corresponds to a set of cases. The
set might be very large, even if it could be defined precisely; and given the open
texture of the rules, specifying the corresponding set of cases would be difficult.
The natural structure of the tax domain is based on rules and cases, and following
that structure makes the system easier to design and maintain.
CHIRON’s case-based and hierarchical planners are equal, interdependent
reasoners. Each cooperates with the other. The hierarchical planner calls the casebased
module for guidance in choosing adaptations and refining the plan to satisfy
the open-textured predicates, and the case-based planner uses the hierarchical
planner to help with indexing, controlling adaptations, and combining plans. This
tightly-coupled architecture exploits our knowledge about the potential interactions
of the case-based and hierarchical planners.
Again, this decision reflects the structure of the domain. In tax, both the rules
and the cases are named. The cases refer to the rules they interpret, and books discussing
the rules are annotated with references to the cases. And the rules and the
cases complement each other. The rules organize cases into groups and partition the
case base; the cases interpret the rules and suggest rules that might be considered.
Suppose the system had a purely rule-dominant design. Then it would refine
the plan as far as possible using rules, and invoke the case-based planner only
when the rules run out. This is comparable to the algorithm used by Gardner’s
legal analysis system (Gardner 1987). As the rule base grows larger and larger,
however, search control – controlling the choice of plan decompositions – becomes
an important issue. The case-based reasoner could suggest search paths for the
hierarchical planner, if it were invoked earlier.
Suppose the system had a case-dominant design. Then the system would retrieve
similar cases first, and apply rules second, if at all. But rules structure the
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 245
case base. Cases are related to each other by the fact that they interpret the same
rules. Rules also partition the case base. A case-dominant approach fails to take
advantage of this structure.
2.3.2. Open-textured planning
CHIRON’s prototype, cases, and adaptations define a space of possible plans. Any
of the cases in the case base can be obtained by starting with the prototype and
applying the appropriate adaptations. This makes it possible for a variety of different
plans to be generated for the same basic strategy, allowing CHIRON’s plans to
cover a range of possibilities. For example, a house can be the taxpayer’s principal
residence whether he’s lived there for one year or fifty, whether it has one bedroom
or a hundred, and so forth.
This design raises an important question, however: how large should the space
of acceptable plans be? The size of the space is a function of the amount of case
support the system requires a plan to have. The more case support required, the
more plans will be rejected, and the more control will be taken away from the user.
The less case support required, the more dubious plans will be suggested, and the
less helpful the system’s advice will be for the user.
Fairly arbitrarily, I have decided on the following requirements: first, that there
must be at least one most analogous case with the desired result, and second, that
for each fact where the current situation is weaker than the prototype case, there
must be either a favorable case that is at least as weak, or a trend in that direction.
HYPO considers the most analogous, or “most-on-point”, cases to be “[t]he most
analogous cases to a problem situation as determined by relative degrees of overlap
between sets of dimensions shared by cases and problem situation. Measured in
HYPO by claim-lattices” (Ashley 1988, p. 378). CHIRON also determines which
cases are most analogous by constructing a lattice and examining the relative degrees
of overlap between cases; the only difference is that CHIRON uses its case
facts, rather than HYPO’s dimensions, in building the lattice. A case with the desired
result, in this domain, is simply defined as a case that holds for the taxpayer.
There are three things to note about this standard. First, it assumes that it is
possible to argue about the weaknesses in a plan separately. In fact, they may
be cumulative; a plan with several weaknesses, even if there is some support for
each, may be too weak to consider. The system compensates for this possibility by
informing the user about each of the weaknesses; given the information, the user
can decide whether to go ahead.
Second, the system does not attempt to find new creative arguments. For example,
if the taxpayer lived in his house at the time of sale in every §1034 case, the
system would not generate a plan that involves living somewhere else. It would be
interesting to model this kind of creativity, but it is outside the scope of this article.
Third, the system uses the timestamps on the cases. I define “trend” simply as
two cases, one decided earlier than the other, where both decided in favor of the
taxpayer, both were weaker than the prototype along a given dimension, and the
246 KATHRYN E. SANDERS
second was weaker than the first. Taking into account the temporal dimension of
the casebase is rare in CBR systems. Cuthill also uses timestamps: her system,
CHASER, includes a preference for more recent cases in its retrieval algorithm
(Cuthill 1992; Cuthill and McCartney 1993). Reasoning about the time of cases
seems intuitive in legal reasoning systems; it could be useful for any system in a
domain that changes over time.
2.3.3. Detailed fact representation
As described above, CHIRON has a rich, detailed representation for the facts of
its cases. This representation is used in indexing, adaptation, and comparison of
cases. Cases are indexed under all of their facts, so CHIRON can retrieve more
cases than a system that uses a less detailed or more abstract representation. The
system can find more ways to distinguish two cases. And it has a large repertoire of
possible adaptations. Adaptations of the prototype plan are also based on the facts:
some adaptations consist simply of adding a fact to or subtracting a fact from the
prototype; others involve varying the parameters of a given fact. Any fact in one of
the previous cases can potentially be added to the prototype to make a new plan.
CHIRON’s use of the facts is limited, however. It only considers the facts of cases
that involve the current strategy. This results in some loss of reasoning power. If
the system looked at facts first, it would retrieve some cases from other contexts,
which might be useful. For example, when reasoning about §1034, it might also
retrieve cases under §121, where the taxpayer must also establish that he or she has
sold a principal residence.
The loss of reasoning power seems outweighed by the gain in efficiency, however.
Practically speaking, cases interpreting different plans are only useful if there
is no closer match. Legal provisions are context-dependent; their meaning cannot
be assumed to be the same wherever they appear (Gardner 1987, p. 53). In a welldrafted
statute such as the Internal Revenue Code, language is generally used in a
consistent way, and cases interpreting one section are persuasive in arguing about
another, but still, the results are not necessarily the same. I chose, therefore, to see
how far the system could go using only cases from the current context. For interesting
work on “cross-context remindings” in the domain of tort law, see Cuthill
(1992).
2.3.4. Conservative and aggressive plans
CHIRON associates a conservative plan, called the prototype, with each partial plan.
Because the system is operating in an adversarial domain, it is useful to have a safe
interpretation of the rules: a plan that Gardner’s system would accept as an “easy
question”, or what tax planners refer to as a “safe harbor”. If a taxpayer executes
this plan, she will probably be safe from challenge by the government; the more
she differs from it, the more likely she is to be challenged (and to lose). CHIRON’s
prototypes correspond to these safe-harbor plans.
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 247
It uses the previous cases as guidelines to show how far the prototype can be
adapted. Where the plan differs from the safe-harbor plan, it may be vulnerable to
attack; so in addition to the plan, the system outputs citations to relevant rules and
cases and a discussion of each plan’s strengths and weaknesses, using the model of
legal argumentation developed in Ashley (1991).
CHIRON constructs plans that are as similar to the prototype as possible. This
strategy gives plans a conservative bias, which is consistent with much of tax
planning. It is also consistent with the approach suggested in Kambhampati and
Hendler (1989), that adaptations should in general be conservative.
If a plan is weaker than any of the previous successful cases along some dimension,
the system rejects it, unless there is a trend towards weakening cases
along that dimension. CHIRON could adopt a more aggressive strategy by changing
the facts of the current situation to match weaker pro-taxpayer cases, where the
weakness would be beneficial to the taxpayer. Report-Results does point these
opportunities out to the user, but does not generate plans for them.
Alternatively, CHIRON could adopt an even more conservative strategy by changing
the facts of the current situation to match those of the prototype. For example,
if the taxpayer is not living in a house, CHIRON could suggest that he move in, then
execute the rollover strategy. This change might even make it possible to execute
plans that would be rejected by the current version of CHIRON for lack of support.
Moving into a house and making it his residence would be equally effective if
the taxpayer had been gone for a year (in which case there would be support
for a rollover even without moving back) or ten years (in which case the rollover
strategy would currently be rejected). Many facts cannot be changed, however. If
the taxpayer sold his house five years ago and didn’t buy a second one, he can’t
unsell it and start again. In addition, changing individual facts requires knowledge
about the effect of small changes on the whole plan. It might be, for example, that
in order to move into the house, the taxpayer would have to move six hundred miles
away from his current job, which would mean getting a new job. CHEF-style casebased
planning is based in part on the assumption that the planner doesn’t have that
kind of information about the interactions of parts of a plan; all he has to work with
are entire plans, as executed in the past (Hammond 1986). Instead of making the
changes, then, CHIRON reports the plan’s weaknesses to the user (whether or not
they are grounds for rejecting the plan) and allows the user to decide.
3. Examples
CHIRON solves a cluster of problems having to do with buying, selling, renting,
and owning residential housing. These problems have been chosen in part because
they are simple; most tax planners would agree on at least the obvious solutions. In
addition, they illustrate problems such as timing and satisfaction of open-textured
rules that are typical of the domain. Since they all involve transfers of residential
housing, the commonsense knowledge that must be formalized to handle them
248 KATHRYN E. SANDERS
involves the same cluster of concepts. They also illustrate the problems involved
in reasoning with a dynamic statute: the case base includes cases decided when
the time limit for §1034 rollovers was 18 months and cases decided under a later
version of the statute when the time limit was 24 months; since the system was first
designed, §1034 has been repealed entirely and §121 has been extended to cover
sales formerly covered by §1034.
The examples in this section are excerpts from transcripts, edited to fit within
the margins of an article. Following the usual convention, system output is given
in an easily distinguished typeface.
3.1. CHIRON SOLVES AN EASY PROBLEM
Consider a straightforward example: the taxpayer, whose name is Greenlee, wants
to sell a house. Her tax advisor asks CHIRON for suggestions on how to minimize
tax liability on the sale. She has owned the house, identified by the token 32-
Eleventh-Street, since October 30, 1972, and has occupied the house during that
entire period. She is now forty-two years old. CHIRON takes as input the internal
representation of these facts and the seller’s goals, as follows:
(occurs (individual-return return1) (1994) (1994))
(occurs (taxpayer return1 Greenlee) (1994) (1994))
(occurs (age Greenlee 42) (1994) (1994))
(goal (occurs (selling selling157) ft155 ft156)
Greenlee (9 20 1994) (9 20 1994))
(goal (occurs (seller selling157 Greenlee) ft155 ft156)
Greenlee (9 20 1994) (9 20 1994))
(goal (occurs (object selling157 32-Eleventh-Street)
ft155 ft156) Greenlee (9 20 1994) (9 20 1994))
(occurs (house 32-Eleventh-Street)
(0 0 0) (12 31 99999))
(occurs (physically-occupy Greenlee 32-Eleventh-Street)
(10 30 1972) (9 20 1994))
(occurs (owns Greenlee 32-Eleventh-Street)
(10 30 1972) (9 20 1994))
CHIRON then adds another goal, assumed to be universal to all taxpayers, of
reducing taxes. This goal is instantiated as follows:
(occurs (reduce-taxable-income Greenlee) (1994) (1994))
Next, the Classifier takes the facts of the current situation and applies rules to
determine that the situation involves one type of transaction: a sale of real property
in return for some other unspecified type of property.
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 249
The current situation involves the following
transaction types:
(:SALE :REAL-PROPERTY :PROPERTY)
The Strategy-Retriever then uses case-guided search to determine which
strategies to consider. First, it retrieves the cases in its case base that also involved
a sale of real property. There are four:
Cases retrieved:
"Welch v. Commissioner, T.C.M. (P-H) 79,010 (1979)".
"Trisko v. Commissioner, 29 T.C. 515 (1957)".
"Sayre v. United States, 163 F.Supp. 495 (S.D.W.Va. 1972)".
"Hughston v. Commissioner, T.C.M.(P-H) 50,188 (1950)".
Among them, these cases involved a total of three strategies: a §121 exclusion,
which permitted a taxpayer who was over 55 years old to sell his or her principal
residence and exclude part of the gain from tax, if he or she met certain tests; a
rollover, where the taxpayer sells one principal residence and buys another; and
a like-kind exchange, where the taxpayer exchanges one piece of property for
another similar one. Accordingly, the Strategy-Retriever returns a list of paths
(i.e., sequences of decomposition rules) that the hierarchical planner can follow
to transform the user’s goals into a partial plan for each of these strategies.
CHIRON now attempts to construct a plan for each of these strategies in the
current situation, starting with the like-kind exchange. First, the Strategy-Processor
calls the hierarchical planner, which constructs an abstract plan, including the
taxpayer’s goals and the general system goal of reducing taxes.
Plans and partial plans are implemented as a task network, represented by a set
of assertions stored in DUCK. Every task has an id, a description, start time, end
time, and type. Task types include :root, :plan-start and :plan-end (for the beginning
and end of the whole plan), :dummy-start and :dummy-end (for the beginning and
end of subplans), :goal, :action, :primitive, and :open-textured. The task network
is partially ordered. Where it is known that one task tsk1 directly precedes another
task tsk2, we assert (pred tsk1 tsk2). The task with no parent is the root; but for
convenience, we also assert (root tskn). Open-textured tasks have associated with
them a code section, identifying the portion of the Internal Revenue Code from
which they are derived; so for open-textured tasks, we assert (code-section sometsk
some-section-number).
The top-level task network is not output as part of the transcript, but internally,
it looks like this:
(task tsk1)
(descr tsk1 (reduce-taxes Greenlee))
(start tsk1 (1 1 0))
(end tsk1 (12 31 99999))
250 KATHRYN E. SANDERS
(type tsk1 :goal)
(task tsk2)
(descr tsk2 (object selling156 32-Eleventh-Street))
(start tsk2 FT154)
(end tsk2 FT155)
(type tsk2 :goal)
(task tsk3)
(descr tsk3 (seller selling156 Greenlee))
(start tsk3 FT154)
(end tsk3 FT155)
(type tsk3 :goal)
(task tsk4)
(descr tsk4 (selling selling 156))
(start tsk4 FT154)
(end tsk4 (FT155)
(type tsk4 :goal)
...
The starting and ending dates for the reduce-taxes task indicate that it is always a
goal, from the earliest to the latest known date. The starting and ending dates for
the other tasks, FT154 and FT155, simply indicate that those tasks have the same
starting and ending times, and that it is some unknown time in the future. There
will also be various other bookkeeping tasks: the starting point of the plan, that is a
predecessor to all of these tasks, the ending point of the plan, to which all of these
tasks are predecessors, and so forth.
After asserting the top-level plan, the hierarchical planner then refines it, using
the sequence of decomposition rules suggested by the Strategy-Retriever. When it
encounters a task of type open-textured, it calls the case-based planner.
The case-based planner then instantiates the prototype corresponding to the
open-textured strategy, adding facts and changing parameters as necessary to fit
the current situation. Next, the case-based planner retrieves all the previous cases
where the taxpayer executed, or attempted to execute the current strategy. For each
such case, if it involves any facts comparable to those in the current situation, the
case-based planner computes a mapping between the facts of the current situation
and those of the previous case. The result of the mapping is a list of pairs of facts
that match, a list of pairs of facts that don’t match but are comparable, a list of
facts that appear only in the previous case, and a list of facts that appear only in the
current situation.
The case-based planner then calls the Argument-Builder module to construct
a subset lattice, storing the current situation at the root, and the prototype and
previous cases at nodes determined by the number of facts they share with the
current situation.
The case-based planner then considers each of the facts where the current situation
is weaker than the prototype, and concludes there is insufficient support for
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 251
the plan. Because the intended users of the system are experts, however, the system
doesn’t simply continue to the next plan; instead, Report-Results prints out the
reasons why the plan is being rejected, as follows:
*******************************
RESULTS
*******************************
The strategy :LIKE-KIND-EXCHANGE has been rejected for the
following reasons:
(1) The strategy :LIKE-KIND-EXCHANGE is unlikely to succeed
because there is no support for the plan along the dimension
DURATION-OF-OCCUPANCY.
In the current situation:
(OCCURS (DURATION-OF-OCCUPANCY GREENLEE 23-ELEVENTH-STREET
(35 10 20)) (1 1 0) FT155)
*******************************
The taxpayer has been living in her house for many years. The properties involved
in a like-kind exchange must be “held for productive use in a trade or business or
for investment . . . ”. §1031. Personal use of the property weakens the taxpayer’s
case, and there are no cases in CHIRON’s case base with this weakness, so the
system rejects the plan.
The hierarchical planner then backtracks to the top-level plan of selling a house
and reducing taxes. It then continues with the next strategy, the §121 exclusion.
This strategy is also rejected, because the statute set a minimum age of 55 for
taking advantage of the provision, the taxpayer was only 42, and there was no case
support for a lower age.
With regard to the rollover, however, the taxpayer is in a strong position. Her
case is no weaker, and in some respects stronger, than the prototype for this
strategy. As a result, the case-based planner returns to the hierarchical planner
a set of steps that can be substituted for the open-textured plan step. The hierarchical
planner continues refining the plan until all the steps are primitive, and
Report-Results prints out the plan and relevant citations.
*******************************
RESULTS
*******************************
The strategy :ROLLOVER is suggested.
Code sections involved in plan: (1034).
Then Report-Results prints out the plan. The taxpayer is advised to live in the house
until the date of sale, sell it and buy another on the same date, and occupy the new
252 KATHRYN E. SANDERS
house immediately.4 In addition, the taxpayer should have only one residence at
the time of sale.
This is a very conservative plan, but it can be executed by the taxpayer, given
the simple facts of our example. That is, it can be executed from a tax point of view;
the realities of the housing market may make selling more difficult. Note also that
there is nothing in the plan about the value of the original house and the price of
the new house, factors which also enter into a real-world computation of the tax
savings and general desirability of the plan.
Next, Report-Results prints out arguments for and against the plan, generated by
the Argument-Builder from the case lattice. Like HYPO, CHIRON generates threeply
arguments: in other words, its arguments consist of a point, response, and (if
possible) rebuttal. CHIRON makes one of these three-point arguments for each most
analogous case (as defined in Section 2.3.2, above), whether or not the case held
for the government or the taxpayer.
In this case, which raises no real issues, there are no most-on-point cases holding
for the government, and in fact the prototype is the most-on-point case for the
taxpayer.
The taxpayer can cite the following cases for which there
are no most-on-point counter-examples:
"prototype: rollover"
In support of the plan, the system notes its similarities with the prototype. As in the
prototype, Greenlee has bought and occupied a house, and she will (if she executes
this plan) sell a house that she has been occupying until the date of sale. In both
the prototype and the current situation, the taxpayer has one residence at the time
of sale. And she has been living in her house longer than the minimal amount of
time required by the prototype. Finally, the system notes, the time elapsed between
the sale of one house and purchase of another, if Greenlee follows the plan, will
be shorter than that required by the prototype. This point is always followed by a
citation to the most-on-point case in question. For actual cases, the citation is in
correct legal form.
Next, the system prints out a response for the government. It cannot find any
weaknesses in the plan (which is indeed quite standard), but it does find a case,
Welch v. Commissioner, T.C.M. (P-H) 79,010 (1979), in which the government
won even though in some respects, the taxpayer’s plan was even stronger than the
prototype.
In Welch, the system points out, the government won even though the taxpayer
had occupied his first residence for seven years, longer than the prototype requires,
4 The advice to move immediately results from a lack of subtlety in CHIRON’s knowledge base.
It knows that sooner is better, so it assumes that moving immediately is the safest possible plan. In
fact, sooner is better only if you miss the statutory time limit; within the time limit, it doesn’t make
a difference.
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 253
bought a new piece of property only ten months and 22 days after the sale of his
old residence, and occupied the new property a year, eight months, and two days
after the sale, both less than the two-year interval allowed by the prototype.
Since the government has cited a new case in response, the taxpayer offers a rebuttal.
The best rebuttal here would be that Welch was decided under an old version
of the statute. In the time period covered by this case, the statutory time limit was
eighteen months. Because the taxpayer’s house was not completed within the time
limit, he was unable to occupy it until two months afterwards. The prototype case,
on the other hand, is within the current two-year time limit.
CHIRON has no means of reasoning about the dynamic nature of statutes, however,
so it must look for other distinctions. It notes that the type of property involved
in the two cases is different (the current situation involves selling a house; Welch
just involved some unspecified real property); Welch involved a rental, while the
current situation does not (the taxpayer in Welch temporarily rented a residence
while his new house was being built); and the suggested plan is stronger than
Welch with respect to three key facts: the time between sale of the old residence and
purchase of a new one, the time between sale of the old residence and occupancy
of the new one, and the length of the time during which the taxpayer had occupied
the old residence.
CHIRON has almost no domain knowledge about which distinctions are important,
so it considers them all. CHIRON’s first two arguments are not compelling: the
property types involved in the two situations are not significantly different, and the
fact that the taxpayer in Welch temporarily rented a residence while his new house
was being built should not disqualify him.
The third argument is a strong one, however. The current plan is stronger than
Welch along exactly the dimensions that the government had cited in comparing
Welch to the prototype. Thus, even though the system assumes that all its cases are
decided under the same version of the statute, it provides strong support for the
right conclusion.
The system has now considered all the strategies recommended by the caseguided
search mechanism and recommended one of them. It reminds the user
which strategy was successful, and offers the user a chance to consider other
possibilities. From here on, the user can request any strategy or combination of
strategies he or she would like the system to consider. Requests are processed in a
similar manner, except that when called by the Combiner, instead of backtracking,
the hierarchical planner constructs a plan for the first strategy and then attempts to
add steps to achieve each additional strategy. The final decision on which of these
plans to execute, if any, is left to Ms. Greenlee and her tax advisor.
3.2. CHIRON REJECTS THE SAME PLAN GIVEN WEAKER FACTS
Suppose the input facts are the same as the first example, except that the taxpayer
has already sold her house, the sale occurred four years ago, and she has not yet
254 KATHRYN E. SANDERS
bought a new residence. Since this example also involves a sale of real property,
CHIRON considers the same three strategies as in the first example. Like-kind exchange
is rejected for the same reason as in the first example, because the taxpayer
has occupied the property; and again, the taxpayer is too young to take advantage
of the §121 exclusion. Report-Results also notes two other reasons for rejecting the
§121 exclusion in this case are that the taxpayer only occupied the house during
one of the five years prior to the sale, and that she was away from home for the four
years immediately prior to the sale.
The rollover strategy is also rejected. In the first example, the taxpayer was
occupying the house at the time of sale; here, the taxpayer had not been occupying
the house for four years before the sale. The only case in the case base where the
taxpayer was away from home at the time of sale, Trisko v. Commissioner, 29 T.C.
515 (1957), involved an absence of only three years; this case is even weaker than
Trisko, so the plan is rejected.
3.3. CHIRON RECOMMENDS A PLAN IN SPITE OF WEAKNESSES
Now suppose the input facts are that the taxpayer has been away from home one
year instead of three. Again, CHIRON considers like-kind exchange, rollover, and
the §121 exclusion, and again rejects the like-kind exchange and the exclusion.
Despite the fact that the taxpayer had not lived in the house for a year before the
sale, the system recommends a rollover.
*****************************************************
RESULTS
*****************************************************
The strategy :ROLLOVER is suggested.
Code sections involved in plan: (1034).
CHIRON suggests that the taxpayer buy and occupy another house. It notes that
the taxpayer lived in her house for three years, two months, and twenty-one days
before selling it, and that she had only one residence on the date of sale.
*******************************************
PLAN
*******************************************
((OCCURS (ABSENCE-BEFORE-SALE GREENLEE SELLING163
(1 0 0)) (9 15 1993) (9 15 1993)))
((OCCURS (NUMBER-OF-RESIDENCES GREENLEE 1)
(9 15 1993) (9 15 1993)))
((OCCURS (DURATION-OF-OCCUPANCY GREENLEE 32-ELEVENTH-STREET
(3 2 21)) (1 1 0) (9 15 1993)))
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 255
((OCCURS (BUYER SELLING541 GREENLEE) FT539 FT540))
((OCCURS (PHYSICALLY-OCCUPY GREENLEE UNKNOWN-HOUSE538)
FT536 FT537))
((OCCURS (SALE-TO-PURCHASE-TIME SELLING163 SELLING541
(1 0 7)) (9 15 1993) FT539))
((OCCURS (SALE-TO-OCCUPANCY-TIME SELLING163 SELLING541
(1 0 7)) (9 15 1993) FT536))
((OCCURS (SELLING SELLING541) FT539 FT540))
((OCCURS (OBJECT SELLING541 UNKNOWN-HOUSE538)
FT539 FT540))
*******************************************
Optimistically, the system assumes that the new purchase will occur exactly a
year and seven days after the sale, that is, the day the example was run. We would
need more knowledge about how long it actually takes to buy a house to make
a more accurate guess. If both sale and purchase are in the future, this is not a
problem; and if one of them occurred more than two years ago, the system will
note a weakness in the plan, as it should. For sales that occurred within the past
two years, and particularly over a year ago (so that meeting the time limit may
be impossible), we would like to be able to say “Buy another house within two
years”, and express the fact that the sale-to-purchase time will be between one and
two years, but that would require more sophisticated temporal reasoning than the
system is yet capable of.
Next, CHIRON prints out arguments for and against the success of the suggested
plan. As in the first example, the prototype rollover plan is the only most-on-point
case. The system argues (on behalf of the taxpayer) that the current plan should
succeed like the prototype, because of the facts they have in common.
In response, on behalf of the government, the system admits that the current
situation and the prototype contain the same transactions (two sales of real property
and purchases of real property), but notes that the current situation is weaker than
the prototype, in that the taxpayer, if she executes the suggested plan, would not be
occupying the house at the time of sale.
(4) Weaknesses of the taxpayer’s case compared with
"the rollover prototype":
"the rollover prototype" is distinguishable because
it is stronger for the taxpayer than the current situation
along the shared features:
Current situation:
(OCCURS (ABSENCE-BEFORE-SALE GREENLEE SELLING163
(1 0 0)) (9 15 1993) (9 15 1993))
"the rollover prototype":
256 KATHRYN E. SANDERS
(OCCURS (ABSENCE-BEFORE-SALE GREENLEE SELLING163
(0 0 0)) UT555 UT555)
Next, as in the first example above, CHIRON makes an argument based on Welch
on behalf of the government and rebuts it on behalf of the taxpayer by noting the
differences in transactions. In the first example, the suggested plan was stronger
than Welch along all three of the dimensions where Welch was stronger than the
prototype; here, the suggested plan is stronger than Welch only on one of those
dimensions, sale-to-occupancy-time.
(4) Strengths of the taxpayer’s case compared with "Welch":
"Welch" is distinguishable because
it is stronger for the government than the current situation
along the shared features:
Current situation:
(OCCURS (SALE-TO-OCCUPANCY-TIME SELLING163 SELLING541
(1 0 7)) (9 15 1993) FT536)
"Welch":
(OCCURS (SALE-TO-OCCUPANCY-TIME SELLING1 SELLING2
(1 8 2)) (SEPTEMBER 11 1972) (MAY 13 1974))
Finally, CHIRON warns the user that the suggested plan is weaker than the
prototype.
The current situation is weaker than the prototype
in some respects. There is some support for each
of these weaknesses, however:
(1) The current situation is weaker than the
prototype along the dimension ABSENCE-BEFORE-SALE. In the
current situation
(OCCURS (ABSENCE-BEFORE-SALE GREENLEE SELLING163
(1 0 0)) (9 15 1993) (9 15 1993))
However, the taxpayer has won in the following
cases with a weaker position:
Case: "Trisko v. Commissioner, 29 T.C. 515 (1957)"
Position:
(OCCURS (ABSENCE-BEFORE-SALE TRISKO SELLING1
(3 8 0)) (JUNE 1941) (OCTOBER 10 1951))
*****************************************************
In other words, the system notes that the fact that the taxpayer was not living in
the house at the time of sale raises an issue with respect to this strategy, but there is
case support for that weakness, and gives the user a citation to the supporting case.
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 257
3.4. CHIRON FINDS MORE THAN ONE PLAN TO RECOMMEND
If we change the facts so that the taxpayer qualifies for more than one plan, CHIRON
will recommend both, but may not recommend the combination. For example, if the
input facts are the same as the first example, but the taxpayer is over 55 at the time
she sells her house, CHIRON again rejects the like-kind exchange and recommends
the rollover, for the same reasons as in the first example. The fact that the taxpayer
is older makes no difference to these plans.
Although the new plan is very similar to the §121 prototype, surprisingly, the
system finds most-on-point cases for both sides. For the taxpayer, it finds the
prototype; for the government, a Revenue Ruling (Rev. Rul. 88-29, 1988-1 C.B.
75). Revenue Rulings are cases decided by the Internal Revenue Service after an
administrative hearing. They do not have the precedential effect of a court case,
but they are indicative of the IRS’s position. If they hold in favor of the taxpayer,
they provide useful support. If not, they suggest issues the taxpayer may need to
address.
CHIRON bases its argument on the fact that, in this ruling, the taxpayer won
even though the taxpayer was older, and had owned the property sold longer, than
required by the prototype. The ruling involved a taxpayer in New York City who
sold his rights to a rent-controlled apartment to his landlord and wanted to take the
§121 exclusion for part of the (apparently sizeable) payment, arguing that he had
sold his principal residence. He was sixty years old at the time of sale and still living
in the apartment, where he had lived for 25 years. He had no other residence. The
IRS took the position that these rights, no matter how valuable, were not the kind
of ownership rights required by the statute, and it rejected the taxpayer’s claim.
In rebuttal, CHIRON identifies exactly the right distinction between the current
situation and the Revenue Ruling: they involve the sale of different types of property.
In the current situation, the taxpayer is selling a piece of real property; in the
Revenue Ruling the taxpayer sold intangible rights. This is precisely the rationale
given by the IRS for its decision. Incidentally, this is also the reason why the
Revenue Ruling does not turn up during case-guided search: unlike the current
situation, it does not involve a sale of real estate.
If requested, CHIRON does succeed in constructing a plan that combines both
of these strategies, given these input facts. The system does caution, however, that
there are no cases in the case base where the combination succeeded, and one
(Welch), where it failed.
Here CHIRON is too cautious. If there were several negative cases, the government
might have a strong argument. For example, there are a number of cases
where the taxpayer claims deductions on the grounds that he was away from home
on business, and also claims moving expenses for the return trip. (See, e.g., Goldman
v. Commissioner, 497 F.2d 382 (6th Cir. 1974); Rev. Rul. 74-242, 1974-1
C.B. 69; and Schweighardt v. Commissioner, 54 T.C. 1273 (1970)). In general,
this combination always loses. The combination of rollover and §121, on the other
258 KATHRYN E. SANDERS
hand, was not at all unusual. An elderly taxpayer who wants to sell her house
and move to a smaller place might well have chosen to take advantage of both
strategies.
3.5. CHIRON FAILS TO RECOMMEND A PLAN THAT IT SHOULD RECOMMEND,
AND WHY IT FAILS
To illustrate yet further the problems with changes in the statute, as input, I gave
CHIRON as input the facts of a case, Welch, where the taxpayer lost under an
old version of §1034. Just as in the earlier case, the taxpayer sells a residence
in Minneapolis that she had been occupying for 7 years, buys some land, builds
a new house, and occupies the new house twenty months after the sale of the old
residence. In the meanwhile, like the taxpayer in Welch, she rents a place to live.
To simplify the example, we make one minor change: the taxpayer is 42 and so,
unlike the earlier taxpayer, too young to qualify for the §121 exclusion. This way
the system does not consider the §121 exclusion, and it also rejects the like-kind
exchange on the grounds that the taxpayer has been occupying the property.
A rollover should succeed on these facts; they are no weaker than the prototype,
and under current law they seem quite unproblematic; but CHIRON rejects
the rollover strategy as well. The system correctly identifies Welch as the most-onpoint
case. It would be difficult for any other case to be more on point, since, except
for the taxpayer’s age and the dates, the facts are the same. Because the taxpayer
lost in Welch, the system concludes, the taxpayer will lose in the current situation
as well.
In order to solve this problem correctly, CHIRON would need some way of
reasoning about the dynamic nature of statutes. Welch is still useful as a precedent;
it can be cited, for example, for the proposition that the statutory time limit is
fixed and cannot be stretched. It cannot be cited, as it is here, for the proposition
that a twenty-month delay exceeds the time limit. To capture this aspect of the
relationship between rules and cases, statutes, as well as cases, would need to be
time-stamped, and even then, it would be hard to distinguish between those aspects
of a case that are no longer meaningful and those that are still relevant to the revised
version of the statute. This remains a difficult, but interesting problem for future
work.
4. Related Work
There are no completely “on-point” systems: that is, no other open-textured legal
planners. Nevertheless, CHIRON owes a great deal to previous work. The system’s
method of reasoning about open-textured rules combines ideas from several previous
projects: prototypes from (Bareiss 1988); prototypes and deformations from
(McCarty 1980, 1989a, Schlobohm and McCarty 1989); the use of a set of cases
from (Gardner 1987); the use of cases with concrete facts related by dimensions
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 259
(corporation Chrysler t1)
(issue Chrysler s1 t1)
(stock s1 t1)
(common s1 t1)
(piece-of p1 s1 t1)
(nshares p1 100 t1)
(own Iacocca p1 t1)
Figure 7. How TAXMAN would represent “Iacocca owns 100 shares of Chrysler stock”.
(occurs (corporation Chrysler) t1 t2)
(occurs (issue Chrysler s1) t1 t2)
(occurs (stock s1) t1 t2)
(occurs (common s1) t1 t2)
(occurs (financial-part p1 s1) t1 t2)
(occurs (number-of-shares p1 100) t1 t2)
(occurs (owns Iacocca p1) t1 t2)
Figure 8. How CHIRON would represent “Iacocca owns 100 shares of Chrysler stock”.
from (Ashley 1991; Rissland and Skalak 1989a, b, c; Skalak 1989a, b; Skalak and
Rissland 1992; Rissland et al. 1994); and the use of cases with a rich, detailed
representation from Branting (1990a).
CHIRON’s fact representation language is strongly influenced by TAXMAN (Mc-
Carty 1977) and TAXMAN II (McCarty 1980; McCarty and Sridharan 1982). For
TAXMAN, McCarty designed a representation language that was sufficiently expressive
to state the facts of an input case in detail, yet based on a economical set
of predicates. For example, see how TAXMAN would represent “Iacocca owns 100
shares of Chrysler stock” (Figure 7.) In TAXMAN II, McCarty developed a deontic
logic for representing the concepts of permissions and obligations that occur on
the legal domain. The representation language CHIRON uses for case facts and
useful abstractions from those facts is a modal temporal logic with a formal syntax
and semantics, an extension of the temporal logic developed by Shoham (1988),
modified to incorporate the modal operators “know”, “believe”, “want” (or “goal”),
“obligated”, “permitted”, and “prohibited”. In designing this language, I aimed for
the economy and flexibility of McCarty’s languages.
For purposes of comparison, CHIRON’s representation of the same facts is given
in Figure 8. Aside from the use of “occurs”, the fact that CHIRON’s dates have
starting and ending points (useful when computing duration) and minor variations
in vocabulary such as the use of “owns” rather than “own”, this representation is
very similar to TAXMAN’s.
CHIRON’s hierarchical planner is adapted from a version of Nonlin implemented
in Common Lisp at the University of Maryland (Ghosh et al. 1991). CHIRON’s de-
260 KATHRYN E. SANDERS
composition rules are represented in the same way as Nonlin’s schemas, except that
CHIRON’s slot “steps” is called “expansion" in Nonlin. CHIRON retains the basic
flow of control and architecture of the Maryland Nonlin, but stores most of its data,
including the tasknet, in DUCK, a deductive retriever written at Yale (McDermott
1985). DUCK allows the user to store and retrieve facts, and it supports forward and
backward-chaining. DUCK also provides a clean backtracking mechanism. DUCK
allows the user to create independent but related local databases, known as “datapools”.
Instead of revising a variety of data structures, as the original Maryland
Nonlin did, CHIRON simply establishes a new datapool at each choice point and
when backtracking, pops the most recent datapool off the stack.
In addition, the Maryland Nonlin’s design is modified in two ways to incorporate
the case-based planner: first, the case-based planner is called to help in the
choice of decompositions; and second, the case-based planner is called to suggest
decompositions for certain “open-textured” tasks.
CHIRON incorporates an argument-generation model based on HYPO’s (Ashley
1991). As in HYPO, arguments are based on comparing and contrasting cases using
a subset lattice, although CHIRON uses case facts as the basis for the lattice, where
HYPO uses dimensions. HYPO introduced dimensions as structured objects able
to represent a variety of factors (i.e., stereotypical patterns of facts that tend to
strengthen or weaken a legal conclusion) with different kinds of ranges, binary,
ordered sequences, etc. CHIRON was designed to explore what happens if the case
facts are used for comparison, but a few of its predicates do incorporate domain
knowledge about facts that strengthen or weaken a legal conclusion, and those are
very similar to HYPO’s dimensions.
CHIRON extends HYPO’s model of argumentation slightly by constructing arguments
about proposed strategy combinations, as well as combinations of facts. A
more significant extension is the application of HYPO-style arguments to planning.
Both HYPO and CABARET, the systems that implemented this kind of argumentation
before CHIRON, are legal analysis systems. That is, they take the role of a
trial lawyer, examining the facts of a past situation and building arguments for and
against a particular legal result in that situation. CHIRON is a planner, considering
facts some of which may be past and some are still in the future.
HYPO does not use prototypes, so there is no comparison of the current situation
to a typical one. This is partly necessitated by the difference between their tasks:
since HYPO is taking the role of a trial lawyer, analyzing the facts of a completed
case, the case description can be assumed to be complete. In CHIRON, some of the
facts may be given, but others will be open, so there is a use for the default values
provided by a prototype.
HYPO does, as its name indicates, generate hypothetical cases to explore how
modifications of the facts might change the system’s arguments. To generate hypotheticals,
the system uses a seed case, either the current situation or another
case. This is similar to planning, in that it generates possible cases, but it is more
open-ended. Such a facility might be a very interesting addition to a planner.
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 261
CABARET (Rissland and Skalak 1989a, b, c; Skalak 1989a, b) also used two
equal modules. In that system, however, all the control knowledge was isolated
in a separate control module. For CABARET’s implementation in the tax domain,
its two modules were rule-based and case-based reasoners, but the system was
designed to work with various types of modules, including case-based, rule-based,
and model-based reasoners.
PROLEXS (Oskamp et al. 1989), a legal expert system in the domain of Dutch
landlord-tenant law, also combines rules and cases in a blackboard architecture.
Unlike the other projects described in this section, PROLEXS was designed as a
practical system. It was intended to be used by a law student advising clients in a
legal-aid clinic, and a prototype of the system has been tested in that context. As in
CABARET, there are distinct reasoners for different types of legal knowledge, with
control knowledge isolated in a separate module. Since PROLEXS is based on civil
law, however, its “cases” seem more like rules and are applied like rules, being
used if relevant rather than being compared and contrasted with each other.
In planning, the two lines of work that are most closely related to CHIRON
are hierarchical, nonlinear planning (Sacerdoti 1974, Tate 1976, Tate 1977) and
case-based planning (Hanks and Weld 1992; McCartney and Wurst 1991; Hinrichs
1991; Kambhampati and Hendler 1989; Hammond 1986; Alterman 1986). Hierarchical
planning provides a way of reasoning with the rules on which plans are
based, but is insufficient in this domain, since schemas based on legal rules “run
out”: the most concrete partial plans that can be generated by the planner still contain
one or more abstract, nonexecutable steps, corresponding to the open-textured
provisions of the rules. Case-based planning is also insufficient: it provides a way of
reasoning with legal cases and clarifies many of the issues involved in case representation,
indexing, and adaptation, but does not capture the way that rules structure
and organize the case base. In CHIRON, the hierarchical planner reasons with representations
of open-textured rules and facts; the case-based planner bridges the
gap between the two, as legal cases bridge the gap between open-textured rules
and facts.
CHIRON’s case-guided search is similar to the case-guided search used in (Veloso
1992; Kambhampati and Hendler 1989), except that in CHIRON, cases are used
as the basis for a solution to the current problem, as well as for search control. In the
context of this project, Kambhampati and Hendler suggested a general criterion for
adaptations: that efficient adaptation strategies (which they call “refitting”) should
be conservative: they should change the old plan only as much as necessary to fit
it to the new situation (Kambhampati and Hendler 1989). This is similar to the
conservative bias found in CHIRON.
Finally, CHIRON also shares many ideas with the design for a legal planning
system sketched out by McCarty in joint work with Dean Schlobohm, an estate
planning attorney (Schlobohm and McCarty 1989). They argue that lawyers
construct plans by retrieving prototype plans and transforming them to meet the
clients’ goals. They discuss how trusts and the Internal Revenue Code can be
262 KATHRYN E. SANDERS
represented using McCarty’s representation language. Unlike CHIRON’s design,
this design does not incorporate any explicit facility for reasoning with cases.
5. Conclusions
What has been learned so far from this project? One way to reflect on this question
is to consider what it would take to reimplement the system in a new domain.
Suppose, for example, we wanted CHIRON to construct plans in the bankruptcy
domain explored in BankXX (Rissland et al. 1994). United States law offers individuals
two means of declaring bankruptcy. Roughly speaking, in the first type
of bankruptcy, the debtor is required to liquidate all of his or her property; in
the second type, known as Chapter 13 bankruptcy after the portion of the law
where it is defined, the debtor proposes a plan to pay over a period of 3–5 years.
BankXX uses heuristics and best-first search, in addition to HYPO-style case-based
techniques, to gather cases, legal theories, stereotypical legal stories, etc. on the
question of whether Chapter 13 personal bankruptcy plans are “proposed in good
faith” (Rissland et al. 1994).
The new bankruptcy CHIRON (BankCH) would have the same basic design and
flow of control as CHIRON. The major change would be in the knowledge base,
and the major task involved in re-implementation would be knowledge acquisition.
Much of CHIRON’s vocabulary concerning property ownership, sales, income, and
debts, could be reused in this domain. It would be necessary to revise the toplevel
goal (currently reduce-taxes), analyze the bankruptcy law to determine which
aspects of the cases to use in case-guided search, and build the domain-knowledge
module about which features and dimensions strengthen or weaken a case (though
here we could build on some of the work already done for BankXX).
One of the lessons learned in building CHIRON, although not a new or surprising
lesson, is that a richer knowledge representation has both advantages and disadvantages.
CHIRON’s algorithms for indexing, adapting, comparing, and contrasting
cases benefit from its rich fact representation. For example, it was able to distinguish
cases involving the sale of a house and the sale of rights to a rent-controlled
apartment on the basis that the latter involves a sale of intangible property rights.
However, the detailed case representation makes knowledge acquisition timeconsuming
and difficult. CHIRON’s 24 cases took several hours apiece to represent
by hand, and BankCH’s cases would be comparable. Representing the rules and
domain knowledge take a substantial amount of time as well. A feature-vector representation
makes it easier to translate from case report to representation. Second,
a richer representation language makes it harder to enforce consistency in the case
representations.
A detailed, flexible case representation significantly increases the computational
cost of matching cases. It remains simpler than full subgraph isomorphism (which
is NP-complete (Garey and Johnson 1979)), but still more complex than the cost of
comparing feature-vector-based cases. However, CHIRON’s response times (typic-
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 263
ally three to five minutes per problem) are tolerable and, as reported in (Sanders et
al. 1997), preliminary results using a synthetic casebase indicate that representing
structured cases on a parallel machine significantly improves the time required for
matching and reasoning with structured cases.
This project illustrates the difficulty of reasoning with cases and rules in a domain
where both are changing. When the statute changes, some aspects of previous
case law may remain relevant, but others do not. The system would benefit from an
ability to reason about dynamic legal systems; but that, as noted by both Gardner
(1989) and Bratley et al. (1991), is a general problem with legal reasoning systems.
Finally, the project illustrates a way of constructing plans from representations
of rules and cases, where the cases extend, limit, and partially define the rules, the
system distinguishes between conservative and aggressive plans, and considerable
control is given to the expert user.
Acknowledgements
The author gratefully acknowledges the criticisms and encouragement of Eugene
Charniak, Tom Dean, Jim Hendler, Leslie Kaelbling, Robert McCartney, Leora
Morgenstern, Edwina Rissland, David Skalak, and the members of the PLUS
group at the University of Maryland. The reviewers’ comments were thorough,
thought-provoking and helpful. This project was part of my thesis work, which was
supported in part by National Science Foundation Presidential Young Investigator
Award IRI-8957601 to Thomas Dean, by the Air Force and the Advanced Research
Projects Agency of the Department of Defense under Contract No. F30602-91-C-
0041, by the Office of Naval Research and the Advanced Research Projects Agency
under grant ONR N00014-91-J-4052, ARPA Order 8225, by the National Science
Foundation in conjunction with the Advanced Research Projects Agency of the Department
of Defense under Contract No. IRI-8905436, by IBM grants 17290066,
17291066, 17292066, and 17293066, and by National Science Foundation grant
IRI-8801253.
264 KATHRYN E. SANDERS
Appendix
A. CHIRON’s predicates
absence-before-sale age
apartment area
bank bathroom
building buyer
car cash
condominium cooperative
corporation distance
dormitory duration-of-occupancy
duration-of-ownership duration-of-rental
121-elections-before-sale elementary-school
employment employer
employee farm
files financial-part
floortype government
half-bath house
individual individual-return
intangible-personal-property joint-return
kitchen labor
laboratory lawyer
lessee lessor
location maintain
near number-of-bathrooms
number-of-residences object
121-occupancy-time 121-ownership-time
older owns
parent payment
physically-occupy private-school
public-school real-property
rent-controlled renting
research room
sale-to-occupancy-time sale-to-purchase-time
seller selling
services signs
spatial-part student
tangible-personal-property tax-exempt
taxpayer tenants-rights
trailer university
war-veteran
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 265
References
Alterman, R. (1986). An Adaptive Planner, Proceedings of the Fifth National Conference on Artificial
Intelligence, Philadelphia, Pennsylvania, (reprinted in Proceedings of aWorkshop on Case-Based
Reasoning, 1988, and Readings in Planning, edited by James Allen, James Hendler, and Austin
Tate, Morgan Kaufman, 1990),
Ashley, K. D. (1988). Modelling Legal Argument: Reasoning with Cases and Hypotheticals,
Technical Report 88-01, University of Massachusetts, Amherst, Department of Computer and
Information Science. (Ph.D. Thesis).
Ashley, K. D. (1991). Modelling Legal Argument: Reasoning with Cases and Hypotheticals, MIT
Press.
Ashley, K. D. and Rissland, E. L. (1987). Compare and Contrast, A Test of Expertise. In Proceedings
of the Sixth National Conference on Artificial Intelligence, 273–278, Seattle, Washington.
Bareiss, R. (1988). Protos: A Unified Approach to Concept Representation, Classification, and
Learning, Technical Report AI88-83, University of Texas at Austin. (Ph.D. Thesis).
Bench-Capon, T. and Sergot, M. J. (1988). Towards a Rule-Based Representation of Open Texture in
Law. In Computer Power and Legal Language, 39–60, Quorum Books: New York.
Berman, D. H. and Hafner, C. D. (1991). Incorporating Procedural Context into a Model of
Case-Based Legal Reasoning, Proceedings of the Third International Conference on Artificial
Intelligence and Law, 12–20, Oxford.
Branting, L. K. (1988). The Role of Explanations in Reasoning from Legal Precedents, Proceedings
of a Workshop on Case-Based Reasoning, 94–103.
Branting, L. K. (1989a). Integrating Generalizations with Exemplar-Based Reasoning, Proceedings
of a Workshop on Case-Based Reasoning, 224–229.
Branting, L. K. (1989b). Integrating Generalizations with Exemplar-Based Reasoning, Proceedings
of the Eleventh Annual Conference of the Cognitive Science Society, 139–146, Ann Arbor:
Michigan.
Branting, L. K. (1989c). Representing and Reusing Explanations of Legal Precedents. In Proceedings
of the Second International Conference on Artificial Intelligence and Law, 103–110, Vancouver,
British Columbia.
Branting, L. K. (1990a). Integrating Rules and Precedents for Classification and Explanation:
Automating Legal Analysis. Technical Report AI90-146, Artificial Intelligence Laboratory,
Department of Computer Sciences, University of Texas at Austin. (Ph.D. Thesis).
Branting, L. K. (1990b). Personal communication.
Branting, L. K. (1994). Personal communication.
Bratley, P., Fremont, J., Mackaay, E., and Poulin, D. (1990). Coping with Change. In Proceedings of
the Third International Conference on Artificial Intelligence and Law, 69–76, Oxford, England.
Chirelstein M. A. (1988). Federal Income Taxation: A Law Student’s Guide to the Leading Cases
and Concepts, Foundation Press: Westbury, NY.
Cuthill, B.: (1992). Situation Analysis, Precedent Retrieval, and Cross-Context Reminding in Case-
Based Reasoning, Technical Report CE-TR-92-3, University of Connecticut Department of
Computer Science and Engineering. (Ph.D. Thesis).
Cuthill, B. and McCartney, R. (1993). Issue Spotting in Legal Cases. In Proceedings of the Fourth
International Conference on Artificial Intelligence and Law, 245–253. Amsterdam.
Fikes, R. and Nilsson, N. J. (1971). STRIPS: A New Approach to the Application of Theorem Proving
to Problem Solving. Artificial Intelligence 2: 189–208.
Gardner, A.: (1987). An Artificial Intelligence Approach to Legal Reasoning, MIT Press: Cambridge,
MA.
Gardner, A. (1989). Representing Developing Legal Doctrine. In Proceedings of the Second International
Conference on Artificial Intelligence and Law, 16–22. Vancouver, British Columbia.
266 KATHRYN E. SANDERS
Garey, M. R. and Johnson, D. S. (1979). Computers and Intractability,W. H. Freeman and Company:
New York.
Ghosh, S., Hendler, J., Kambhampati, S., and Kettler, B.: 1991, Common Lisp Implementation of
NONLIN USER MANUAL. Computer Science Department, University of Maryland, College
Park, Maryland.
Golding, A. R. and Rosenbloom, P. S. (1991). Improving Rule-Based Systems Through Case-Based
Reasoning, In Proceedings of the Ninth National Conference on Artificial Intelligence, 22–27.
Anaheim: California.
Gordon, T. F. (1993). The Pleadings Game: An Artificial Intelligence Model of Procedural Justice,
Fachbereich Informatik der Technischen Hochschule, Darmstadt. (Ph.D. Thesis).
Hafner, C. D. (1978). An Information Retrieval System Based on a Computer Model of Legal
Knowledge. Ph.D. Thesis, University of Michigan.
Hammond, K. J. (1986). Case-Based Planning: An Integrated Theory of Planning, Learning, and
Memory. Technical Report YALEU/CSD/RR 488, Yale University Department of Computer
Science. (Ph.D. Thesis).
Hammond, K. J. (1989). On Functionally Motivated Vocabularies: An Apologia. In Proceedings of a
Workshop on Case-Based Reasoning, 52–56.
Hanks, S. and Weld, D. S. (1992). Systematic Adaptation for Case-Based Planning. In Proceedings
of the First International Conference on Artificial Intelligence Planning Systems, 96–105.
Hart, H. L. A. (1961). The Concept of Law, Clarendon Press: Oxford.
Hendler, J. A.: 1995, Experimental AI Systems. Journal of Experimental and Theoretical Artificial
Intelligence 7: 1–5. (editorial).
Hinrichs, T. R. (1991). Problem Solving in Open Worlds: A Case Study in Design. Technical Report
GIT-CC-91/36, Georgia Institute of Technology. (Ph.D. Thesis).
Hinrichs, T. R. and Kolodner, J. (1991). The Roles of Adaptation in Case-Based Design. In
Proceedings of the Ninth National Conference on Artificial Intelligence. Anaheim: California.
Hughes, G. and Cresswell, M. J. (1968). An Introduction to Modal Logic, Methuen: London.
Kambhampati, S. and Hendler, J. A. (1989). Control of Refitting During Plan Reuse. In Proceedings
of the Eleventh International Joint Conference on Artificial Intelligence, 943–948. Detroit,
Michigan.
Klein, G. R. and Calderwood, R. (1988). How Do People Use Analogues to Make Decisions? In
Proceedings of a Workshop on Case-Based Reasoning, 209–223.
Lakatos, I.: Proofs and Refutations: The Logic of Mathematical Discovery. Cambridge University
Press: Cambridge, England.
Lakoff, G. (1987). Women, Fire, and Dangerous Things: What Categories Reveal About the Mind,
University of Chicago Press, Chicago, IL.
Levi, E. H. (1949). An Introduction to Legal Reasoning, University of Chicago Press, Chicago, IL.
Lifschitz, V. (1987). Formal Theories of Action. In Proceedings of the 1987 Workshop on the Frame
Problem in Artificial Intelligence.
Llewellyn, K. N. (1930). The Bramble Bush: Our Law and Its Study. Oceana Publications, Dobbs
Ferry: New York.
Loui, R. P., Norman, J., Olson, J., and Merrill, A. (1993). A Design for Reasoning with Policies,
Precedents, and Rationales. In Proceedings of the Fourth International Conference on Artificial
Intelligence and Law, 202–211, Amsterdam.
McCartney, R. (1993). Planning from Partial and Multiple Episodes in a Case-Based Planner. In
Proceedings of the 1993 AAAIWorkshop on Case-Based Reasoning, 94–100. Washington, D.C.
McCartney, R. and Sanders, K. E. (1990). The Case for Cases: A Call for Purity in Case-Based
Reasoning. In Proceedings of the AAAI Symposium on Case-Based Reasoning, 12–16.
McCartney, R. and Wurst, K. R. (1991). DEFARGE: A Real-Time Execution Monitor for a Case-
Based Planner. In Proceedings of the Workshop on Case-Based Reasoning, 233–244.
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 267
McCarty, L. T. (1977). Reflections on TAXMAN: An Experiment in Artificial Intelligence and Legal
Reasoning, Harvard Law Review 90: 837–893.
McCarty, L. T. (1980). The TAXMAN Project: Towards a Cognitive Theory of Legal Argument. In
B. Niblett (ed.), Computer Science and Law, 23–43. Cambridge University Press: Cambridge,
England.
McCarty, L. T. (1983). Permissions and Obligations, Proceedings of the Eighth International Joint
Conference on Artificial Intelligence, 287–294. Karlsruhe, West Germany.
McCarty, L. T. (1986). Permissions and Obligations: An Informal Introduction, Technical Report
LRP-TR-19, Laboratory for Computer Science Research, Rutgers University.
McCarty, L. T. (1989a). Computing with Prototypes (Preliminary Report). In Proceedings of the
Bar-Ilan Symposium on the Foundations of Artificial Intelligence.
McCarty, L. T. (1989b). A Language for Legal Discourse: I. Basic Features. In Proceedings of the
Second International Conference on Artificial Intelligence and Law, 180–189. Vancouver, British
Columbia.
McCarty, L. T. and Sridharan, N. S. (1982). A Computational Theory of Legal Argument, Technical
Report LRP-TR-13, Laboratory for Computer Science Research, Rutgers University.
McDermott, D. (1985). The DUCK Manual, Technical Report YALEU/CSD/RR 399, Yale University
Department of Computer Science.
Oskamp, A., Walker, R. F., Schrickx, J. A., and van den Berg, P. H. (1989). Prolexs, Divide and
Rule: A Legal Application. In Proceedings of the Second International Conference on Artificial
Intelligence and Law, 54–62. Vancouver, British Columbia.
Redmond, M. (1990). Distributed Cases for Case-Based Reasoning; Facilitating Use of Multiple
Cases. In Proceedings of the Eighth National Conference on Artificial Intelligence, 304–309.
Boston, MA.
Riesbeck, C. K. and Schank, R. C. (1989). Inside Case-based Reasoning, Lawrence Erlbaum
Associates: Hillsdale, NJ.
Rissland (Michener), E. (1978). Epistemology, Representation, Understanding and Interactive Exploration
of Mathematical Theories, Department of Mathematics, Massachusetts Institute of
Technology. (Ph.D. Thesis).
Rissland, E. L. (1982). Examples in the Legal Domain: Hypotheticals in Contract Law, Proceedings
of the Fourth Annual Conference on Cognitive Science, Ann Arbor, Michigan.
Rissland, E. L., Daniels, J. J., Rubinstein, Z. B., and Skalak, D. B. (1993a). Case-Based Diagnostic
Analysis in a Blackboard Architecture, Proceedings of the Eleventh National Conference on
Artificial Intelligence, Washington, D.C.
Rissland, E. L., Daniels, J. J., Rubinstein, Z. B., and Skalak, D. B. (1993b). Diagnostic Case Retrieval
Guided by Evaluation and Feedback, Proceedings of the AAAI Workshop on Case-Based
Reasoning, Washington, D.C.
Rissland, E. L., Skalak, D. B., and Friedman, M. T. (1994). Heuristic Harvesting of Information
for Case-Based Argument, Proceedings of the Twelfth National Conference on Artificial
Intelligence, 36–43. Seattle, Washington.
Rissland, E. L. and Soloway, E.M. (1980). Overview of an Example Generation System, Proceedings
of the First National Conference on Artificial Intelligence, 256–258.
Rissland, E. L. and Skalak, D. B. (1989a). Case-Based Reasoning in a Rule-Governed Domain, Proceedings
of the Fifth IEEE Conference on Artificial Intelligence Applications, Miami, Florida,
45–53.
Rissland, E. L. and Skalak, D. B. (1989b). Interpreting Statutory Predicates, Proceedings of
the Second International Conference on Artificial Intelligence and Law, Vancouver, British
Columbia, 46–53.
Rissland, E. L. and Skalak, D. B. (1989c). Combining Case-Based and Rule-Based Reasoning: A
Heuristic Approach, Proceedings of the Eleventh International Joint Conference on Artificial
Intelligence, 524–530. Detroit, Michigan.
268 KATHRYN E. SANDERS
Rissland, E. L. and Skalak, D. B. (1991). CABARET: Rule Interpretation in a Hybrid Architecture.
International Journal of Man-Machine Studies, 839–887.
Rosch, E. and Mervis, C. B. (1975). Family Resemblance: Studies in the Internal Structure of
Categories. Cognitive Psychology 7: 573–605.
Sacerdoti, E. (1974). Planning in a Hierarchy of Abstraction Spaces. Artificial Intelligence 7: 231–
272.
Sanders, K. E. (1989a). A Logic for Emotions: A Basis for Reasoning about Commonsense Psychological
Knowledge, Proceedings of the Eleventh Annual Conference of the Cognitive Science
Society, Ann Arbor, Michigan.
Sanders, K. E. (1989b). A Logic for Emotions: A Basis for Reasoning about Commonsense
Psychological Knowledge, Brown University Computer Science Department, 89–23.
Sanders, K. E. (1990). The Truth, the Whole Truth, and Nothing but the Truth: Case Representations
for Legal Reasoning Systems, Proceedings of the AAAI Workshop on Legal Reasoning, Boston,
MA.
Sanders, K. E. (1991a). Representing and Reasoning about Open-Textured Predicates, Proceedings
of the Third International Conference on Artificial Intelligence and Law, Oxford, England, 137–
144.
Sanders, K. E. (1991b). Within the Letter of the Law: Planning Among Multiple Cases, Proceedings
of the DARPA Workshop on Case-Based Reasoning, Washington, D.C.
Sanders, K. E. (1993). If It Worked Last Time, It Should Work Again: Justifying and Validating
Case-Based Plans, Proceedings of the AAAI Workshop on Case-Based Reasoning.
Sanders, K. E. (1994). CHIRON: Planning in an Open-Textured Domain, Computer Science
Department, Brown University, 94–38. (Ph.D. Thesis).
Sanders, K. E., Kettler, B. P., and Hendler, J. A. (1997). The Case for Structure-Based Representations,
Proceedings of the Second International Conference on Case-Based Reasoning,
Providence, Rhode Island.
Schlobohm, D. and McCarty, L. T. (1989). EPS II: Estate Planning with Prototypes, Proceedings of
the Second International Conference on Artificial Intelligence and Law, 1–10. Vancouver, British
Columbia.
Sergot, M. J., Sadri, F., Kowalski, R. A., Kriwaczek, F., Hammond, P., and Cory, H. T. (1986). The
British Nationality Act as a Logic Program. Communications of the Association for Computing
Machinery 29(5): 370–386.
Shoham, Y. (1988). Reasoning about Change: Time and Causation from the Standpoint of Artificial
Intelligence, MIT Press: Cambridge, MA.
Skalak, D. B. (1989a). Options for ControllingMixed Paradigm Systems, Proceedings of aWorkshop
on Case-Based Reasoning, 318–323.
Skalak, D. B. (1989b). Taking Advantage of Models for Legal Classification, Proceedings of the
Second International Conference on Artificial Intelligence and Law, Vancouver, 234–241. British
Columbia.
Skalak, D. B. and Rissland, E. L. (1992). Arguments and Cases: An Inevitable Intertwining. Artificial
Intelligence and Law: An International Journal 1(1), 3–44.
Statsky, W. P. and Wernet, Jr., R. J. (1989). Case Analysis and Fundamentals of Legal Writing, West
Publishing Company: St. Paul, MN.
Tate, A. (1976). Project Planning Using a Hierarchic Non-Linear Planner, Technical Report 245,
Department of Artificial Intelligence, University of Edinburgh.
Tate, A. (1977). Generating Project Networks, Proceedings of the Fifth International Joint Conference
on Artificial Intelligence, 888–893.
Veloso, M. M. (1992). Learning by Analogical Reasoning in General Problem Solving, Technical
Report CMU-CS-92-174, School of Computer Science, Carnegie Mellon University (Ph.D.
Thesis).
CHIRON: PLANNING IN AN OPEN-TEXTURED DOMAIN 269
Waismann, F. (1965). Verifiability. In A. Flew (ed.), Logic and Language: First and Second Series,
122–151. Anchor Books: Garden City, NJ. (First Published in Proceedings of the Aristotelian
Society, 119-150 (1945).)
Waltz, D. L. (1989). Is Indexing Used for Retrieval?, Proceedings of a Workshop on Case-Based
Reasoning, 41–44.
Waterman, D. A. (1986). A Guide to Expert Systems, Addison-Wesley Publishing Company:
Reading, MA.
Zito-Wolf, R. and Alterman, R. (1992). Multicases: A Case-Based Representation for Procedural
Knowledge, Proceedings of the Fourteenth Annual Conference of the Cognitive Science Society.
Artificial Intelligence and Law 9: 225–269, 2001.
© 2002 Kluwer Academic Publishers. Printed in the Netherlands.
225