Defining Salience in Case-Based Arguments

Kevin D. Ashley

Introduction

In the field of artificial intelligence and law, a persistent challenge has been the development of systems capable of sophisticated reasoning rather than simple information retrieval. Early knowledge-based systems often suffered from a "static description" problem: when asked to describe an object or a case, they would provide a fixed list of attributes regardless of the context. However, human experts—particularly lawyers—demonstrate a dynamic ability to tailor descriptions to suit specific argumentative goals. This paper explores the computational definition of "salience," the process by which a system identifies and emphasizes the most relevant features of a case to support a particular viewpoint.

The Problem of Static Knowledge Bases

In legal discourse, an attorney does not merely list facts; they strategically highlight certain aspects of a case while downplaying others. A single past precedent can be framed as a binding authority by one side and an irrelevant outlier by the other. For an AI to argue effectively, it must move beyond static representations. The core research question addressed is: How can a computer program dynamically determine which features of a case are salient based on the current argument's context and the arguer's perspective?

Methodology: The HYPO System

The author utilizes HYPO, an adversarial case-based reasoning (CBR) program designed to operate within the domain of trade secret law. Unlike modern connectionist models, HYPO employs symbolic logic and set theory to navigate a Case Knowledge Base (CKB).

The system's reasoning is built upon two primary constructs:

  • Factors: These are generalized collections of facts that typically favor either the plaintiff or the defendant. HYPO implements thirteen specific factors (e.g., "secrets disclosed to outsiders").
  • Dimensions: These represent the magnitude or "strength" of a factor. A dimension allows the system to reason about degrees—for instance, comparing a case with 10 disclosures to one with 10,000.

Contextual Roles and Salience

HYPO defines salience not as an inherent property of a fact, but as a function of the case's role in an argument. The paper identifies four primary roles:

  1. Cited Case: A case used as authority to support a point. Here, salient features are the similarities between the cited case and the current problem.
  2. Distinguished Case: A case cited by an opponent that must be neutralized. Salience here lies in the "relevant differences"—factors present in the problem but missing in the cited case, or vice versa.
  3. Counterexample: A case with similar facts but an opposite legal outcome. Salience is defined by the similarities that make the counterexample "on-point" and the differences that explain the divergent outcome.
  4. Target Case: A case used to suggest hypothetical "what-if" scenarios. Salience involves identifying features that, if changed, would significantly alter the strength of the argument.

Data and Implementation

The framework was tested using a CKB of 30 actual legal cases involving trade secret misappropriation, such as Data General v. Digital Computer and Crown Industries. While the dataset is small by modern standards, each case was meticulously hand-curated to include the complex symbolic relationships necessary for legal reasoning.

Key Insights: The Dynamics of Argument

A central insight of this work is that the absence of a factor can be just as salient as its presence. In legal distinguishing, if a precedent had a pro-plaintiff factor that the current case lacks, that "non-feature" becomes a primary point of contention for the defense. Furthermore, salience is sensitive to magnitude; HYPO recognizes that a case involving fewer disclosures is a "stronger" precedent for a plaintiff than one with many, and it adjusts its descriptions to reflect this "slant."

Results and Evaluation

The paper demonstrates HYPO's effectiveness through qualitative analysis of its generated summaries. For the Data General case, the system can produce four distinct descriptions:

  • As a Plaintiff's Cite: Emphasizing that the plaintiff prevailed despite some disclosures.
  • As a Defendant's Distinction: Highlighting that the disclosures in Data General were more restricted than in the current case.
  • As a Counterexample: Using it to "trump" weaker precedents by showing it is more "on point."
  • As a Target for Hypotheticals: Using it as a model for how the plaintiff's position could be improved.

The author notes that these generated descriptions compare favorably with the reasoning found in actual court opinions, suggesting that the symbolic definitions of salience accurately capture a core component of human legal expertise.

Conclusion and Future Work

While HYPO's logic for selecting salient features is sophisticated, its natural language generation remains primitive, relying on "canned" phrases. Future work involves integrating this salience framework with more advanced natural language processors. Additionally, while the current model is specific to trade secret law, the underlying principles of contextual salience are potentially applicable to broader domains of adversarial reasoning and general discourse.

Summary updated

BibTeX
@inproceedings{ashley1989defining,
  author = {Kevin D. Ashley},
  title = {{Defining Salience in Case-Based Arguments}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1989},
}
Figure