Introduction
This 1987 classic from the AAAI proceedings explores the intersection of Artificial Intelligence and legal reasoning. Authors Kevin D. Ashley and Edwina L. Rissland introduce HYPO, a system designed to move beyond simple rule-based logic by embracing the adversarial, case-heavy nature of the law.
The Challenge of "Open-Textured" Domains
In the late 1980s, most expert systems relied on rigid rule-based architectures. However, the authors argue that true expertise—especially in fields like law, strategic planning, and ethics—isn't just about following rules; it’s about the ability to compare a new problem to prior experiences.
Legal domains are "open-textured," meaning rules are often incomplete, vague, or lead to inconsistent results. In trade secret law, for example, there isn't a simple formula to decide if a secret was "misappropriated." Instead, lawyers look for "on-point" cases to justify their position or distinguish away damaging precedents. The challenge was building a system that could perform this dynamic indexing and relevancy assessment.
Methodology: Beyond If-Then Rules
The authors developed several key mechanisms to enable HYPO to reason like a lawyer:
- Dimensions: Rather than simple keywords, HYPO uses "dimensions" to represent legal relationships. A dimension (like Disclose-Secrets) links a cluster of facts to a legal conclusion. It includes "focal slots" (e.g., the number of people a secret was disclosed to) that determine how strongly the dimension favors the plaintiff or defendant.
- Case-Analysis-Record: When a new case (the "current fact situation" or cfs) is entered, HYPO generates a record of applicable dimensions and "near-misses"—dimensions where most, but not all, prerequisites are met.
- Claim Lattice: This is the system's core data structure. It is a rooted graph that organizes cases from the Case Knowledge-Base (CKB) based on how many dimensions they share with the cfs. Cases closer to the root are "Most-On-Point Cases" (MOPCs).
- 3-Ply Arguments: HYPO doesn't just output a "win/loss" prediction. It generates a three-part argument: a Point (citing a pro-side case), a Response (distinguishing that case or citing a counter-precedent), and a Rebuttal (responding to the distinction or the counter-precedent).
Key Insights: Relevance as Argument
- Relevance is Argumentative: In Case-Based Reasoning (CBR), a case isn't "relevant" just because it's mathematically similar; it's relevant if it can be used to build a persuasive argument.
- The Power of "Near-Misses": HYPO uses near-miss dimensions to suggest "hypotheticals." By asking "what if the plaintiff had required a non-disclosure agreement?", the system can find cases that would have been relevant, helping a user understand the strengths and weaknesses of their position.
- Adversarial Reasoning: The paper posits that CBR is inherently adversarial. There is rarely one "right" answer; instead, there are better and worse justifications. HYPO’s success is measured by its ability to present these alternatives.
Results and Evaluation
The authors validated HYPO by comparing its output to actual court opinions. In a test case based on Crown Industries, Inc. v. Kawneer Co., HYPO correctly identified Midland-Ross as the defendant's best case. The system's generated reasoning closely mirrored the actual judge's logic: both concluded that the plaintiff's sale of the product constituted a public disclosure that defeated the trade secret claim.
Furthermore, HYPO was able to generate "hypothetical variants" of the case to show exactly how the plaintiff could have strengthened their legal standing. As shown in the figure above, adding a dimension like "Disclosure Restrictions" changes the claim lattice, promoting different cases to "Most-On-Point" status and altering the argumentative landscape.
Limitations and Future Work
The authors acknowledge several hurdles for scaling this approach, including high indexing costs for constructing the library of dimensions and the depth of inferencing required to navigate the claim lattice as the knowledge base grows. Future work suggested by the paper includes improving search efficiency and refining the "evaluation functions" used to compare different claim lattices.
Summary updated