Abstract
Legal reasoning involves balancing conflicting factors, a task traditionally handled in AI through numerical weighting. This paper presents HYPO, a case-based reasoning (CBR) system that adopts a "least commitment" approach to avoid premature and unjustifiable numerical assignments. By utilizing symbolic dimensions and a claim-lattice structure, HYPO identifies most-on-point precedents and generates adversarial arguments. We demonstrate how HYPO performs sensitivity analysis via boundary cases, providing a model for AI that aligns with the qualitative nature of legal advocacy.
Introduction
In complex domains like law, decisions depend on multiple, often conflicting factors. Traditional AI approaches often rely on linear evaluation functions, assigning numerical weights to factors to calculate a final score. However, this approach is fundamentally flawed for legal reasoning for three primary reasons:
- Context Dependency: The importance of a factor changes based on the presence of others.
- Lack of Justification: Numerical scores lack the rhetorical power of citing specific precedents in a courtroom.
- Premature Commitment: Fixing weights early prevents the exploration of alternative lines of reasoning that may emerge later.
The HYPO system addresses these issues by postponing the "weighting" of factors, instead using a symbolic "least commitment" approach to compare past precedents.
Methodology: The HYPO System
HYPO operates in the domain of trade secrets law, representing cases through "dimensions"—factual features with legal significance (e.g., the number of disclosures). The system follows a three-phase process:
1. Clustering and the Claim-Lattice
HYPO organizes the Case Knowledge Base (CKB) into a "claim-lattice." This structure allows the system to identify Most-On-Point (MOP) cases—precedents that share the maximum number of factual dimensions with the current problem situation.
2. Interpreting Outcomes
Once MOP cases are identified, HYPO analyzes their outcomes. If a cluster of factors consistently leads to a win for the plaintiff, it is interpreted as favoring that side. If outcomes are split, the system identifies a "tie" requiring deeper analysis.
3. Criticizing and Testing
HYPO challenges its own interpretations by searching for counter-examples. This includes "trumping cases" (cases with similar factors but an additional one that flips the result) and "boundary cases" where factors are at extremes.
Results: Symbolic Sensitivity Analysis
HYPO demonstrates that sensitivity analysis can be performed without numerical weights. By examining dimensions, the system can identify how changes in factual magnitude affect legal outcomes. A key example is the Data General case. In a dispute involving disclosures to outsiders, HYPO can cite Data General as a boundary counter-example: even though the plaintiff disclosed secrets to 6,000 people, they still won because those disclosures were protected by confidentiality agreements.
By using this symbolic approach, HYPO generates arguments grounded in case law rather than abstract probabilities, allowing for more persuasive legal advocacy.
Conclusion
HYPO shifts the focus of legal AI from adjudication (deciding a winner) to advocacy (generating the best possible arguments). While a final decision might eventually require a commitment to weighting, HYPO demonstrates that the most critical phase of legal reasoning—the comparison and interpretation of precedents—is best served by a symbolic, least-commitment approach.
Summary updated