Introduction
In legal reasoning, classification is not merely a labeling task but a process of justification. Automated systems often encounter a "generality gap": the disconnect between abstract legal rules and the specific facts of real-world cases. While rules provide general definitions—for instance, "negligence requires a breach of reasonable care"—they often fail to specify what constitutes "reasonable care" in a particular context. Conversely, precedents offer concrete examples but are difficult to apply to new cases with differing factual patterns.
The GREBE (GeneratoR of Exemplar-Based Explanations) system addresses this by treating rules and precedents as "impartial" equals. Rather than imposing a hierarchy, GREBE navigates the "open-textured" nature of legal language by simultaneously back-chaining through both rules and cases to construct the most persuasive argument.
Methodology: The GREBE System
GREBE employs an impartial control strategy to find the strongest "warrant" (justification) for a classification. It utilizes three primary mechanisms to bridge the generality gap:
- Term Reformulation: The system uses rules to substitute an abstract term for another that has a more direct match in the precedent database.
- Case Elaboration: GREBE uses rules to infer missing facts in a new case, making it functionally similar to a known precedent. For instance, it might infer that "sandwiches" in a current case serve the same role as "ice water" in a past case regarding essential items for employment.
- Structure Matching: This technique maps entity relationships between new and past cases to determine structural similarity, ensuring that the logic of the precedent applies to the current facts.
Experimental Evaluation
The system was tested on 18 hypothetical worker's compensation cases based on Texas law. These "hypos" were designed to test reasoning depth rather than simple pattern matching. The knowledge base included statutory rules and actual legal precedents, such as Janak v. Texas Employer’s Ins. Co.
To benchmark performance, GREBE's outputs were compared against those of five law students from the University of Texas. An experienced attorney graded all responses blindly on a 4-point scale (0-4).
Results and Discussion
The evaluation demonstrated that GREBE performs at a level comparable to, and in some metrics superior to, human law students.
As shown in the results, GREBE achieved a higher overall score (2.00, equivalent to a 'C') compared to the students (1.78, equivalent to a 'C-'). GREBE significantly outperformed students in "Explanation Quality" (2.22 vs. 1.78), suggesting that its ability to systematically assemble rules and precedents into a coherent argument provides a more robust justification than manual student analysis. Furthermore, GREBE received four 'B' grades, while the students collectively received only one.
Conclusion
GREBE demonstrates that an impartial approach to rules and precedents is highly effective for reasoning in complex, ambiguous domains. By treating precedents as operationalized versions of abstract rules, the system successfully bridges the generality gap, producing human-level legal justifications. Future work will focus on expanding the knowledge base to handle a wider variety of legal distinctions and exploring more sophisticated natural language generation.
Summary updated