MECE-Driven Neuro-Symbolic Framework for Explainable Legal Inference

摘要

Existing learning-based approaches lack explicit legal reasoning structures, while retrieval-augmented generation (RAG) suffers from sound deductive reasoning over legal elements and conditions due to fragmentation. In this paper, we present a neuro-symbolic framework for explainable legal statute inference that addresses the core problem in legal AI: how to preserve normative structure and verifiable reasoning when applying large language models to statute application.To resolve this, we propose three legally grounded innovations. Methodologically, we formalize principles for legal knowledge engineering, including MECE-driven ontology construction that preserves statutory hierarchies and logical operators, and atomic decomposition that transforms case narratives into standardized legal elements at appropriate reasoning granularity. Architecturally, we design a hybrid neuro-symbolic pipeline that integrates vector-based recall, LLM-based semantic filtering, and graph-based logical reasoning to determine applicable statutes with complete and traceable inference paths. Empirically, we provide a comprehensive evaluation that demonstrates the foundational nature of explicit normative modeling for reliable legal AI.We apply this framework on Taiwan’s criminal property law (Articles 320–348), constructing a knowledge graph comprising 79 legal elements, 94 crime definitions, 36 statutes, and 646 explicit relations. Evaluation on 325 real-world cases achieves 76.62\% accuracy, outperforming a traditional RAG (62.77\%) by 13.85 percentage points. Error analysis reveals five recurrent failure modes, including hierarchical misclassification and temporal or causal reasoning errors, indicating that many failures stem from violations of legal structure rather than surface-level semantic mismatch. These findings suggest that explicit modeling of normative structure and reasoning granularity is a foundational requirement for reliable and explainable legal statute inference.