An Ontology-Guided Multi-Anchor Graph Retrieval Framework for Traffic Legal Liability Determination

Technical Summary: Ontology-Guided Multi-Anchor Graph Retrieval for Traffic Legal Liability Determination

Problem Statement
Traffic law liability determination is a complex legal process requiring the simultaneous identification of interdependent statutory provisions across multiple dimensions, such as accident scenes, violation types, responsible parties, and liability allocation rules. Existing Retrieval-Augmented Generation (RAG) approaches face a "multi-dimensional retrieval bottleneck." Standard dense retrieval compresses multi-faceted legal queries into a single vector, causing the dominant semantic dimension (e.g., the specific violation) to crowd out other legally relevant dimensions. Similarly, knowledge graph-based systems that traverse from a single starting entity often fail to retrieve provisions anchored in other ontological dimensions if the initial extraction misses a key entity. Consequently, answers generated by these single-axis architectures are often legally incomplete, as they fail to capture the full spectrum of co-applicable statutes required for a valid liability determination.

Methodology: OMAGR and TrafficOmni-RAG
To address this bottleneck, the authors propose OMAGR (Ontology-Guided Multi-Anchor Graph Retrieval), instantiated as the core retrieval engine of TrafficOmni-RAG, an end-to-end legal QA system. The methodology operates through the following stages:

1. Legal Ontology and Knowledge Graph Construction:
   The system is built upon a six-dimensional legal ontology ($O = \{D_1, ..., D_6\}$) derived from six Chinese traffic law sources (including the Road Traffic Safety Law and judicial interpretations). The dimensions are:

   • Retrieval Anchors: Accident Scene ($D_1$), Violation Type ($D_2$), Responsible Party ($D_3$), and Liability Category ($D_4$).
   • Outcome Dimensions: Legal Consequence ($D_5$) and Damage Type ($D_6$).
     A Neo4j knowledge graph ($G$) is constructed containing 527 provisions. Provisions are nodes connected by typed edges representing ontology associations, inter-provision cross-references, and statutory hierarchies.

2. Parallel Multi-Anchor Retrieval:
   Unlike single-path traversal, OMAGR decomposes a user query into structured legal anchors aligned with the ontology. It executes parallel graph retrievals simultaneously across the four anchor dimensions ($D_1$–$D_4$). This ensures that every legally relevant dimension receives an independent retrieval pathway, preventing the dominant semantic signal from suppressing other critical dimensions.

3. Supporting Mechanisms:

   • Query Expansion: Generates paraphrased formulations to improve lexical and topical recall within each anchor's pathway.
   • Two-Hop Relational Expansion: Traverses inter-provision cross-reference edges from the top violation results to recover co-applicable statutes (e.g., penalties linked to procedural articles) that are not directly reachable via primary ontology edges.
   • Dense Retrieval: Provides a semantic fallback for provisions lacking strong graph connectivity.
   • Fusion and Refinement: All retrieved sets are merged using Reciprocal Rank Fusion (RRF). A cross-encoder reranker selects the top 8 provisions, followed by Maximal Marginal Relevance (MMR) to remove near-duplicates and hierarchical context pruning to optimize token allocation.

4. Generation:
   The refined provision set serves as the context for a generator LLM (GLM-4-Flash). The system employs citation-grounded generation, instructing the model to embed verbatim statutory phrases as evidence and prohibiting citations outside the retrieved set to minimize hallucination.

Key Contributions

1. OMAGR Framework: The proposal of a parallel multi-anchor graph retrieval architecture designed specifically to resolve the multi-dimensional retrieval bottleneck in legal domains.
2. TrafficLaw-QA Benchmark: The introduction of a new, expert-validated benchmark dataset containing 200 questions and 527 legal provisions. The dataset is annotated under a double-blind protocol (Cohen's $\kappa = 0.82$) and specifically targets scenarios requiring multi-statute grounding across interdependent dimensions.
3. Empirical Validation: Demonstration through experiments and ablation studies that the proposed system substantially improves Context Precision and Faithfulness compared to baselines.

Experimental Results
Evaluated on the TrafficLaw-QA dataset using RAGAS metrics, TrafficOmni-RAG outperformed three baselines: Naive RAG (dense vector only), TrafficRAG (hybrid BM25-dense), and LlamaIndex KG (single-entity graph traversal).

• Context Precision: TrafficOmni-RAG achieved the highest score (0.915), significantly outperforming LlamaIndex KG (0.722). The authors attribute this to the multi-anchor design retrieving provisions from all relevant dimensions rather than clustering around a single entity.
• Faithfulness: The system achieved 0.787, surpassing Naive RAG (0.625) and LlamaIndex KG (0.751).
• Ablation Studies: Removing the multi-anchor component (reverting to single-anchor retrieval) caused the largest drop in performance (Context Precision fell from 0.915 to 0.862), confirming that the parallel retrieval structure is the primary driver of success.
• Trade-offs: While TrafficRAG achieved slightly higher Context Recall (0.896 vs. 0.885) due to BM25 keyword matching, TrafficOmni-RAG prioritized precision, which the authors argue is functionally superior for legal QA to reduce hallucination risks from irrelevant context.

Significance and Claims
The paper claims that parallel multi-anchor retrieval effectively resolves the multi-dimensional retrieval bottleneck inherent in traffic law liability determination. By ensuring that independent evidence is gathered for every legally relevant dimension before fusion, the framework offers a promising direction for legal research where single-axis architectures fail to capture the interdependent nature of statutory provisions. The authors position TrafficOmni-RAG not merely as an improvement in retrieval metrics, but as a necessary structural shift to handle the complexity of legal reasoning where a single accident activates provisions across multiple statutory categories simultaneously.

Limitations
The authors acknowledge several limitations: the quality of anchor extraction is dependent on the LLM parsing step (with an observed F1 of 64.3% in blind evaluation), the system uses a fixed two-hop expansion radius which may miss deeper dependency chains, and the current implementation is exclusive to Chinese traffic law, requiring further evaluation for cross-jurisdictional generalizability.