面向多版本法规问答的版本感知知识图谱增强检索方法

Version-Aware Knowledge-Graph-Enhanced Retrieval for Multi-Version Legal Question Answering

摘要: 法律法规在多次修订后，同一条款编号在不同版本中内容存在差异，且条款之间交叉引用频繁。现有基于BM25和稠密向量的检索增强生成(RAG)系统在两类问题上表现不佳：其一为时效性问题(即给定日期下应当适用条款的哪个版本)，其二为版本差异问题(即条款在不同版本之间发生了哪些变化)。本文分析上述不佳的根本原因，提出一套有针对性的解决方案。首先，构建跨版本法律知识图谱，包含条款引用(references)、版本修订(amends)和未变延续(inherits)三类边。其次，将问题划分为精确、版本差异、推理和时效四类，每类配置固定的跳数和边类型策略。当问题中出现具体条款编号时，跳过稠密种子检索，按article_no_int字段直接执行SQL查询以获取同条款的多版本内容，从而规避稠密检索因抽象关键词导致的语义偏移。本文在《中华人民共和国城乡规划法》四个版本(1989、2007、2015、2019，共256条)上构建67题基准开展评估。实验结果显示，本方法Recall@5达到0.963, MRR达到0.934，引用F1达到0.897，均高于BM25(0.761/0.692/0.718)和稠密检索(0.918/0.872/0.867)。在版本差异类别上，两类基线方法的Recall@5仅为0.167和0.500，本方法达到1.000，引用F1由0.133和0.500提升至0.933。在时效性类别上，本方法与稠密检索持平于Recall@5 = 1.000和引用F1 = 0.972。为验证泛化性，本文在《测绘法》(32题，修订密集)和《土地权属争议调查处理办法》(15题，修订稀疏)两部法律上补充实验。前者四项指标均达到1.000，后者退化为稠密检索水平但不劣化。双侧Wilcoxon配对检验显示，本方法相对BM25在六项指标上均p < 0.01(效应量r ∈ [0.659, 0.883])。本文实验环境完全可复现：嵌入采用本地Qwen3-Embedding-4B，生成采用Gemma-4-31B。

Abstract: Legal statutes are amended repeatedly. Across versions the same article number can carry different text and cross-references. BM25 and dense RAG both fail on two practical question types in this setting: temporal applicability (i.e., which version of an article should be applied on a given date) and version differences (i.e., how an article has changed across different versions). This paper analyzes the root causes of these failures and proposes a targeted solution. First, we build a cross-version legal knowledge graph containing references, amends, and inherits edges. Second, we classify questions into four types (precise, version difference, reasoning, and temporal applicability) and configure fixed hop and edge type policies for each. Finally, when an explicit article number appears in the question, we bypass dense seeding retrieval and directly execute a SQL query on the article_no_int field to obtain multi-version contents of the same article, thereby circumventing the semantic shift in dense retrieval caused by abstract keywords. On the Chinese Urban-Rural Planning Law (4 versions, 256 articles, 67-question benchmark), our method reaches Recall@5 = 0.963, MRR = 0.934, Hit@1 = 0.896, Citation F1 = 0.897, against 0.761/0.692/0.612/0.718 for BM25 and 0.918/0.872/0.821/0.867 for dense retrieval. On version-diff queries, the Recall@5 of the two baselines is only 0.167 and 0.500, while our method reaches 1.000, and the citation F1 is improved from 0.133 and 0.500 to 0.933. On the temporal applicability category, our method ties with dense retrieval at Recall@5 = 1.000 and citation F1 = 0.972. To verify generalization, we conduct supplementary experiments on two additional laws: the Surveying and Mapping Law (32 questions, densely amended) and the Measures for the Investigation and Handling of Land Ownership Disputes (15 questions, sparsely amended). The former achieves 1.000 on all four metrics, while the latter regresses to the dense retrieval level without degradation. Two-sided Wilcoxon paired tests show that our method outperforms BM25 on all six metrics with p < 0.01 (effect size r in [0.659, 0.883]). The experimental setup is fully reproducible, using local Qwen3-Embedding-4B (MLX backend) for retrieval and Gemma-4-31B (Google AI Studio free tier) for generation.