关键词: 推荐算法, 类案推荐, KeyBERT算法, 法条知识

Abstract: With the surge in judicial cases, traditional text matching methods struggle to meet the precise retrieval demands of massive legal documents due to inefficiency, weak generalization capability and insufficient interpretability. The professionalism and complexity of legal texts require models to not only capture semantic information but also incorporate structured domain knowledge. This study proposes a case recommendation algorithm integrating legal knowledge and case characteristics through constructing a legal article knowledge base and multi-dimensional similarity computation framework, aiming to enhance recommendation accuracy and interpretability. This assists judicial professionals in efficiently locating similar cases, improves transparency and consistency of legal reasoning, and provides a technically valuable pathway for judicial intelligence with both theoretical significance and practical value. Thus, effectively integrating case facts into legal provisions to build an interpretable case recommendation model becomes crucial for enhancing judicial efficiency and decision consistency.
   This paper presents a multi-dimensional feature fusion-based precedent recommendation model with the following core framework: (1) Data acquisition and preprocessing: crawling 215 criminal judgment documents on intentional injury from China Judgments Online and extracting factual descriptions as raw data. Text segmentation using THULAC with dual filtering through general and legal-domain-specific stop word lists (e.g., “public security bureau”, “review”) optimizes text representation. (2) Keyword extraction and legal knowledge base construction: KeyBERT algorithm extracts top-10 case keywords, filtered through BERT’s semantic understanding. Transforming criminal law provisions into element-based structures (e.g., decomposing “fraud crime” into elements like “defrauding public/private property” and “large amount”), stored in Elasticsearch as structured knowledge. (3) Semantic matching and similarity computation: XS-BERT (legally optimized pre-trained model) generates semantic vectors for keywords and legal elements. A weighted similarity function integrates three dimensions: charge overlap (Jaccard index), legal knowledge similarity (vector inner product) and sentence difference (normalized distance). (4) Recommendation and validation: using DCG as core metric, comparative experiments with TF-IDF and Word2Vec baseline models demonstrate superior retrieval accuracy and interpretability.
   The experimental results show significant advantages in DCG@5, DCG@10, and DCG@20 metrics over traditional methods. By integrating legal knowledge bases into deep learning, this model effectively addresses semantic gaps and logical inconsistencies in conventional legal text processing. The algorithm not only improves recommendation precision but also enhances credibility through structured legal provision matching, offering an efficient and reliable solution for judicial AI systems. Future work will extend to multi-offense joint recommendation, courtroom debate perspective integration and cross-jurisdictional adaptability optimization. The case recommendations demonstrate high consistency in charges, legal provisions and sentencing patterns with real cases, validating practical feasibility. This approach assists judicial professionals in rapidly locating similar precedents while enhancing decision interpretability, providing a technically referential framework balancing efficiency and precision. Subsequent research could incorporate external knowledge like trial arguments to further optimize multi-dimensional recommendation mechanisms.

Key words: recommendation algorithm, similar cases recommendation, KeyBERT algorithm, legal provisions knowledge