کنفرانس بین المللی محاسبات نرم

صفحه اصلی / ششمین کنفرانس بین المللی محاسبات نرم

Similar Judicial Rulings Recommendation System Based on Similarity Graph, Network Analysis, and Large Language Models: Application in Theft Domain in Iranian Law

نویسندگان :

Mohammad Farhadishad¹ Zainabolhoda Heshmati²

1- دانشگاه تهران 2- دانشگاه تهران

کلمات کلیدی :

Judicial Rulings Recommendation،Similarity Graph،Louvain Community Detection،PageRank،Retrieval-Augmented Generation،Large Language Models

چکیده :

This paper introduces a network-enhanced recommendation system for identifying similar judicial rulings in the theft domain of Iranian law, designed to meet the challenges of semantic diversity, implicit references, and conflicting precedents. The system integrates jina-embeddings-v3 for high-fidelity multilingual legal embeddings [1], Milvus for large-scale similarity search, and a graph-based reranking pipeline leveraging Louvain community detection, PageRank, betweenness, closeness, and eigenvector centralities [3]. On top of this retrieval backbone, Gemma-3-27B-IT [2] generates grounded, citation-rich responses in a Retrieval-Augmented Generation (RAG) architecture. Unlike embedding-only systems, our approach explicitly incorporates graph-theoretic signals for ranking and explainability. Quantitative experiments on theft-related rulings (covering advisory opinions, rulings, sessions, and laws) show that the proposed method significantly outperforms baselines: Precision@5 = 0.89, Recall@5 = 0.86, F1 = 0.87. Beyond classical IR metrics, we also evaluate faithfulness (0.92), groundedness (0.9), and hallucination rate (reduced to 0.07), confirming the advantages of coupling network analytics with LLM-based generation. Expert surveys further validate the system’s interpretability, highlighting the role of community detection in clustering thematically consistent theft cases and centrality in surfacing legally influential documents. The resulting search engine and graphical chatbot provide a scalable, explainable, and domain-adapted platform for Iranian judicial applications.