Similar-Case Retrieval for Government Hotlines via an Appeal-Driven Event Graph and Dual-Route Hybrid Retrieval

Abstract

To address the colloquial expressions, redundant descriptions, and dispersed structural information commonly found in historical work orders from 12345 government hotlines, traditional keyword matching and purely semantic retrieval methods often fail to identify cases that are both semantically relevant and structurally comparable, limiting the effectiveness of similar-case retrieval in complex hotline scenarios. This paper proposes a similar-case retrieval method for government hotlines based on an appeal-driven event graph and dual-route hybrid retrieval. Specifically, key elements, including location, time, core issue, and core appeal, are extracted from historical work orders to construct an appeal-driven complaint event knowledge graph. On this basis, a dual-route hybrid retrieval mechanism that integrates dense vector recall and graph-based logical recall is designed, and a structure-aware fusion strategy with cross-encoder reranking is further introduced to improve candidate ranking quality. Experiments on a real government hotline complaint dataset and the public LeCaRD legal case retrieval dataset show that the proposed method outperforms multiple baseline methods. On the complaint dataset, it achieves a P@5 of 0.613 and a Mean Average Precision(MAP) of 0.633, while on LeCaRD it reaches a P@5 of 0.435 and a MAP of 0.487. The results demonstrate that the proposed method can effectively improve the accuracy and structural comparability of similar-case retrieval in complex government hotline scenarios.