Dialogue Topic Segmentation Enhancement Method Based on Collaborative Graph Convolutional Networks and Large Language Models

Faculty of Computer and Mathematical Sciences, the Hong Kong Polytechnic University, 999077 Hong Kong, China

^* Corresponding author: 23108099d@connect.polyu.hk

Abstract

With the rapid development of dialogue systems and Retrieval- Augmented Generation (RAG) technologies, accurate dialogue topic segmentation has become a core link in improving context understanding and response quality. However, existing methods have some limitations: semantic similarity-based methods (e.g., Sentence-BERT) only capture local correlations and easily miss indirectly related segments, while Large Language Model (LLM)-based judgments, though precise in reasoning, are computationally expensive and slow, unsuitable for large-scale initial screening. Thus, this paper proposes a GCN-LLM collaborative topic segmentation framework for addressing these issues. It constructs a graph structure using LLM ‘ s “topic continuity” judgments as the “gold standard”, designs a lightweight Graph Convolutional Network (GCN) to aggregate neighbor features and generate “enhanced embeddings” to overcome locality limitations, and adopts an “assist- not-replace” strategy (GCN optimizes initial screening features, LLM determines final results). Experiments on the dialseg_711.json dataset show the framework achieves an average F1-score of 0.8393, significantly outperforming existing methods (the highest F1 of comparison methods is 0.6606), and its graph dynamic update mechanism adapts to real-time dialogues. This study provides a new precision-efficiency-balanced solution for long-dialogue topic segmentation, especially suitable for context-sensitive RAG systems and intelligent customer service scenarios.