STAR: Syntax- and Topic-Aware Role Dialogue Summarization

Dialogue summarization aims to refine a dialogue into a concise and focused summary of the text. Existing approaches select the most salient dialogue content after topic division, but this often ignores the coherence of the dialogue context and the tolerance of topic division. In this paper, we propose the Syntax- and Topic-Aware Role (STAR) model, aimed at optimizing dialogue summary through syntactic analysis and topic assignment. The STAR improves dialogue summarization by accurately dividing dialogue topics and enhancing logical relationships between utterances, ensuring comprehensive and contextually coherent summarization. Specifically, STAR's topic-aware role performs semantics-based topic segmentation, which assigns a topic to each utterance and considers its significance and influence in the dialogue network. Meanwhile, its syntax-aware role utilizes syntactic analysis to determine the syntactic importance of each utterance in the dialogue context. Finally, we combine the two to guide the model in generating summaries that more accurately identify key thematic elements in the dialogue and remain sensitive to contextual logic. Validation on three public benchmark datasets, CSDS, MC and SAMSUM, shows that our proposed method outperforms the strong baseline. Further analysis shows that STAR effectively improves the accuracy and quality of summarization. Our code is publicly available at: https://anonymous.4open.science/r/STAR-B8B8.