DIGSum: dynamic interaction graph representation for dialogue summarization

The rapid growth of online social platforms has generated a substantial amount of dialogue data, highlighting the increasing significance of dialogue summarization technologies in information retrieval. However, existing methods struggle to effectively address the inherent complexity of multi-party dialogues and their information flow patterns. To this end, we propose DIGSum, a novel framework for dialogue summarization that employs dynamic interaction graphs. DIGSum operates by extracting key information from both utterance and dialogue levels. It utilizes an innovative utterance-dialogue interaction graph module that integrates local details with global information flow. We evaluated DIGSum on two benchmark datasets, SAMSum and DialogSum. Experiments on benchmark datasets, SAMSum and DialogSum, demonstrate the framework's outstanding performance and robust generalization capabilities.