A novel model on improving Chinese dialogue summarization with multi-perspective information enhancement

Dialogue summarization aims to distill lengthy dialogues into concise texts that encapsulate key information. Typically, dialogues are accompanied by topic variations and characterized by multi-role participation. However, existing methods have difficulty in accurately tracking the topic information and suffer from error propagation, and they do not fully explore the interaction information between different roles in multiple dimensions. In this paper, we propose a novel model on Improving Chinese Dialogue Summarization with Multi-perspective Information Enhancement. Specifically, the model tracks topic changes through an innovative topic-guided tracker, guiding the learning of topic information via topic loss implicitly. Then, the model adopts a unique adaptive attention fusion method that combines contextual attention distribution and role attention distribution to synthesize both global and local fine-grained interaction information of different roles, and obtain a hybrid attention distribution with multi-dimensional interaction information. Finally, we combine the feature of discourse context with the hybrid attention distribution to generate summary. Experimental results have shown that our model largely outperforms strong baselines on two public dialogue summarization datasets, CSDS and MC. Further analyses have demonstrated that our model can effectively incorporate multi-perspective information enhancement to significantly improve the quality of summary.¹