Leveraging speaker-aware structure and factual knowledge for faithful dialogue summarization

摘要

Currently, sequence/graph-to-sequence models for abstractive dialogue summarization are being studied extensively. However, previous methods strive to integrate complex events spanning multiple utterances, and the generated summaries are often filled with incorrect facts. In this study, we first utilize the speaker-aware structure to model the information interaction process in the dialogue, which shows an excellent ability to settle the cross-sentence dependency. Then, we incorporate the factual representations via a dual-copy decoder to obtain summaries conditioned on both the tokens from source sequences and the factual knowledge from our designed fact graph, which enhances the factual consistency for dialogue summarization. We also propose some fact-level factual consistency metrics. Adequate experimental results demonstrate that our model outperforms the state-of-the-art baselines by a significant margin on the SAMSum and DialSumm datasets. A comprehensive analysis also proves the effectiveness of our model. Furthermore, human judges confirm that the outputs of our model contain more informative and faithful information.