FusionSum: Abstractive summarization with sentence fusion and cooperative reinforcement learning

摘要

When summarizing an article, humans are habituated to fuse multiple related sentences to make the summary more concise and coherent. But most of the previous work focuses on the grammaticality of the fusion process and neglects the mechanism behind which sentences should be fused together. And there also lacks an effective training method for bridging the modules in the model to approach a global optimization. In this paper we propose FusionSum, a novel framework that imitates the behaviors of humans in summarization by explicitly modeling the sentence grouping and fusion process. It consists of an entity-aware sentence grouping module to identify salient sentences and combine them into groups, after which a unified sentence fusion module rewrites each group into a summary sentence. Furthermore, we also study the collaboration problem between these two modules and propose cooperative reinforcement learning method, which plays a maximax two-player game to approach global optimization. Automatic evaluation shows that our model significantly outperforms the strong baselines on three prevalent corpora. Human evaluation further demonstrates the summaries generated by our model are more concise and coherent.