Abstract
Coherence plays a critical role in producing a high-quality summary from adocument. In recent years, neural extractive summarization is becomingincreasingly attractive. However, most of them ignore the coherence ofsummaries when extracting sentences. As an effort towards extracting coherentsummaries, we propose a neural coherence model to capture the cross-sentencesemantic and syntactic coherence patterns. The proposed neural coherence modelobviates the need for feature engineering and can be trained in an end-to-endfashion using unlabeled data. Empirical results show that the proposed neuralcoherence model can efficiently capture the cross-sentence coherence patterns.Using the combined output of the neural coherence model and ROUGE package asthe reward, we design a reinforcement learning method to train a proposedneural extractive summarizer which is named Reinforced Neural ExtractiveSummarization (RNES) model. The RNES model learns to optimize coherence andinformative importance of the summary simultaneously. Experimental results showthat the proposed RNES outperforms existing baselines and achievesstate-of-the-art performance in term of ROUGE on CNN/Daily Mail dataset. Thequalitative evaluation indicates that summaries produced by RNES are morecoherent and readable.