Abstractive document summarization via multi-template decoding

摘要

Most previous abstractive summarization models generate the summary in a left-to-right manner without making the most use of target-side global information. Recently, many researchers seek to alleviate this issue by retrieving target-side templates from large-scale training corpus, yet have limitations in template quality. To overcome the problem of template selection bias, one promising direction is to get better target-side global information from multiple high-quality templates. Hence, this paper extends the encoder-decoder framework by introducing a multi-template decoding mechanism, which can utilize multiple templates retrieved from the training corpus based on the semantic distance. In addition, we introduce a multi-granular attention mechanism by simultaneously taking into account the importance of words in templates and the importance of different templates. Extensive experiment results on CNN/Daily mail and English Gigaword show that our proposed model significantly outperforms several state-of-the-art abstractive and extractive baseline models.