Hybrid Non-Local Network for Abstractive Summarization
Zepeng Hao,Yupu Guo,Cheng Gong,Honghui Chen
DOI: https://doi.org/10.1088/1757-899x/692/1/012049
2019-01-01
IOP Conference Series Materials Science and Engineering
Abstract:In the field of abstractive automatic summarization, the Seq2Seq model constructed by the RNN network is widely used and has achieved perfect results. However, due to the Seq2Seq model and the internal structure of the RNN network, only short-term and local information can be referenced in the process of generating summary, resulting the generated summary contains less information, and may even happen that the generated summary does not match the corresponding document. There have been many previous attempts to improve the model by using the attention mechanism, but only to use long-term information within a certain range. In our paper, we present a abstractive summarization model that enhances the global feature, which called Hybrid Non-Local Network model for single document summarization task, consisting of a non-local block and a traditional Seq2Seq module. Our model uses the non-local network block to transform the Seq2Seq model to better capture global feature. For example, the relationship between words and words, sentences and sentences in a large span of the original document, making the generated summary of higher quality, more informative, more accurate and rich representation of the original document. The experiments in our paper are performed on the CNN/DailyMail corpus. The proposed model was improved by 1.37, 0.47, and 1.16 on R-1 score, R-2 score, and R-L score, separately, compared to the baseline model.