Coverage-Based Variational Generative Decoder for Abstractive Text Summarization

Neural abstractive summarization models, based on the attentional sequence-to-sequence model, can generate simple summarization from source text. However, these models inevitably face with repetition problem, and ignore the rich structure information contained in target summarization. To alleviate these problems, we propose a generation model that combines the coverage mechanism and neural variational inference to make full use of all information in sentences. First, we employ Variation Auto-Encoders(VAEs) for generating latent variables, which obtain the latent structure information in target summarization. Second, we introduced coverage mechanism in final decoding process, which is expected greatly improve the weights of those important parts in the encoder and effectively reduce repetition problem. Abstractive summaries are generate by combining the generated latent structure information and the coverage information. Experimental results on LCSTS dataset show that our model outperforms baseline methods.