A Hybrid Attention Based Autoencoder for Chinese Text Abstractive Summarization

Salient information selection is one of major difficulties for generating summaries with adequate salient information. To address this serious problem in Chinese text summarization task, we put forward a novel method for abstractive summarization. In this paper, we try to adjust the local and global information between the source text and summary, and construct a hybrid attention mechanism. The hybrid attention mechanism can dynamically leverage both of the local and global information. For conveniently quantifying their contributions, we use a gating scalar to integrate both sources of the local and global attention. Then, we combine the hybrid attention mechanism with a supervise autoencoder. We evaluate our model on a popular Chinese social media dataset, LCSTS. Experiments results on the LCSTS demonstrate that the proposed method is effective in the aspect of summary informativeness.