A Hierarchical Attention Seq2seq Model with CopyNet for Text Summarization

We proposed a Hierarchical Attention Seq2seq (HAS) Model to abstractive text summarization, and show that they achieve state-of-the-art performance on two different corpora. In our opinion, the location of the passage expresses special meaning due to people's habits. Just as people usually put the main content in the middle, and expressing the summarization at the beginning, expresses conclusion at the ending. So our model splits the long text into different paragraphs, then train different encoder networks for these paragraphs according different locations. Our work shows that many of our proposed models contribute to further improvement in performance.