Keyword-fusion Pointer-Generator Network for Policy Text Title Summarization

In this paper, we study the electricity price policy text title summarization problem. Comparing with conventional summarization tasks, title summarization of policy text has an extra characteristic. Policy texts always contain many professional keywords. In order to retain the main information in title summarization as much as possible, we propose keyword-fusion pointer-generator network with additional consideration of keywords of policy text. We incorporate keywords information from the original policy texts into our model by a new attention mechanism called keyword-fusion attention mechanism so that keywords can be generated in the title. What's more, our keyword-fusion pointer-generator network contains a more useful coverage vector using exponentially weighted averages method in order to solve the problem of repetition. Experimental results show that our model outperforms the other baselines.