Keyword-Aware Encoder for Abstractive Text Summarization

Tianxiang Hu, Jingxi Liang, Wei Ye, Shikun Zhang

Abstract: Text summarization aims to produce a brief statement covering main points. Human beings would intentionally look for key entities and key concepts when summarizing a text. Fewer efforts are needed to write a high-quality summary if keywords in the original text are provided. Inspired by this observation, we propose a keyword-aware encoder (KAE) for abstractive text summarization, which extracts and exploits keywords explicitly. It enriches word representations by incorporating keyword information and thus leverages keywords to distill salient information. We construct an attention-based neural summarizer equipped with KAE and evaluate our model extensively on benchmark datasets of various languages and text lengths. Experiment results show that our model generates competitive results comparing to state-of-the-art methods.