Neural Query-Biased Abstractive Summarization Using Copying Mechanism

Tatsuya Ishigaki, Hen-Hsen Huang, Hiroya Takamura, Hsin-Hsi Chen, Manabu Okumura

Abstract: This paper deals with the query-biased summarization task. Conventional non-neural network-based approaches have achieved better performance by primarily including the words overlapping between the source and the query in the summary. However, recurrent neural network (RNN)-based approaches do not explicitly model this phenomenon. Therefore, we model an RNN-based query-biased summarizer to primarily include the overlapping words in the summary, using a copying mechanism. Experimental results, in terms of both automatic evaluation with ROUGE and manual evaluation, show that the strategy to include the overlapping words also works well for neural query-biased summarizers.