Neural abstractive summarization fusing by global generative topics

Various efforts have been dedicated to automatically generate coherent, condensed and informative summaries. Most concentrate on improving the capability of generating neural language models locally, but do not consider global information. In real cases, a summary is comprehensively influenced by the full content of the source text and is especially guided by its core sense. To seamlessly integrate global semantic representation into a summarization generation system, we propose to incorporate a neural generative topic matrix as an abstractive level of topic information. By mapping global semantics into a local generative language model, the abstractive summarization is capable of generating succinct and recapitulative words or phrases. Extensive experiments on DUC-2004 and Gigaword datasets convincingly validate the proposed model.