A Neural Topic Summarizer: using Topic Model to enrich Abstractive Summaries

Abstract

Nowadays, Neural Networks are widely used forabstractive and extractive summarization, since they are able tocreate human-like summaries. To the best of our knowledge,few existing neural models for summarization use a priori wordfeatures such as POS tags, which require time to be generated. Inthis paper, we present a model called NeTSumm that merges theSequence-to-Sequence model with topics, computed on-the-fly, toextract hidden thematic structures which influence the generationof the summaries. Despite our model did not reach state-ofthe-art results, it was able to better discover actual relationsbetween words.