A Deep Learning Approach to overcome the issues in Abstractive Summarization

Abstract

Generating abstractive summaries has always been difficult since the new words in the summary may disregard the meaning of the text. Abstractive summarization generally faces two major issues: Lack of facts in the summary and repeated words in the summary. This study focused on the challenges of generating novel/nearly accurate summaries. Webis-TLDR-17 corpus is used for this study. The corpus comprises of unstructured social media posts from the social media site, Reddit (2006-2016). In this study, three different models were trained: the pointer generator network (with and without coverage) and the Seq2Seq model as baseline for comparison of the generated summaries. ROGUE Evaluation method was used to calculate the quality of the generated summary. Final generated summaries show the outperformance of the proposed models.
IndexTerms—Attention;Webis-TLDR-17;