Improving Abstractive Summarization With Segment-Augmented And Position-Awareness

Abstract

Detecting the relationship between phrases to then derive the salient information is always an art of abstractive summarizing text. In this work, we present an improved version of the extractor-abstractor system called SEGMENT, in which the extractor identifies words and phrases included in the target summary and the abstractor leverages these features into generating a fluent summary. We provide a segment embedding layer to enrich information for the abstractor that increases the cohesion among phrases. In the extractor, we combined the filtering mechanism and the position awareness to improve the quality of information selectivity. Our method demonstrates potential improvements in CNN/DM dataset and exceeds state-of-the-art 5.1% in ROUGE-1 and 5% in ROUGE-2.