Semantic argument frequency-based multi-document summarization

Abstract

Semantic role labeling (SRL) aims to identify the constituents of a sentence, together with their roles with respect to the sentence predicates. In this paper, we introduce and assess the idea of using SRL on generic multi-document summarization (MDS). We score sentences according to their inclusion of frequent semantic phrases and form the summary using the top-scored sentences. We compare this method with a term-based sentence scoring approach to investigate the effects of using semantic units instead of single words for sentence scoring. We also integrate our scoring metric as an auxiliary feature to a cutting edge summarizer with the intention of examining its effects on the performance. The experiments using datasets from the Document Understanding Conference (DUC) 2004 show that the SRL-based summarization outperforms the term-based approach as well as most of the DUC participants.