Indian Journal of Science and Technology

Abstract

Background/Objective: Multi-document summarization produces a concise summary from several online topically related documents. A major challenge in this domain is usually the information overlap in documents emanating from various sources. This paper introduces an optimized semantic technique for multi-document abstractive summarization. Methods/Statistical Analysis: Linguistic and semantic approaches are usually employed for abstractive summarization of multiple documents. Linguistic approaches lack semantic representation of source text while semantic approaches mostly rely on human experts to construct domain ontology and rules; which require immense time and effort. The technique in this paper utilizes the benefits of semantic role labeling, clustering and Particle Swarm Optimization (PSO) to rank predicate argument structures (semantic representation) in each cluster using optimized features. Findings: The summary quality is susceptible to the text features i.e., different features have varied importance towards summary generation. Therefore, optimal features weights obtained using PSO integrated in the semantic technique to rank semantic representation improved summarization results. The performance of the technique is evaluated against the benchmark summarization systems using pyramid evaluation measures (mean coverage score, precision and F-measure). A PairedSamples T-test is carried out to validate the summarization results. Applications/Improvements: Experiment of this research is performed with DUC-2002, a benchmark data set for text summarization. Experimental results confirm that the proposed technique yields better results than other comparison summarization models in terms of mean coverage score and average F-measure.
Keywords: Language Generation, Multi-Document Abstractive Summarization, Particle Swarm Optimization, Semantic Similarity Measure, Semantic Role Labeling (SRL)