Multi-document abstractive summarization based on predicate argument structure

Abstract

The proposed work is based on abstractive summarization which is the division of text summarization. It developed a summary of the multi-document using the semantic relationship between the input documents rather than what we get exactly from the input document. It is very necessary because of the difficulty of generating abstract manually and also a challenging task. In our system, summary is generated based on the predicate argument structure of the sentences. Semantic role labeling is utilized to obtain the predicate argument structure. Main steps involved in the proposed system: Predicate argument structure of the sentences is extracted to represent text semantically as the first step. Next, it group semantically similar predicate argument structure using hybrid approach of K-mean and agglomerative hierarchical clustering by utilizing semantic similarity measures. K-mean is selected due to its run time efficiency and agglomerative hierarchical clustering is selected due to its quality. We extract twelve features of the predicate argument and feature selection is made randomly in the optimization stage. Then top ranked predicate argument structures taken from the optimization phase. Sentences for summary is selected from top ranked predicate argument structure by utilizing language generation. The Proposed study is evaluated by Document Understanding Conference 2002 (DUC 2002). We observed that the proposed work saved the computation time and provides better result than existing systems.