Using variant directional dis(similarity) measures for the task of textual entailment

Abstract

Textual entailment (TE) is a task used to determine degree of semantic inference between a pair of text fragments in many natural language processing applications. In literature, a single document summarization framework has exploited TE to establish degree of connectedness between pair of sentences in a text summarization method. Despite noteworthy performance of the method, the extensive resource requirements and slow speed of the TE tool render it impractical to generate summaries in real time scenarios. This has stimulated the authors to propose the use of available directional dis(similarity) (distance and similarity) measures in place of TE system. The present paper aims to find a suitable directional measure which can successfully replace the TE system and decrease the overall runtime of the summarization method. Therefore, state-of-the-art directional dis(similarity) measures are implemented in the same summarization framework to present a comparative analysis of performance of all the measures. The experiments are conducted on DUC 2002 dataset and the results are evaluated using ROUGE tool to find the most suitable directional measure of textual entailment.