Recognizing textual entailment with statistical methods

Abstract

In this paper we propose a new cause-effect non-symmetric measure applied to the task of Recognizing Textual Entailment .First we searched over a big corpus for sentences which contains the discourse marker "because" and collected cause-effect pairs. The entailment recognition is based on measure the cause-effect relation between the text and the hypothesis using the relative frequencies of words from the cause-effect pairs. Our measure outperformed the baseline method, over the three test sets of the PASCAL Recognizing Textual Entailment Challenges (RTE). The measure shows to be good at discriminate over the "true" class. Therefore we develop a meta-classifier using a symmetric measure and a non-symmetric measure as base classifiers. So, our meta-classifier has a competitive performance. © 2010 Springer-Verlag Berlin Heidelberg.