Textual Entailment as a Directional Relation

This paper presents three methods for solving the problem of textual entailment, obtained from an equal number of text-to-text similarity metrics. The first method starts with the directional measure of text-to-text similarity presented in Corley and Mihalcea (2005), and integrates word sense disambiguation and several heuristics. The second method exploits the relations between the cosine directional measures of similarity as means to identify textual entailment. Finally, the third method relies on the directional variant of Levenshtein distance between two words. Each “word” in this method is a string consisting of all the words concatenated. In all these methods the decision about an entailment relation depends on the relation established between these measures of similarity. The methods are applied and evaluated on the whole set of text-hypothesis pairs included in the PASCAL RTE-1 development dataset (RTE-1, 2005). The corresponding accuracy and statistics are presented for each method.