Semantic and Logical Inference Model for Textual Entailment

We compare two approaches to the problem of Textual Entailment: SLIM, a compositional approach modeling the task based on identifying relations in the entailment pair, and BoLI, a lexical matching algorithm. SLIM’s framework incorporates a range of resources that solve local entailment problems. A search-based inference procedure unifies these resources, permitting them to interact flexibly. BoLI uses WordNet and other lexical similarity resources to detect correspondence between related words in the Hypothesis and the Text. In this paper we describe both systems in some detail and evaluate their performance on the 3rd PASCAL RTE Challenge. While the lexical method outperforms the relation-based approach, we argue that the relation-based model offers better long-term prospects for entailment recognition.