Definition and Analysis of Intermediate Entailment Levels

摘要

In this paper we define two intermediatemodels of textual entailment, which correspondto lexical and lexical-syntactic levelsof representation. We manually annotateda sample from the RTE dataset accordingto each model, compared the outcomefor the two models, and exploredhow well they approximate the notion ofentailment. We show that the lexicalsyntacticmodel outperforms the lexicalmodel, mainly due to a much lower rateof false-positives, but both models fail toachieve high recall. Our analysis alsoshows that paraphrases stand out as adominant contributor to the entailmenttask. We suggest that our models and annotationmethods can serve as an evaluationscheme for entailment at these levels.