Solving Textual Entailment with the Theorem Prover for Natural Language

Abstract

We present a theorem prover for natural language and show how it processes various types of textual entailment problems. The prover itself is based on a tableau system for natural logic that employs logical forms similar to linguistic expressions. With respect to the problems drawn from textual entailment datasets, a wide-range of the judgments of the prover are discussed, including both correct and incorrect ones. The analysis shows that the false proofs, which are extremely rare, are mainly due to the wrong lexical senses or the noisy gold labels of the dataset. Knowledge sparsity is identified as the main reason for the failure in proof search.