Evaluating and enhancing cross-domain rank predictability of textual entailment datasets

Abstract

Textual Entailment (TE) is the task of recognizing entailment, paraphrase, and contradiction relations between a given text pair. The goal of textual entailment research is to develop a core inference component that can be applied to various domains, such as IR or NLP. Since the domain that a TE system applies to may be different from its source domain, it is crucial to develop proper datasets for measuring the cross-domain ability of a TE system. We propose using Kendall's tau to measure a dataset's cross-domain rank predictability. Our analysis shows that incorporating “artificial pairs” into a dataset helps enhance its rank predictability. We also find that the completeness of guidelines has no obvious effect on the rank predictability of a dataset. To validate these findings, more investigation is needed; however these findings suggest some new directions for the creation of TE datasets in the future.