Hunting for Entailing Pairs in the Penn Discourse Treebank

摘要

Given the growing amount of resources developed in the NLP community, it is crucial to exploit as much as possible annotated data and tools across different research domains. Past works on discourse analysis have been conducted in parallel with research on semantic inference and, although the two fields of study are intertwined, there have been only few initiatives to put them into relation. Our work addresses the issue of interoperability by investigating the connection between implicit Restatement relations in the Penn Discourse Treebank (PDTB) and Textual Entailment. We compare the performance of two TE systems on the Restatement pairs and we argue that TE is a subclass of Restatement through a manual validation of the pairs. Furthermore, we observe that entailing pairs extracted from the PDTB add interesting and additional levels of complexity to TE, since inference relation relies less on lexical-syntactic variations, and more on reasoning.