Recognising Entailment within Discourse

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 770–778, Beijing, August 2010 Recognising Entailment within Discourse Shachar Mirkin§, Jonathan Berant†, Ido Dagan§, Eyal Shnarch§ § Computer Science Department, Bar-Ilan University † The Blavatnik School of Computer Science, Tel-Aviv University Abstract Texts are commonly interpreted based on the entire discourse in which they are sit- uated. Discourse processing has been shown useful for inference-based applica- tion; yet, most systems for textual entail- ment – a generic paradigm for applied in- ference – have only addressed discourse considerations via off-the-shelf corefer- ence resolvers. In this paper we explore various discourse aspects in entailment in- ference, suggest initial solutions for them and investigate their impact on entailment performance. Our experiments suggest that discourse provides useful informa- tion, which significantly improves entail- ment inference, and should be better ad- dressed by future entailment systems. 1 Introduction This paper investigates the problem of recognising textual entailment within discourse. Textual En- tailment (TE) is a generic framework for applied semantic inference (Dagan et al., 2009). Under TE, the relationship between a text (T) and a tex- tual assertion (hypothesis, H) is defined such that T entails H if humans reading T would infer that H is most likely true (Dagan et al., 2006). TE has been successfully applied to a variety of natural language processing applications, includ- ing information extraction (Romano et al., 2006) and question answering (Harabagiu and Hickl, 2006). Yet, most entailment systems have thus far paid little attention to discourse aspects of in- ference. In part, this is the result of the unavail- ability of adept tools for handling the kind of dis- course processing required for inference. In addi- tion in the main TE benchmarks, the Recognising Textual Entailment (RTE) challenges, discourse played little role. This state of affairs has started to change with the recent introduction of the RTE Pilot “Search” task (Bentivogli et al., 2009b), in which assessed texts are situated within complete documents. In this setting, texts need to be inter- preted based on their entire discourse (Bentivogli et al., 2009a), hence attending to discourse issues becomes essential. Consider the following exam- ple from the task’s dataset: (T) The seven men on board were said to have as little as 24 hours of air. For the interpretation of T, e.g. the identity and whereabouts of the seven men, one must consider T’s discourse. The preceding sentence T’, for in- stance, provides useful information to that aim: (T’) The Russian navy worked desperately to save a small military submarine. This example demonstrates a common situation in texts, and is also applicable to the RTE Search task’s setting. Still, little was done by the task’s participants to consider discourse, and sentences were mostly processed independently. Analyzing the Search task’s development set, we identified several key discourse aspects that af- fect entailment in a discourse-dependent setting. First, we observed that the coverage of available coreference resolution tools is considerably lim- ited. To partly address this problem, we extend …

Content truncated to 500 words