Using Discourse Commitments to Recognize Textual Entailment

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 337–344 Manchester, August 2008 Using Discourse Commitments to Recognize Textual Entailment Andrew Hickl Language Computer Corporation 1701 North Collins Boulevard Suite 2000 Richardson, Texas 75080 USA andy@languagecomputer.com Abstract In this paper, we introduce a new frame- work for recognizing textual entailment (RTE) which depends on extraction of the set of publicly-held beliefs – known as dis- course commitments – that can be ascribed to the author of a text (t) or a hypothesis (h). We show that once a set of commit- ments have been extracted from a t-h pair, the task of recognizing textual entailment is reduced to the identification of the com- mitments from a t which support the infer- ence of the h. Our system correctly identi- fied entailment relationships in more than 80% of t-h pairs taken from all three of the previous PASCAL RTE Challenges, with- out the need for additional sources of train- ing data. 1 Introduction Systems participating in the PASCAL Recogniz- ing Textual Entailment (RTE) Challenges (Dagan et al., 2005) have successfully employed a variety of “shallow” techniques in order to recognize in- stances of textual entailment, including methods based on: (1) sets of heuristics (Vanderwende et al., 2006), (2) measures of term overlap (Jijkoun and de Rijke, 2005) (or other measures of seman- tic “relatedness” (Glickman et al., 2005), (3) the alignment of graphs created from syntactic or se- mantic dependencies (Haghighi et al., 2005), or (4) statistical classifiers which leverage a wide range of features, including the output of paraphrase gen- eration (Hickl et al., 2006), inference rule genera- c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. tion (Szpektor et al., 2007), or model building sys- tems (Bos and Markert, 2006). While these relatively “shallow” approaches have shown much promise in RTE for entailment pairs where the text and hypothesis remain short, we expect that performance of these types of sys- tems will ultimately degrade as longer and more syntactically complex entailment pairs are consid- ered. For example, given a “short” t-h pair (as in (1)), we might expect that a feature-based compar- ison of the t and the h would be sufficient to iden- tify that the t textually entailed the h. (1) Short t-h Pair a. Text: Mack Sennett was involved in the produc- tion of “The Extra Girl”. b. Hypothesis: “The Extra Girl” was produced by Sennett. The additional information included in a longer t (like the one in (2)) can make for a much more challenging entailment computation. While the ev- idence supporting the h is included in the t, sys- tems must be able to establish that (1) Mack Sen- nett was involved in producing a Mabel Normand vehicle and (2) that “The Extra Girl” and the Ma- bel Normand vehicle refer to the same film. (2) Long t-h Pair a. Text: “The Extra Girl” (1923) is the story of a small-town girl, Sue …

Content truncated to 500 words