LEDIR: An Unsupervised Algorithm for Learning Directionality of Inference Rules

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 161–170, Prague, June 2007. c⃝2007 Association for Computational Linguistics LEDIR: An Unsupervised Algorithm for Learning Directionality of Inference Rules Rahul Bhagat, Patrick Pantel, Eduard Hovy Information Sciences Institute University of Southern California Marina del Rey, CA {rahul,pantel,hovy}@isi.edu Abstract Semantic inference is a core component of many natural language applications. In re- sponse, several researchers have developed algorithms for automatically learning infer- ence rules from textual corpora. However, these rules are often either imprecise or un- derspecified in directionality. In this paper we propose an algorithm called LEDIR that filters incorrect inference rules and identi- fies the directionality of correct ones. Based on an extension to Harris’s distribu- tional hypothesis, we use selectional pref- erences to gather evidence of inference di- rectionality and plausibility. Experiments show empirical evidence that our approach can classify inference rules significantly better than several baselines. 1 Introduction Paraphrases are textual expressions that convey the same meaning using different surface forms. Tex- tual entailment is a similar phenomenon, in which the presence of one expression licenses the validity of another. Paraphrases and inference rules are known to improve performance in various NLP applications like Question Answering (Harabagiu and Hickl 2006), summarization (Barzilay et al. 1999) and Information Retrieval (Anick and Tipir- neni 1999). Paraphrase and entailment involve inference rules that license a conclusion when a premise is given. Deciding whether a proposed inference rule is fully valid is difficult, however, and most NL systems instead focus on plausible inference. In this case, one statement has some likelihood of being identical in meaning to, or derivable from, the other. In the rest of this paper we discuss plau- sible inference only. Given the importance of inference, several re- searchers have developed inference rule collec- tions. While manually built resources like Word- Net (Fellbaum 1998) and Cyc (Lenat 1995) have been around for years, for coverage and domain adaptability reasons many recent approaches have focused on automatic acquisition of paraphrases (Barzilay and McKeown 2001) and inference rules (Lin and Pantel 2001; Szpektor et al. 2004). The downside of these approaches is that they often result in incorrect inference rules or in inference rules that are underspecified in directionality (i.e. asymmetric but are wrongly considered symmet- ric). For example, consider an inference rule from DIRT (Lin and Pantel 2001): X eats Y ⇔ X likes Y (1) All rules in DIRT are considered symmetric. Though here, one is most likely to infer that “X eats Y” ⇒ “X likes Y”, because if someone eats something, he most probably likes it1, but if he likes something he might not necessarily be able to eat it. So for example, given the sentence “I eat spicy food”, one is mostly likely to infer that “I like spicy food”. On the other hand, given the sentence “I like rollerblading”, one cannot infer that “I eat rollerblading”. In this paper, we propose an algorithm called LEDIR …

Content truncated to 500 words