ISP: Learning Inferential Selectional Preferences

Proceedings of NAACL HLT 2007, pages 564–571, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics ISP: Learning Inferential Selectional Preferences Patrick Pantel†, Rahul Bhagat†, Bonaventura Coppola‡, Timothy Chklovski† , Eduard Hovy† †Information Sciences Institute University of Southern California Marina del Rey, CA {pantel,rahul,timc,hovy}@isi.edu ‡ITC-Irst and University of Trento Via Sommarive, 18 – Povo 38050 Trento, Italy coppolab@itc.it Abstract Semantic inference is a key component for advanced natural language under- standing. However, existing collections of automatically acquired inference rules have shown disappointing results when used in applications such as textual en- tailment and question answering. This pa- per presents ISP, a collection of methods for automatically learning admissible ar- gument values to which an inference rule can be applied, which we call inferential selectional preferences, and methods for filtering out incorrect inferences. We evaluate ISP and present empirical evi- dence of its effectiveness. 1 Introduction Semantic inference is a key component for ad- vanced natural language understanding. Several important applications are already relying heavily on inference, including question answering (Moldovan et al. 2003; Harabagiu and Hickl 2006), information extraction (Romano et al. 2006), and textual entailment (Szpektor et al. 2004). In response, several researchers have created re- sources for enabling semantic inference. Among manual resources used for this task are WordNet (Fellbaum 1998) and Cyc (Lenat 1995). Although important and useful, these resources primarily contain prescriptive inference rules such as “X di- vorces Y ⇒ X married Y”. In practical NLP appli- cations, however, plausible inference rules such as “X married Y” ⇒ “X dated Y” are very useful. This, along with the difficulty and labor-intensiveness of generating exhaustive lists of rules, has led re- searchers to focus on automatic methods for build- ing inference resources such as inference rule collections (Lin and Pantel 2001; Szpektor et al. 2004) and paraphrase collections (Barzilay and McKeown 2001). Using these resources in applications has been hindered by the large amount of incorrect infer- ences they generate, either because of altogether incorrect rules or because of blind application of plausible rules without considering the context of the relations or the senses of the words. For exam- ple, consider the following sentence: Terry Nichols was charged by federal prosecutors for murder and conspiracy in the Oklahoma City bombing. and an inference rule such as: X is charged by Y ⇒ Y announced the arrest of X (1) Using this rule, we can infer that “federal prosecu- tors announced the arrest of Terry Nichols”. How- ever, given the sentence: Fraud was suspected when accounts were charged by CCM telemarketers without obtaining consumer authorization. the plausible inference rule (1) would incorrectly infer that “CCM telemarketers announced the ar- rest of accounts”. This example depicts a major obstacle to the ef- fective use of automatically learned inference rules. What is missing is knowledge about the ad- missible argument values for which an inference rule holds, which we call Inferential Selectional Preferences. For example, inference rule (1) should only be applied if X is a Person and Y …

Content truncated to 500 words