Animacy Detection with Voting Models

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 55–60, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Animacy Detection with Voting Models Joshua L. Moore∗ Dept. Computer Science Cornell University Ithaca, NY 14853 jlmo@cs.cornell.edu Christopher J.C. Burges Erin Renshaw Wen-tau Yih Microsoft Research One Microsoft Way Redmond, WA 98052 {cburges, erinren, scottyih} @microsoft.com Abstract Animacy detection is a problem whose solu- tion has been shown to be beneficial for a number of syntactic and semantic tasks. We present a state-of-the-art system for this task which uses a number of simple classifiers with heterogeneous data sources in a voting scheme. We show how this framework can give us direct insight into the behavior of the system, allowing us to more easily diagnose sources of error. 1 Introduction Animacy detection has proven useful for a va- riety of syntactic and semantic tasks, such as anaphora and coreference resolution (Orˇasan and Evans, 2007; Lee et al., 2013), verb argument dis- ambiguation (Dell’Orletta et al., 2005) and depen- dency parsing (Øvrelid and Nivre, 2007). Existing approaches for animacy detection typically rely on two types of information: linguistic databases, and syntactic cues observed from the corpus. They usu- ally combine two types of approaches: rule based systems, and machine learning techniques. In this paper we explore a slightly different angle: we wish to design an animacy detector whose decisions are interpretable and correctable, so that downstream semantic modeling systems can revisit those deci- sions as needed. Thus here, we avoid defining a large number of features and then using a ma- chine learning method such as boosted trees, since such methods, although powerful, result in hard-to- interpret systems. Instead, we explore combining interpretable voting models using machine learning ∗ Work performed while visiting Microsoft Research. only to reweight their votes. We show that such an approach can indeed result in a high perform- ing system, with animacy detection accuracies in the mid 90% range, which compares well with other re- ported rates. Ensemble methods are well known (see for example, Dietterich (2000)) but our focus here is on using them for interpretability while still main- taining accuracy. 2 Previous Work 2.1 Definitions of Animacy Previous work uses several different definitions of animacy. Orˇasan and Evans (2007) define animacy in the service of anaphora resolution: an NP is con- sidered animate “if its referent can also be referred to using one of the pronouns he, she, him, her, his, hers, himself, herself, or a combination of such pro- nouns (e.g. his/her )”. Although useful for the task at hand, this has counterintuitive consequences: for example, baby may be considered animate or inan- imate, and ant is considered inanimate (Ibid., Fig- ure 1). Others have argued that animacy should be captured by a hierarchy or by categories (Aissen, 2003; Silverstein, 1986). For instance, Zaenen et al. (2004) propose three levels of animacy (human, other animate and inanimate), which cover ten cat- egories of noun phrases, with categories like ORG …

Content truncated to 500 words