Improving Verb Clustering with Automatically Acquired Selectional Preferences

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 638–647, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Improving Verb Clustering with Automatically Acquired Selectional Preferences Lin Sun and Anna Korhonen University of Cambridge, Computer Laboratory 15 JJ Thomson Avenue, Cambridge CB3 0GD, UK ls418,alk23@cl.cam.ac.uk Abstract In previous research in automatic verb classification, syntactic features have proved the most useful features, although manual classifications rely heavily on se- mantic features. We show, in contrast with previous work, that considerable ad- ditional improvement can be obtained by using semantic features in automatic clas- sification: verb selectional preferences ac- quired from corpus data using a fully unsu- pervised method. We report these promis- ing results using a new framework for verb clustering which incorporates a re- cent subcategorization acquisition system, rich syntactic-semantic feature sets, and a variation of spectral clustering which performs particularly well in high dimen- sional feature space. 1 Introduction Verb classifications have attracted a great deal of interest in natural language processing (NLP). They have proved useful for various important NLP tasks and applications, including e.g. parsing, word sense disambiguation, semantic role label- ing, information extraction, question-answering, and machine translation (Swier and Stevenson, 2004; Dang, 2004; Shi and Mihalcea, 2005; Za- pirain et al., 2008). Verb classes are useful because they offer a powerful tool for generalization and abstraction which can be beneficial when faced e.g. with the problem of data sparsity. Particularly useful can be classes which capture generalizations over a range of (cross-)linguistic properties, such as the ones proposed by Levin (1993). Being defined in terms of similar meaning and (morpho-)syntactic behaviour of words, Levin style classes gener- ally incorporate a wider range of properties than e.g. classes defined solely on semantic grounds (Miller, 1995). In recent years, a variety of approaches have been proposed for automatic induction of verb classes from corpus data (Schulte im Walde, 2006; Joanis et al., 2008; Sun et al., 2008; Li and Brew, 2008; Korhonen et al., 2008; ´O S´eaghdha and Copestake, 2008; Vlachos et al., 2009). This work opens up the opportunity of learning and tuning classifications tailored to the application and do- main in question. Although manual classification may always yields higher accuracy, automatic verb classification is cost-effective and gathers statisti- cal information as a side-effect of the acquisition process which is difficult for humans to gather but can be highly useful for NLP applications. To date, both supervised and unsupervised ma- chine learning (ML) methods have been proposed for verb classification and used to classify a vari- ety of features extracted from raw, tagged and/or parsed corpus data. The best performing features on cross-domain verb classification have been syn- tactic in nature (e.g. syntactic slots, subcategoriza- tion frames (SCFs)). Disappointingly, semantic features have not yielded significant additional im- provement, although they play a key role in man- ual and theoretical work on verb classification and could thus be expected to offer a considerable con- tribution to classification performance. Since the accuracy of automatic verb …

Content truncated to 500 words