Supervised Learning of a Probabilistic Lexicon of Verb Semantic Classes

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1328–1337, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Supervised Learning of a Probabilistic Lexicon of Verb Semantic Classes Yusuke Miyao University of Tokyo Hongo 7-3-1, Bunkyo-ku, Tokyo, Japan yusuke@is.s.u-tokyo.ac.jp Jun’ichi Tsujii University of Tokyo University of Manchester National Center for Text Mining Hongo 7-3-1, Bunkyo-ku, Tokyo, Japan tsujii@is.s.u-tokyo.ac.jp Abstract The work presented in this paper explores a supervised method for learning a prob- abilistic model of a lexicon of VerbNet classes. We intend for the probabilis- tic model to provide a probability dis- tribution of verb-class associations, over known and unknown verbs, including pol- ysemous words. In our approach, train- ing instances are obtained from an ex- isting lexicon and/or from an annotated corpus, while the features, which repre- sent syntactic frames, semantic similarity, and selectional preferences, are extracted from unannotated corpora. Our model is evaluated in type-level verb classifica- tion tasks: we measure the prediction ac- curacy of VerbNet classes for unknown verbs, and also measure the dissimilarity between the learned and observed proba- bility distributions. We empirically com- pare several settings for model learning, while we vary the use of features, source corpora for feature extraction, and disam- biguated corpora. In the task of verb clas- sification into all VerbNet classes, our best model achieved a 10.69% error reduction in the classification accuracy, over the pre- viously proposed model. 1 Introduction Lexicons are invaluable resources for semantic processing. In many cases, lexicons are neces- sary to restrict a set of semantic classes to be as- signed to a word. In fact, a considerable number of works on semantic processing implicitly or explic- itly presupposes the availability of a lexicon, such as in word sense disambiguation (WSD) (Mc- Carthy et al., 2004), and in token-level verb class disambiguation (Lapata and Brew, 2004; Girju et al., 2005; Li and Brew, 2007; Abend et al., 2008). In other words, those methods are heavily de- pendent on the availability of a semantic lexicon. Therefore, recent research efforts have invested in developing semantic resources, such as WordNet (Fellbaum, 1998), FrameNet (Baker et al., 1998), and VerbNet (Kipper et al., 2000; Kipper-Schuler, 2005), which greatly advanced research in seman- tic processing. However, the construction of such resources is expensive, and it is unrealistic to pre- suppose the availability of full-coverage lexicons; this is the case because unknown words always ap- pear in real texts, and word-semantics associations may vary (Abend et al., 2008). This paper explores a method for the supervised learning of a probabilistic model for the VerbNet lexicon. We target the automatic classification of arbitrary verbs, including polysemous verbs, into all VerbNet classes; further, we target the esti- mation of a probabilistic model, which represents the saliences of verb-class associations for polyse- mous verbs. In our approach, an existing lexicon and/or an annotated corpus are used as the training data. Since VerbNet classes are designed to rep- resent the distinctions in the syntactic frames that verbs can take, …

Content truncated to 500 words