A Comparison of Selectional Preference Models for Automatic Verb Classification

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 511–522, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics A Comparison of Selectional Preference Models for Automatic Verb Classification Will Roberts and Markus Egg Institut für Anglistik und Amerikanistik, Humboldt University 10099 Berlin, Germany {will.roberts,markus.egg}@anglistik.hu-berlin.de Abstract We present a comparison of different selec- tional preference models and evaluate them on an automatic verb classification task in German. We find that all the models we compare are effective for verb clustering; the best-performing model uses syntactic information to induce nouns classes from unlabelled data in an unsupervised man- ner. A very simple model based on lexical preferences is also found to perform well. 1 Introduction Selectional preferences (Katz and Fodor, 1963; Wilks, 1975; Resnik, 1993) are the tendency for a word to semantically select or constrain which other words may appear in a direct syntactic re- lation with it. Selectional preferences (SPs) have been a perennial knowledge source for NLP tasks such as word sense disambiguation (Resnik, 1997; Stevenson and Wilks, 2001; McCarthy and Car- roll, 2003) and semantic role labelling (Erk, 2007); and recognising selectional violations is thought to play a role in identifying and interpreting meta- phor (Wilks, 1978; Shutova et al., 2013). We focus on the SPs of verbs, since determining which argu- ments are typical of a given verb sheds light on the semantics of that verb. In this study, we present the first empirical com- parison of different SP models from the perspective of automatic verb classification (Schulte im Walde, 2009; Sun, 2012), the task of grouping verbs to- gether based on shared syntactic and semantic prop- erties. We cluster German verbs using features captur- ing their valency or subcategorisation, following prior work (Schulte im Walde, 2000; Esteve Ferrer, 2004; Schulte im Walde, 2006; Sun et al., 2008; Korhonen et al., 2008; Li and Brew, 2008), and investigate the effect of adding information about verb argument preferences. SPs are represented by features capturing lexical information about the heads of arguments to the verbs; we restrict our focus here to nouns. We operationalise a selectional preference model as a function which maps such an argument head to a concept label. We submit that the primary characteristic of such a model is its granularity. In our baseline condition, all nouns are mapped to the same label; this effectively captures no information about a verb’s SPs (i.e., we cluster verbs using sub- categorisation information only). On the other ex- treme, each noun is its own concept label; we term this condition lexical preferences (LP). Between the baseline and LP lie a spectrum of models, in which multiple concepts are distinguished, and each concept label can represent multiple nouns. Our main hypothesis is that verb clustering will work best using a model of such intermediate gran- ularity. This follows the intuition that verbs would seem to select for classes of nouns; for instance, we suppose that essen ‘eat’ would tend to prefer as a …

Content truncated to 500 words