Semi-supervised Verb Class Discovery Using Noisy Features

Semi-supervised Verb Class Discovery Using Noisy Features Suzanne Stevenson and Eric Joanis Department of Computer Science University of Toronto � suzanne,joanis � @cs.toronto.edu Abstract We cluster verbs into lexical semantic classes, using a general set of noisy features that cap- ture syntactic and semantic properties of the verbs. The feature set was previously shown to work well in a supervised learning setting, us- ing known English verb classes. In moving to a scenario of verb class discovery, using cluster- ing, we face the problem of having a large num- ber of irrelevant features for a particular cluster- ing task. We investigate various approaches to feature selection, using both unsupervised and semi-supervised methods, comparingthe results to subsets of features manually chosen accord- ing to linguistic properties. We find that the un- supervised method we tried cannot be consis- tently applied to our data. However, the semi- supervised approach (using a seed set of sam- ple verbs) overall outperforms not only the full set of features, but the hand-selected features as well. 1 Introduction Computational linguists face a lexical acquisition bot- tleneck, as vast amounts of knowledge about individual words are required for language technologies. Learn- ing the argument structure properties of verbs—the se- mantic roles they assign and their mapping to syntac- tic positions—is both particularly important and difficult. A number of supervised learning approaches have ex- tracted such informationabout verbs from corpora, includ- ing their argument roles (Gildea and Jurafsky, 2002), se- lectional preferences (Resnik, 1996), and lexical semantic classification (i.e., groupingverbs according to their argu- ment structure properties) (Dorr and Jones, 1996; Lapata and Brew, 1999; Merlo and Stevenson, 2001; Joanis and Stevenson, 2003). Unsupervised or semi-supervised ap- proaches have been successful as well, but have tended to be more restrictive, in relying on human filtering of the results (Riloff and Schmelzenbach, 1998), on the hand- selection of features (Stevenson and Merlo, 1999), or on the use of an extensive grammar (Schulte im Walde and Brew, 2002). We focus here on extending the applicability of un- supervised methods, as in (Schulte im Walde and Brew, 2002; Stevenson and Merlo, 1999), to the lexical seman- tic classification of verbs. Such classes group together verbs that share both a common semantics (such as trans- fer of possession or change of state), and a set of syntactic frames for expressing the arguments of the verb (Levin, 1993; FrameNet, 2003). As such, they serve as a means for organizing complex knowledge about verbs in a com- putational lexicon (Kipper et al., 2000). However, cre- ating a verb classification is highly resource intensive, in terms of both required time and linguistic expertise. De- velopment of minimally supervised methods is of particu- lar importance if we are to automatically classify verbs for languages other than English, where substantial amounts of labelled data are not available for training classifiers. It is also necessary to consider the probable lack of sophisti- cated grammars or text processing tools for extracting ac- curate features. We have previously shown that …

Content truncated to 500 words