Automatic Induction of Semantic Verb Classes that incorporate
Selectional Preferences

Semantic verb classifications, i.e., groupings of verbs according to
semantic properties, are of great interest to both theoretical and
computational linguistics. In theoretical linguistics, verb classes
are a useful means to organise verbs with respect to common
properties, such as meaning components (Koenig and Davis, 2001), or
shared argument structure (Levin, 1993). In computational linguistics,
semantic verb classifications represent a valuable source of
underspecification, by generalising over the verbs according to their
shared properties. Specifically regarding the pervasive problem of
data sparseness in the processing of natural language, such
classifications have been used in computational applications such as
word sense disambiguation (Dorr and Jones, 1996; Kohomban and Lee,
2005), machine translation (Prescher et al., 2000; Koehn and Hoang,
2007), document classification (Klavans and Kan, 1998), and
statistical lexical acquisition in general (Merlo and Stevenson, 2001;
Schulte im Walde, 2006).

This talk presents a novel approach to a semantic classification of
verbs, that incorporates selectional preferences as common verb
properties. Similarly to previous related work (Pereira et al., 1993;
Rooth et al., 1999), we rely on the Expectation-Maximisation (EM)
Algorithm as a soft-clustering technique, and model verb
classification by probabilistic class membership of verbs and their
semantic properties. In contrast to earlier work, we choose a more
complex set of semantic properties: rather than directly using
bilexical head dependencies between verbs and (direct object) nouns as
clustering dimensions, we abstract over the noun dimension by
selectional preferences. Consequently, a semantic class generalises
over verb senses (as one dimension), and selectional preferences (as a
second dimension). The talk will introduce the classification approach
in some detail, by providing (a) an overview of the parameters of the
clustering technique (the input, the probabilistic model, and the
implementation), and (b) potential interpretations and applications,
relying on examples for English and German verbs.