"Automatic Induction of Semantic Classes for German Verbs"

Verb entries constitute an essential part of the lexicon, since the
verb is central to the meaning and the structure of a sentence. For
that reason, lexical verb information represents the core in
supporting NLP-tasks such as lexicography, parsing, machine
translation, and information retrieval. Since manually built extensive
lexica are resource-consuming, automatic processes support the
induction of lexicon-relevant information, such as subcategorisation
frames for English (Brent 1993, Manning 1993, Briscoe/Carroll 1997,
Carroll/Rooth 1998) and German (Eckle 1999, Wauschkuhn 1999, Schulte
im Walde 2002), and conveniently observable morpho-syntactic and
semantic properties for English verbs (Dorr/Jones 1996, Lapata 1999,
Stevenson/Merlo 1999, Schulte im Walde 2000, McCarthy 2001).

A verb classification provides an especially valuable principled basis
for filling gaps in available lexical knowledge, since it generalises
over both syntactic and semantic properties of verbs. The idea of a
classification is based on a long-standing linguistic hypothesis which
asserts a tight connection between the syntactic behaviour of a verb
and its meaning components: to a certain extent, the lexical meaning
of a verb determines its behaviour, particularly with respect to the
choice of its arguments. The theoretical foundation has been
established in extensive work on semantic verb classes such as Levin
(1993) for English and Vazquez et al. (2000) for Spanish.

I will describe the application of standard clustering techniques to
the task of automatically inducing semantic classes for German verbs:
In a first step, a robust statistical parser was used to obtain
lexical verb information from corpus data, referring to syntactic
frame definitions as well as selectional preferences. In a second
step, this lexical information was used to automatically infer
semantic labels for the German verbs. The resulting knowledge on the
verb classification represents lexical verb information which (i)
generalises over both syntactic and semantic properties of the verbs,
and (ii) can be used to predict the respective properties of verbs
which were so far unknown to the lexicon.

A concluding discussion will concentrate on the following aspects of
the induction process:
- Can we transfer the classification procedure to other lexical units
  than verbs?
- Which are possible applications of the lexical class information?
- Where are the limits of the automatic clustering process, especially
  concerning the semantic content of the lexical verb entries?