German Semantic Verb Classes - Manual and Automatic Acquisition

       Talk given by Sabine Schulte im Walde
       January 22, 2004

In this talk, I will investigate the potential and the limits of an
automatic acquisition of semantic classes for German verbs. Referring
to the tight connection between the lexical meaning of a verb and its
behaviour, cf. Levin (1993), I induce a verb classification on basis
of verb features describing verb behaviour and expect the resulting
behaviour-classification to agree with a semantic classification to a
certain extent. A statistical grammar model for German serves as
source for a German verb description at the syntax-semantic interface.
The automatic induction of the German verb classes is performed by the
k-Means algorithm, a standard unsupervised clustering technique. The
algorithm uses the syntactico-semantic descriptions of the verbs as
empirical verb properties and learns to induce a semantic
classification from this input data.

A main focus in the talk will be on the manual vs. the automatic
acquisition of German semantic verb classes: 
(a) I manually defined a small-scale verb classification which was
    used to evaluate the reliability and performance of the clustering
    experiments. The classification contains 43 German semantic verb
    classes with 168 partly ambiguous German verbs. 
(b) Within the SALSA project at CoLi in Saarbrücken, frame-semantic
    descriptions for German verbs and nouns are developed, as based on
    Fillmore's scenes-and-frames semantics, in order to semantically
    annotate the German TIGER corpus. 
(c) Depending on the setup of the clustering parameters, the automatic
    cluster analysis presents verb clusters which differ according to
    the underlying verb properties and according to the clustering
    methodology.
I will discuss the assumptions and properties underlying the diverse
kinds of verb classifications. To which extent do human beings agree
in the manual definitions? Which is the potential and where are the
limits of an automatic acquisition of semantic classes?