Unsupervised Semantic Role Labelling
Suzanne Stevenson
Department of Computer Science
University of Toronto
Mittwoch, 23.06.2001, 12 Uhr c.t., D6-135
Semantic interpretation is the overarching goal of computational
linguistics. One important step in achieving a meaningful
representation of a text is to determine its predicate-argument
relations -- that is, for each entity in the text, what role it plays
in the event in which it participates. For example, in the sentence,
"Jo returned to London", "Jo" is the Agent of the returning action,
and "London" is its Destination. In this work, we have developed an
unsupervised method for labelling the semantic arguments of verbs with
participant roles such as these. Our bootstrapping algorithm makes
initial unambiguous role assignments on the basis of knowledge from a
verb lexicon. The algorithm then iteratively updates the probability
model on which additional assignments are based. A novel aspect of
our approach is the use of classes of information as the basis for
backing off in our probability model. In preliminary results, we
achieve 30-50% reduction in the error rate over an informed baseline,
indicating the potential of our approach for a task that has
heretofore relied on large amounts of manually generated training
data.
This work is in collaboration with Robert Swier, University of Toronto.