"Identifying Semantic Relations and Functional Properties 
 of Human Verb Associations"

This talk presents a study which is based on human associations evoked
by German verbs in a web experiment. The aim of the study is to
identify the various types of semantic associations provided by
humans, and to distinguish and quantify their relations to the target
verbs:

1. In a preparatory step, we distinguish the responses with respect to
the major part-of-speech tags: nouns, verbs, adjectives, adverbs.

2. For each verb associate, we look up the semantic relation between
the target and response verbs using the lexical taxonomy GermaNet
(Kunze, 2004).

3. For each noun associate, we investigate the kinds of linguistic
functions that are realised by the noun with respect to the target
verb. The analysis is based on an empirical grammar (Schulte im Walde,
2003).

The investigations serve two purposes: (a) The notion of semantic verb
relations is crucial for many NLP applications such as thesaurus
extraction, word sense discrimination, summarisation, and question
answering. We use GermaNet to identify and quantify the classical
verb-verb relations (e.g. synonymy, hypernymy) as indexed by the
semantic associations provided by native speakers; verb-verb pairs not
covered by GermaNet can help detect missing links in the taxonomy. In
addition, we demonstrate that the associations provide a useful basis
for defining non-classical relations (e.g. temporal order, causality)
on a theoretical basis, and for adding to NLP resources. (b) In
data-intensive lexical semantics, words are commonly modelled by
distributional vectors, and relatedness of words is measured by vector
similarity. We characterise the types of nouns for a distributional
verb description as based on the noun associates provided in our web
experiment, assuming that the properties of the elicited noun concepts
from the semantic associations are relevant conceptual roles of verbs.

This is joint work with Alissa Melinger.