PUB - Publikationen an der Universität Bielefeld

Extrapositions are discontinuous constructions of spontaneous speech where constituents are extraposed into the Nachfeld. Parallelism constraints between the source sentence and the extraposed constituents are claimed to govern all possible interpretations induced by an extraposed constituent. The aim of this article is to integrate visual context information into the parsing model to avoid possible overgeneralizations which might arise from contradictory information given by the source sentence and the extraposed constituent. Therefore, a careful approach to uncertainty is needed because the system is concerned with erroneous, vague, and incomplete data from vision and speech recognition. Bayesian Networks provide an adequate environment to model these uncertainties. The conditional probabilities of the Network are estimated from psycholinguistic experiments.