Workshop organized as a part of the Annual Meeting of the Cognitive Science Society (CogSci 2009)
In the last decade, corpus-based distributional models of semantic similarity and association have slipped into the mainstream of cognitive science and computational linguistics. On the basis of the contexts in which a word is used, they claim to capture certain aspects of word meaning and human semantic space organization. In computational linguistics, these models have been used to automatically retrieve synonyms (Lin, 1998) or to find the multiple senses of a word (Schütze, 1998), among other tasks. In cognitive science, they have been applied to the modelling of semantic priming (Burgess, Livesay, & Lund, 1998; Landauer & Dumais, 1997), semantic dyslexia (Buchanan, Burgess, & Lund, 1996), categorization and prototypicality (Louwerse, Hu, Cai, Ventura, & Jeuniaux, 2005), and many other phenomena. Yet, despite their claims to model human language behaviour, relatively little is known about the precise relationship between these distributional models and human semantic knowledge. While they offer a credible account of (thematic or general) similarity of unary predicates such as concrete nouns, the question remains if and how more complex knowledge can be modelled using distributional information. This workshop therefore wants to focus on new challenges to distributional approaches that lie beyond the traditional modelling of concrete concepts.
Three current challenges to distributional semantics take a central position in our workshop. These are the discovery of verb meaning, the modelling of different aspects of semantics and the combination of different types of data. We want to address each of these, with specific attention to the relationship between distributional models and human semantic cognition.