Thematic Roles and Semantic Space Insights from Distributional Semantic Models

The goal of this work is to use Distributional Semantic Models (Sahlgren, 2006; Turney, 2010) to get insights into the nature of thematic roles. In particular, we investigate whether the semantic representation produced by Distributional Semantic Models (henceforth, DSMs) is sensitive to effects of typicality involving thematic roles, and we quantify their relative prominence in the semantic representation encoded in the distributional space. Corpus-based modeling of selectional preferences and thematic fit is a well established field of research (see Erk et al., 2007 and references therein). What is peculiar to our approach is its attempt to model thematic fit data without taking into account syntactic relations, on the basis of distributional relatedness in bag-of-words DSMs. In this abstract we will show that (a) DSMs that make no use of syntax show good performances in a task related to selectional preference and (b) that the distribution of DSMs’ performance across thematic relations shows patterns which are compatible with some general assumptions in theoretical linguistics.