Distributional models of verb meaning: syntactic versus lexical contexts

Over the last decade or so, distributional methods have become the mainstay of semantic modelling in Computational Linguistics. As such, they have also been applied the automatic modelling of verb meaning. However, more than with other lexical categories, the research into verb semantics has taken its inspiration from the idea that a verb's meaning is strongly linked to its syntactic behaviour and more specifically, to its selectional preferences. Depending on how they use these selectional preferences, distributional models of verb meaning come in two flavours. The first approach has its historical origins in the linguistic research tradition into verb valency and frame sem-natics and is in principle purely syntactical in nature. A verb's semantic category is said to be inferrable from its distribution over subcategorization (subcat) frames, i.e. the possible combinations of syntactic verb arguments like subject, direct object, indirect object etc. Additionally, this purely syntactic information can be extended with some high-level semantic information like the animacy of the verb arguments (see Schulte im Walde (2006) for an overview). Whereas this first, syntax-oriented approach is specifically geared towards verbs, the second approach is more generally applicable to all lexical categories and is a direct implementation of the ideas of Harris (1954). These so-called word space models use other words as context features with a specific implementation using only those context words that co-occur in a given dependency relation to the target word (see Padó and Lapata (2007) for an overview). In the first approach, one context feature is a possible combination of syntactic arguments that a verb can govern. In the second approach, one specific context feature corresponds to one lexeme plus its syntactic relation to the target verb. Whereas the first approach is mostly used to automatically induce Levin-style verb classes, the second approach is typically applied to retrieve semantic equivalents for specific verbs (but see Li and Brew (2008) for a comparison of the two methods on the task of inducing Levin-style classes). In this presentation we will try to have a closer look at the kind of semantic information that is captured by these two distinct types of distri-butional methods for verb meaning. For a sample of 1000 frequent Dutch