Bilexical dependencies have been commonly used to help identify the most likely parses of a sentence. The probability of a word occurring as the dependent of a given head within a particular structure provides a measure of semantic plausibility that complements the purely syntactic part of the parsing model. Here, we attempt to use the distributional information within these bilexical dependencies to construct representations that decompose into semantic and syntactic components. In particular, we compare two different approaches to composing vectors to explore how syntactic and semantic representations should interact within such a model. Our results suggest a tensor product approach has advantages, which we believe could be exploited in making more effective use of the information captured in these bilexical dependencies.
展开▼