In this paper we present a corpus-driven approach to the detection of syntax-lexis interactions. Our approach is based on the output of a syntactic parser. We have parsed the British National Corpus and constructed a database of lexical dependencies. Such a large-scale approach allows for a detailed investigation of patterns and constructions associated with individual lexical items found in argument positions.
We then address the methodological problems of such an approach: precision errors (unwanted instances) and recall errors (missed instances) and offer a detailed evaluation. We investigate the interaction between syntax and lexis in verb-subject and verb-object structures as well as the active-passive alternation. We show that our approach provides relatively clean data and allows for a corpus-driven investigation of rare collocations.