Can we determine the semantics of collocations without using semantics?

The extraction of collocations from corpora has been actively worked on since the late eighties. However, so far, an important task of collocation processing, namely the semantic interpretation of the collocate, did not receive much attention, although the semantics of a given word when used as collocate very often varies from the semantics of this word when used in a free co-occurrence. In this paper, we tackle this problem. Our aim is the automatic semantic disambiguation of collocates, or, more precisely, the classification of collocations with respect to the typology of lexical functions (LFs) introduced in the Explanatory Combinatorial Lexicology. The two main questions underlying our research that seeks a scalable solution which does not depend on any external semantic resources are: (i) how well can we semantically classify collocates without the use of explicit semantic features; (ii) to what extent can we dispense with explicit lexical information when classifying collocates. To answer these two questions, we carried out machine learning experiments in which we used different training feature sets and LF typologies of different abstraction. So far, we worked on Spanish verb-noun and noun-adjective collocations from the lexicographic field of emotion nouns. However, our approach is, strictly speaking, language-independent.