Inferences for enrichment of collocation databases by means of semantic relations

A text consists of words that are syntactically linked and semantically combinable—like “political party,” “pay attention,” or “stone cold.” Such semantically plausible combinations of two content words, which we hereafter refer to as collocations, are important knowledge in many areas of computational linguistics. We present the structure of a lexical resource that provides such knowledge—a collocation database (CBD). Since such databases cannot be complete under any reasonable compilation procedure, we consider heuristic-based inference mechanisms that predict new plausible collocations based on the ones present in the CDB, with the help of a WordNet-like thesaurus: If an available collocation combines the entries A and B, and B is ‘similar’ to C, then A and C are supposed to constitute a collocation of the same category. Also, we describe the semantically induced morphological categories suiting for such inference, as well as the heuristics for filtering out wrong hypotheses. We discuss the experience in inferences obtained with CrossLexica CDB.