Comparison of collocation extraction measures for document indexing

Automatic extraction of collocations from a corpus is a well-known problem in the field of natural language processing. It is typically carried out by employing some kind of a statistical measure that indicates whether or not two words occur together more often than by chance. As there is an abundance of these measures proposed by various authors, we have compared some of them on a task of extracting collocations from a corpus of Croatian legal documents for the purpose of document indexing. We propose and evaluate extensions of these measures for collocations consisting of three words