Unsupervised Learning of P NP P Word Combinations

We evaluate the possibility to learn, in an unsupervised manner, a list of idiomatic word combinations of the type preposition + noun phrase + preposition (P NP P), namely, such groups with three or more simple forms that behave as a whole lexical unit and have semantic and syntactic properties not deducible from the corresponding properties of each simple form, e.g., by means of, in order to, in front of. We show that idiomatic P NP P combinations have some statistical properties distinct from those of usual idiomatic collocations. In particular, we found that most frequent P NP P trigrams tend to be idiomatic. Of other statistical measures, log-likelihood performs almost as good as frequency for detecting idiomatic expressions of this type, while chi-square and point-wise mutual information perform very poor. We experiment on Spanish material.