| Resumen |
For extracting collocations from the Internet, it is necessary to numerically estimate the cohesion between potential collocates. Mutual Information cohesion measure (MI) based on numbers of collocate occurring closely together (N 12) and apart (N 1, N 2) is well known, but the Web page statistics deprives MI of its statistical validity. We propose a family of different measures that depend on N 1, N 2 and N 12 in a similar monotonic way and possess the scalability feature of MI. We apply the new criteria for a collection of N 1, N 2, and N 12 obtained from AltaVista for links between a few tens of English nouns and several hundreds of their modifiers taken from Oxford Collocations Dictionary. The ‘noun–its own adjective’ pairs are true collocations and their measure values form one distribution. The ‘noun–alien adjective’ pairs are false collocations and their measure values form another distribution. The discriminating threshold is searched for to minimize the sum of probabilities for errors of two possible types. The resolving power of a criterion is equal to the minimum of the sum. The best criterion delivering minimum minimorum is found. |
