Outlier Detection in Automatic Collocation Extraction

In this paper we have analysed different association measures between words, generally used for the automatic extraction of collocations in textual corpus. Specifically, they have been considered: relative frequency, mutual information, z-score, t-score and Dunning's test. The volume of handled corpus (300000000 words) requires reviewing of the usual approach to this matter, so a solution that is based on methods used to detect statistical outliers is proposed. It is evident from the results that a lot of free combinations extracted with collocations coming from the comparison of words with very different frequencies of use. For this reason, they are applied considering that each word generates a different sample, instead of generating rankings which come from corpus considered as a single sample. The experiment is also performed on a corpus with a much smaller amount of words and the results are reported so contrasted with those obtained with the full corpus. The conclusions and contributions arising give response automatic extraction of collocations from a textual corpus regardless its volume.