Similarity between the Association Measures: a Case Study of Noun Phrases

Collocation extraction has gained much attention in natural language processing, its results are important in various areas of applied linguistics. The research focuses on a comparison between over a dozen of association measures based on a subset of the Russian Web corpus. The paper studies the automatically extracted Adj-Noun collocations. The aim of the experiments is two-fold. First, to examine the difference between statistical measures and second to find the most effective one for the Russian data. The former assumes the calculation of the Spearman's rank correlation coefficient and the latter implies the evaluation of the extracted lists against a Russian dictionary, i.e. identifying automatically extracted and manually collected collocations. The results are not such straightforward, one can distinguish between groups of measures that demonstrate a relative interchangeability. Also the produced bigrams can be considered as collocations by experts and thus may enrich dictionaries.