Collocation is of great importance in dictionary compilation and natural language processing.Collocation extraction is one of the principal applications of corpus linguistics.Automatic extraction of bi-grams as candidate collocations is studied on Penn Treebank using the criteria of log likelihood,chi square and mutual information as association measure.The experimental results show the feasibility of the statistical methods.On the other hand,collocations extracted show different characteristics because of the different distribution assumptions by the three criteria.