A multi-dimensional comparison of the effectiveness and efficiency of association measures in collocation extraction

Because of the ubiquity and importance of collocations in language use/learning, how to effectively and efficiently identify collocations has been a topic of interest. Although some studies have evaluated many of the existing association measures (AMs) used in the automatic identification of collocations, the results so far have been inconsistent and unclear due to various limitations of the existing studies. Hence, this study makes a multi-dimensional evaluation of the effectiveness and efficiency of seven major AMs in the identification of three types of collocations across five genres and seven corpora of different sizes. The results indicate that while a few AMs, such as Log Likelihood Ratio and Cubic Mutual Information (MI3), are consistently more effective and efficient than the other five AMs being examined, no one AM alone may be adequate in the identification of different types of collocations across different genres and corpus sizes. Research implications are also discussed.