Since association measures (AM) have been utilized to extract collocation candidates from corpora, it is crucial to identify a suitable AM for the extraction of pedagogically useful collocations as the initial step in creating a collocations list for L2 learners of English. Furthermore, dispersion is an important aspect of collocations selection in language learning research. Therefore, this study explores which AM is suitable for identifying pedagogically useful English collocations, considering both frequency and dispersion. Adjective-noun, noun-noun, verb-object, adverbadjective, and adverb-verb pairs were extracted from the syntactically parsed British National Corpus, and their AMs (MI, MI2, MI3, t-score, zscore, FD51 logDice, log-likelihood) were calculated. Dispersion was calculated using Gries’ DP. AMs were evaluated against a gold standard which was defined using Oxford Collocations Dictionary and New General Service List. To assess AMs’ performances, the precision and recall were computed by FD51 setting varying frequency and DP thresholds. The results show that t-score is the optimal AM for adjective/noun-noun, verb-object, and adverb-verb structures and logDice is optimal for adverb-adjective collocations. Applying frequency and DP thresholds at an appropriate level was also found to be quite effective and an adequate DP threshold is mostly 0.7.
Kohei FUKUDA (Mon,) studied this question.