Estimating the Likelihood of Words Being Known with Corpus Analysis and K-Means Clustering Algorithm

This study investigates the extent to which corpora-based frequency indices predict the likelihood of words being known. Five hundred and twenty ESL students in China’s top-tier universities participated in this study. Their knowledge of target words to a meaning-recall level represented their likelihood of knowing these words. The target words were 932 content words retrieved from a text, of which the lexical coverage and readability level were appropriate for the participants. An online test was developed for the target words and administered to the participants. MATLAB was used to perform k-means clustering for participants’ answers and classify the likelihood of words being known into the most appropriate number of clusters. SPSS was used to perform the Kruskal-Wallis test Spearman correlation, and ordinal logistic regression. In addition to the classification of the likelihood of words being known, results showed significant differences and moderate correlations between corpora-based frequency indices and the classification. Moreover, base words’ frequency ranks on Nation’s (2012) BNC/COCA list were found to best correlate with and predict the likelihood of words being known. Future research is recommended to extend this study by classifying more words’ likelihood of being known to more learners at various levels of ESL proficiency.