| Summary: | This paper aims at an automatic evaluation of second language (L2) learners’ proficiencies and tries to analyze English conversation data having 94 statistics and Global Scale scores of the Common European Framework of Reference (CEFR) given to each participant. The CEFR defines Range, Accuracy, Fluency, Interaction and Coherence as 5 subcategories, which constitute the CEFR Global Scale score. The statistics were classified into the CEFR’s 5 subcategories. We used the Principal Component Analysis (PCA), an unsupervised machine learning method, on each subcategory and obtained the participants’ principal component scores (PC scores) of the 5 subcategories for estimation parameters. We predicted the participants’ CEFR Global scores using the Multiple Regression Analysis (MRA). The proposed prediction method using the PC scores was compared with conventional methods with the 94 statistics. Based on the coefficients of determination (R2), the value of the proposed method (0.82) was nearly equivalent to one of values obtained by the conventional methods. Meanwhile, as for standard deviation, the proposed method showed the smallest value in the comparison. The results indicated usability of the PCA and PC scores calculated from the CEFR subcategory data for objective evaluation of L2 learners’ English proficiencies.
|
|---|
