Language：Japanese   Publisher：Faculty of Languages and Cultures, Kyushu University  

This study aims to reveal the importance of criterial features for the CEFR levels using CVLA, an online application for estimating the CEFR level of a given text. The indicators used in this study are BperA, VperSent, ARI, and AvrDiff, which are employed in CVLA. The data used in this study is a corpus that consists of CEFR-based English textbooks. The values of the four indicators of each text were calculated using CVLA, and then a decision tree model that predicts the CEFR level of each text was created to identify which indicator has the strongest influence on the prediction. Our results show that VperSent marked the highest variable importance, which means that it contributed the most to predicting the CEFR level of a text. However, the prediction accuracy was about 55%, which is not sufficiently reliable. Therefore, additional decision tree models were made for the adjacent levels (A1-A2, A2-B1, B1-B2, and B2-C1) to examine the most important indicator for distinguishing the neighboring levels. Consequently, VperSent was the strongest on A1-A2 and A2-B2 decision trees, BperA and AvrDiff on B1-B2, and ARI on B2-C1. This means that when the text level goes up from A1 to A2 and A2 to B1, the constructions of the sentences become more complex. The results also show that the distinction between B1 and B2 levels relies on the vocabulary level. Lastly, it was shown that the difference between B2 and C1 lies mainly in the length of sentences.

DOI： 10.15017/4773109

CiNii Research