Automatic scoring method for open answer task in the SJ-CAT speaking test considering utterance difficulty level

In this paper, we propose an automatic scoring method for the open answer task of the Japanese speaking test SJ-CAT. The proposed method first extracts a set of features from an input answer utterance and then estimates a vocabulary richness score by human raters, which ranges from 0 to 4, by employing SVR (support vector regression). We devised a novel set of features, namely text statistics weighted by word reliability, to assess the abundance of vocabulary and expression, and degree of word relevance based on the hierarchical distance in a thesaurus to evaluate the suitability of vocabulary. We confirmed experimentally that the proposed method provides good estimates of the human richness score, with a correlation coefficient of 0.92 and an RMSE (root mean square error) of 0.56. We also showed that the proposed method is relatively robust to differences among examinees and among questions used for training and testing.