Automatic CEFR Level Prediction for Estonian Learner Text

This paper reports on approaches for automatically predicting a learner’s language proficiency in Estonian according to the European CEFR scale. We used the morphological and POS tag information extracted from the texts written by learners. We compared classification and regression modeling for this task. Our models achieve a classification accuracy of 79% and a correlation of 0.85 when modeled as regression. After a comparison between them, we concluded that classification is more effective than regression in terms of exact error and the direction of error. Apart from this, we investigated the most predictive features for both multiclass and binary classification between groups and also explored the nature of the correlations between highly predictive features. Our results show considerable improvement in classification accuracy over previously reported results and take us a step closer towards the automated assessment of Estonian learner text.