Abstract:

: Students’ after-class self-evaluated comments are useful for understanding students’ learning and reflecting teacher’s teaching. Researchers and engineers have attempted to apply educational data mining techniques, such as text analysis, sentiment analysis, machine learning, and deep learning to develop classification mechanisms of students’ self-evaluated comments. This study was carried out to develop aspect and sentiment classification mechanisms to automatically classify students’ self-evaluated comments into seven aspect categories and three sentiment categories. We investigated the impact of whether we should exclude nonsense data or not, the impact of different feature extraction methods, and the impact of different classification models on classification accuracy. The results showed that the combination of bidirectional encoder representations from transformers (BERT) word embedding feature extraction and Random Forest classification showed the best accuracy (90.7%) on aspect classification when including nonsense data, whereas the combination of BERT-word embedding feature extraction and Random Forest classification had the best accuracy (93.2%) on aspect classification when excluding nonsense data. Including nonsense data reduced the classification accuracies. In addition, the combination of one-word bag-of-words feature extraction and Random Forest classification presented the best accuracy (99.5%) with regard to sentiment classification.

Keywords: