Advancing Aspect-Based Sentiment Analysis in Course Evaluation: A Multi-Task Learning Framework With Selective Paraphrasing

Aspect-based sentiment analysis (ABSA) is essential for extracting valuable perspectives from textual data, particularly within educational contexts where understanding student feedback is vital for course evaluation. Multi-task Learning (MTL) models, which concurrently learn several related tasks, have shown promise in enhancing the performance of ABSA by leveraging shared representations. This research explores using MTL with three prominent pre-trained language models (PLMs): MTL-BERT, MTL-RoBERTa, and MTL-XLNet. To further enhance ABSA performance, we integrate a data augmentation method—selective paraphrasing—with the MTL-based PLMs, including SP-MTL-BERT, SP-MTL-RoBERTa, and SP-MTL-XLNet, aimed at enriching the training dataset without compromising the integrity of aspect terms. Additionally, a nuance control mechanism is integrated into the selective paraphrasing process to preserve sentiment intensity and polarity, ensuring semantic consistency and minimizing unintended sentiment drift in the augmented data. However, the scarcity of diverse and comprehensive training data can hinder the effectiveness of these models. For this study, we develop a custom academic dataset by collecting student feedback data on course evaluations comprising more than 11,000 comments from a public university. We conducted experiments on MTL-based models for the two sub-tasks of ABSA, aspect extraction, and sentiment classification, both with and without the integration of selective paraphrasing. The experimental findings indicate that our approach substantially improves performance across all models. Specifically, the SP-MTL-BERT model achieved the highest performance, showing an improvement of +11.0 in aspect recall, +10.8 in aspect F1-score, and +3.7 in sentiment precision, establishing it as the best-performing model in our study. Moreover, comparative analysis with baseline approaches and other data augmentation techniques, such as back translation and easy data augmentation, demonstrates the superior performance of our proposed approach, particularly in improving aspect classification recall and F1 scores.