Enhanced Cognitive Distortions Detection and Classification Through Data Augmentation Techniques

Cognitive distortions detrimentally affect mental health by distorting reality and influencing emotions and behavior. While the detection and classification of such irrational thinking patterns grow in significance, limited data resources (and thereby limited work) exist for such task. In this study, we are motivated by the work in [5], where a CNN model using BERT embeddings is selected to detect and classify cognitive distortions. We explore various data augmentation techniques, such as Easy Data Augmentation, word embedding substitution, and back-translation to enhance the quality of the training dataset and fine-tune additional embeddings from RoBERTa and GPT-2 to improve the performance of these tasks. Our experimental results demonstrate a significant increase in the F-score by 1.88% for detection and 5.9% for classification. These enhancements increase the potential for building a supportive tool for individuals and mental health professionals.