Formulating Automated Responses to Cognitive Distortions for CBT Interactions

One of the key ideas of Cognitive Behavioural Therapy (CBT) is the ability to convert negative or distorted thoughts into more realistic alternatives. Although modern machine learning techniques can be successfully applied to a variety of Natural Language Processing tasks, including Cognitive Behavioural Therapy, the lack of a publicly available dataset makes supervised training difficult for tasks such as reforming distorted thoughts. In this research, we constructed a small CBT dataset via crowd-sourcing, and leveraged state of the art pre-trained architectures to transform cognitive distortions, producing text that is relevant and more positive than the original negative thoughts. In particular, the T5 transformer approach to multitask pre-training on a sequence-to-sequence framework, allows for higher flexibility when fine-tuning on the CBT dataset. Human evaluation of the automatically generated responses showcases results that are not far behind from the overall quality of the ground truth scores.