DeCoDE: Detection of Cognitive Distortion and Emotion Cause Extraction in Clinical Conversations

Despite significant evidence linking mental health to almost every major development issue, individuals with mental disorders are among those most at risk of being excluded from development programs. We outline a novel task of detection of Cognitive Distortion and Emotion Cause extraction of associated emotions in conversations. Cognitive distortions are inaccurate thought patterns, beliefs, or perceptions that contribute to negative thinking, which subsequently elevates the chances of several mental illnesses. This work introduces a novel multi-modal mental health conversational corpus manually annotated with emotion, emotion causes, and the presence of cognitive distortion at the utterance level. We propose a multitasking framework that uses multi-modal information as inputs and uses both external commonsense knowledge and factual knowledge from the dataset to learn both tasks at the same time. This is because commonsense knowledge is a key part of understanding how and why emotions are implied. We achieve commendable performance gains on the cognitive distortion detection task (+3.91 F1%) and the emotion cause extraction task (+3 ROS points) when compared to the existing state-of-the-art model.