CODIPAD: a dataset for the detection of cognitive distortions in patient discourse

The CODIPAD dataset (Cognitive Distortions in Patient Discourse) is an expert-annotated resource designed to facilitate research in natural language processing and mental health by supporting the automatic detection and classification of cognitive distortions in written patient discourse. It provides expert-annotated, fine-grained textual fragments that enable both binary (distortion vs. no distortion) and multiclass classification tasks aligned with Cognitive Therapy theory. The CODIPAD dataset consists of 5,055 expert-annotated fragments of written patient discourse. The data is organized into 11 classes, including 10 types of cognitive distortions based on cognitive therapy theory and a no-distortion category, enabling binary and multi-class classification tasks. All fragments were annotated by clinical psychologists using a consensus-based expert annotation protocol that allows for detailed segmentation of patient discourse into multiple labeled units. To address class imbalance, synthetic fragments were generated for underrepresented distortion types, reviewed and approved for inclusion by a clinical expert, and then annotated using the same expert annotation process as the organic data. Each record includes the patient’s complete original text, the extracted fragment, a unique fragment identifier, an origin field distinguishing organic from synthetically generated fragments, and expert labels for dominant and secondary cognitive distortions.