TheraMind AI: A Multilingual and Multidomain Diagnostic Assistant for Psychotherapy

TheraMind AI will serve as a modular platform for conversing, capable of sensing emotions and aimed at facilitating psychotherapy-related dialogues within a multilingual and diverse cultural context. The architecture of the system will incorporate emotion detection, retrieval of evidence-based knowledge, and the synthesis of safe responses. Rather than functioning as a fully-fledged diagnostic tool, this work focuses on validating an initial analytical component of the platform, specifically concerning the automated identification of emotions in brief text conversations. An emotion detection module is suggested, utilizing a multilingual DistilBERT model that has been fine-tuned on datasets annotated with emotional content. Despite the issue of class imbalance, the study addresses the semantic connections among various fine-grained emotional labels by aligning the original labels with an aggregated set of broader, high-level emotional categories. The outcome indicates a balanced performance index across the classes, as evidenced by a prominent diagonal in the confusion matrices and elevated macro and weighted F1-scores. This finding confirms the capability of multilingual Transformers to effectively recognize the various emotional states that can be easily identified using affect-sensing dialogue systems, which serves as a crucial basis for a dialogue platform in psychotherapy.