Explainable Transformer Models for Human Emotion Recognition: A Multi-Method Explainability Study in the Context of Mental Health

The ability to identify emotions based on written text is one of the core areas of Natural Language Processing (NLP) and has many applications in areas such as mental health monitoring, sentiment analysis, and dialogue systems. This study proposes an explainable emotion recognition (EER) framework built on a fine-tuned RoBERTa-base model trained on the Emotions for NLP dataset with an accuracy of 92.4% and a weighted F1 score of 92.5%. To interpret the decision process of the EER model, we systematically applied four complementary explainable artificial intelligence (XAI) techniques to provide explanations and insights into how the model makes its predictions: SHAP for global token-level feature attribution, LIME for local instance-level explanations, multi-head attention visualization for structural interpretability, and integrated gradients via Captum for axiom-satisfying gradient-based attribution. Each of these four methods provides complementary multi-perspective views of EER model behavior, which can help increase model transparency, identify potential biases, and enable the responsible use of transformer-based models in critical environments (e.g., those requiring formal clinical documentation). Our experiments consistently show that the EER model identifies tokens as having the highest emotional expression level as the strongest predictive feature across methodological perspectives, with strong evidence of cross-methodological agreement regarding the semantic coherence of learned representations. Our findings have direct implications for the responsible implementation of AI-based emotion recognition systems in mental health support systems, where model user-interface transparency, bias mitigation, and clinical trust are necessary to ensure quality patient care.