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

Recognizing emotions from written text is a very important part of Natural Language Processing (NLP) and is commonly used for feeling or sentiment analysis or keeping track of someone’s mental health status. This study uses a readable emotion-detecting framework with a RoBERTa-base model that has been modified and trained specifically for the Emotions for NLP dataset and provides an accuracy of 0.924% and f1 score of 0.925%. The main contributions of this study are the use of four different techniques that will help understand how the model works: SHAP (SHapley Additive exPlanations) provides global token credit attribution; LIME (Linear Interpretable Model-Agnostic Explanation) provides instance-level explanations; multi-head Attention Visualization provides structural interpretability; and Integrated Gradients via Captum provides gradient-based attribution using integration. The combination of these four techniques works together to improve transparency, help identify bias in the models, and support the responsible use of this model. Finally, the developers of this model performed many experiments that demonstrated the consistency with which the model could identify important emotional tokens (words or phrases) as predictive indicators of emotion.