Transparent Characterization of Affective Cues in Conversational Corpora:An Interpretable Transformer-Based Modeling Framework

Conversational corpora provide emotionally rich and contextually grounded data, making them a compelling source for affective computing tasks. These exchanges capture the shifting tones and contexts of natural speech, mirroring the complex emotional cues found in everyday human interaction. Taking advantage of the representational power of transformer-based architectures, this work utilizes RoBERTa as the foundational model for affective state classification. The optimized model discriminates and categorizes specific affective states within conversational data. The affective states considered in this work include joy, sadness, anger, disgust, fear, surprise, and neutral. Among the three transformer-based models evaluated in this, the RoBERTa based approach yielded the highest performance, by attaining an accuracy of 0.79 and a macro F1-score of 0.77 outperforming the baseline techniques.To elucidate the decision-making processes of these powerful yet opaque classifiers, this work integrates interpretability frameworks to systematically attribute each prediction to the specific tokens, dialogue turns, and contextual patterns driving model decisions. The integration of linguistic features with attention mechanisms facilitates a balance between predictive precision and interpretability, significantly advancing the field of transparent conversational analytics.