Explainable Context-Aware Emotion Recognition in Conversations Through Attribution Analysis, Layer-Wise Probing and Representation Geometry

Abstract

Emotion recognition in conversations is a fundamental capability for advanced human–machine interaction. Its main aim is to enable dialogue systems to adapt their content, tone and pacing to users’ affective state and improve engagement, satisfaction and trust. In this study, we implement and evaluate emotion classification models and compare single-utterance and context-aware formulations that incorporate information from neighboring dialogue turns, with the aim to understand when and how context helps disambiguate emotional utterances. We also investigate how fine-tuning reshapes internal representations and study the geometry of CLS embeddings. Finally, we apply interpretability techniques that combine token-level attribution with layer-wise diagnostics to characterize how emotional cues accumulate across the Transformer stack and how they redistribute after fine-tuning. The findings indicate that contextual formulations consistently outperform single-utterance baselines and that fine-tuning reorganizes the embedding space into more coherent, emotion-aligned clusters. We also examine the impact of class imbalance on model behavior and confusion patterns across emotion categories. The study should be interpreted primarily as an explainability-oriented analysis of context-aware ERC rather than as an attempt to introduce a new state-of-the-art architecture. Its main contribution lies in the intersection of performance, attribution evidence, layer-wise analysis and representation geometry. Together, these analyses show how fine-tuning reorganizes model representations and how conversational context affects predictions.