Emotion Recogniton in Conversations - empirical study

Rufaida Kashif1, Benjamin Piwowarski1, Helena Gomez Adorno2
1Sorbonne University, 2National Autonomous University of Mexico


Abstract

Emotion Recognition in Conversations (ERC) requires modeling complex contextual dependencies across dialog turns. While transformer-based models achieve strong performance on ERC benchmarks, several key design choices including context construction, optimization strategies, and imbalance handling remain insufficiently examined. In this work, we conduct a systematic empirical study of transformer-based ERC models across three benchmark datasets. We analyze the impact of context length and directionality, layer freezing, learning rate scheduling, parameter-efficient fine-tuning, and class imbalance mitigation strategies. Our results show that short-to-medium conversational context and moderate layer freezing provide stable and strong performance, while very long context windows, aggressive freezing, and parameter-efficient adaptation offer limited gains. Furthermore, imbalance-aware losses and data augmentation do not consistently outperform standard cross-entropy training. Overall, our findings provide practical insights into effective and stable design choices for transformer-based conversational emotion recognition.