Abstract

Emotion recognition in conversations plays an important role in human–computer interaction and intelligent systems. Recent approaches in this area mainly rely on multimodal data, combining text, audio, and visual information to improve performance. However, such methods often introduce high computational complexity and are not suitable for real-time or resource-constrained environments. In this study, a lightweight text-based unimodal framework is proposed for emotion recognition in conversational data. The approach focuses on extracting contextual semantic features using a fine-tuned RoBERTa model, followed by a simple classification layer for emotion prediction. Unlike multimodal systems, the proposed framework relies only on textual information, reducing computational cost while maintaining competitive performance. The model is evaluated on the publicly available MELD dataset, and the results demonstrate that the proposed method achieves reliable classification performance with a weighted F1-score of 0.57. The findings indicate that textual information alone can provide sufficient cues for emotion recognition in many conversational scenarios. Overall, this work highlights the effectiveness of a simplified unimodal approach as a practical alternative to complex multimodal systems, especially in applications where efficiency and low latency are critical.

Authors

Preeti¹, Mohit², Manju³, Meenu Sharma⁴
Galgotias University, India¹, Ganga Institute of Technology and Management, India^2,4, Accurate Institute of Management and Technology, India³

Keywords

Emotion Recognition, Transformers, RoBERTa, Unimodal Inference, Affective Computing, Task-Specific Compression, Modality Sufficiency