RoBERTa-based Emotion Detection using Ekman Mapping: A Multi-Label Classification Approach on Go-Emotions Dataset

Abstract

Emotion detection in text is a fundamental task in natural language processing, with applications ranging from mental health monitoring to customer sentiment analysis and human-computer interaction. While recent transformer-based models have significantly improved emotion recognition performance, challenges such as class imbalance, emotion overlap, and the complexity of fine-grained emotion taxonomies remain largely unresolved. In this paper, we propose a multi-label emotion detection framework based on a RoBERTa architecture and evaluate it using the Go-Emotions dataset. To achieve a practical balance between emotional granularity and classification reliability, the original 27 fine-grained emotion categories are systematically mapped to five of Ekman's basic emotions. The proposed approach incorporates a robust preprocessing pipeline and effective class-imbalance handling strategies, enabling reliable multi-label prediction under real-world conditions. Experimental results on the test set demonstrate strong overall performance, achieving a micro-averaged F1-score of 0.83 and a macro-averaged F1-score of 0.76. Per-class analysis further indicates consistent performance across all emotion categories, including less frequent classes. These findings suggest that Ekman-level emotion classification offers an effective and interpretable framework for practical emotion detection systems. © 2026 IEEE.