Addressing Data Imbalance in Transformer-Based Multi-Label Emotion Detection with Weighted Loss
Xia Cui
TL;DR
This work tackles data imbalance in multi-label emotion detection by applying a simple, dynamic class-weighted Binary Cross-Entropy loss to Transformer backbones. Evaluating BERT, RoBERTa, and BART on BRIGHTER English Track A, the approach improves performance on high-frequency emotion classes, with RoBERTa showing robustness to imbalance and BART benefiting most from weighting. The study computes standard metrics including Micro/F1, Macro/F1, ROC-AUC, Accuracy, and Jaccard, and finds that minority-emotion gains are limited, highlighting both the utility and the constraints of weighted loss in imbalanced, multi-label settings. Overall, the method offers a computationally light alternative to resampling while underscoring the remaining challenges in minority-emotion detection across languages and datasets.
Abstract
This paper explores the application of a simple weighted loss function to Transformer-based models for multi-label emotion detection in SemEval-2025 Shared Task 11. Our approach addresses data imbalance by dynamically adjusting class weights, thereby enhancing performance on minority emotion classes without the computational burden of traditional resampling methods. We evaluate BERT, RoBERTa, and BART on the BRIGHTER dataset, using evaluation metrics such as Micro F1, Macro F1, ROC-AUC, Accuracy, and Jaccard similarity coefficients. The results demonstrate that the weighted loss function improves performance on high-frequency emotion classes but shows limited impact on minority classes. These findings underscore both the effectiveness and the challenges of applying this approach to imbalanced multi-label emotion detection.