Improving Arabic Multi-Label Emotion Classification using Stacked Embeddings and Hybrid Loss Function

Multi-label emotion classification (MLEC) for low-resource languages like Arabic faces significant challenges due to class imbalance and label correlations, particularly in accurately predicting minority emotions. This study propose a novel framework combining stacked contextual embeddings, meta-learning, and a hybrid loss function to address these issues. First, we generate enriched embeddings by fine-tuning and stacking three Arabic language models (ArabicBERT, MarBERT, AraBERT). These embeddings are processed by a Bi-LSTM meta-learner for sequence learning, followed by a fully connected network for classification. To mitigate class imbalance and leverage label dependencies, we introduce a hybrid loss integrating contrastive learning (CL), label correlation matrices (LCM), and class weighting (CW). Extensive experiments on the SemEval-2018 Task 1-Ec-Ar dataset demonstrate state-of-the-art performance, with a Jaccard accuracy of 0.81, F1-score of 0.67, and Hamming loss of 0.15. Ablation studies confirm the contributions of each component, while class-wise analysis shows our hybrid loss reduces disparities between majority and minority classes by up to 22%. Beyond Arabic MLEC, this work offers a generalizable framework adaptable to other languages and domains, advancing emotion analysis in low-resource settings.