CONFOBCE-MLTC: A NOVEL CONTRASTIVE FOCAL BINARY CROSS ENTROPY LOSS FUNCTION FOR MULTI-LABEL TEXT-BASED EMOTION CLASSIFICATION

Abstract

Multi-label text classification remains challenging due to class imbalance, overlapping label distributions, and inter-label dependencies. Existing loss functions often address these issues independently, resulting in limited robustness and inconsistent generalization. To address these limitations, this paper proposes ConFoBCE-MLTC, a correlation-aware composite loss function that integrates binary cross-entropy, focal modulation, contrastive learning, and label correlation alignment within a unified optimization framework. The proposed method improves minority-class sensitivity, enhances feature separability, and explicitly models label co-occurrence structures without requiring architectural modifications. The approach was evaluated using multiple transformer-based architectures, including BERT, RoBERTa, DistilBERT, and mBERT, across both high-resource and low-resource datasets, including the newly constructed Hausa Emotion Corpus (HaEmoC_V1). Experimental results demonstrate consistent improvements over conventional loss functions and recent asymmetric loss formulations. In particular, mBERT combined with ConFoBCE-MLTC achieved the strongest overall performance, reaching 82.24 Micro-F1, 80.18 Macro-F1, and 73.31 Jaccard Score on the newly constructed dataset (HaEmoC_V1). Additional ablation and sensitivity analyses confirm the complementary contribution of focal weighting, contrastive representation learning, and label correlation alignment. These findings demonstrate the effectiveness of the proposed framework for multilingual multi-label emotion classification, particularly in low-resource settings.