A hybrid stacked ensemble learning framework for multilabel text emotion detection.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Sentiment analysis traditionally uses binary (positive/negative) classification.
• Accurate emotion detection requires multi-label classification due to complex, overlapping human emotions.
• Transformer models face challenges in low-resource languages and diverse contexts due to data scarcity and generalization issues.

Purpose of the Study:

• To propose Hyb-Stack, a hybrid stacked ensemble framework for enhanced multi-label emotion classification.
• To improve classification accuracy, adaptability, and cross-lingual generalization.
• To address the lack of annotated data in low-resource languages for emotion detection.

Main Methods:

• Developed Hyb-Stack, integrating simple and cross-validation stacking.
• Combined predictions from BERT, DistilBERT, RoBERTa, and mBERT using a Random Forest meta-classifier.
• Evaluated on English (SemEval-2018 Task 1 E-c), Bahasa Indonesia (hate speech), and Hausa (HaEmoC_V1) datasets.

Main Results:

• The EM-9 ensemble (BERT + DistilBERT + mBERT) achieved superior F1-scores: 89.48 (Hausa), 88.19 (Bahasa Indonesia), and 90.67 (English).
• mBERT individually outperformed other base models across all datasets.
• Hyb-Stack surpassed conventional ensemble methods like averaging and weighted averaging.

Conclusions:

• Hyb-Stack effectively enhances multi-label emotion classification across diverse linguistic contexts.
• Combining multiple transformer models with an optimized decision layer is crucial for advancing emotion detection.
• The framework demonstrates significant potential for low-resource language applications in sentiment analysis.