NicholasDobson/Multilabel-Emotion-Analysis-for-African-Languages

# Multilabel Emotion Analysis for African Languages **Transfer Learning + Data Augmentation on BRIGHTER & EthioEmo** **Authors:** Nicholas Dobson, Simon van der Merwe, Naazneen Khan **Course:** Natural Language Processing COS 760 – University of Pretoria (2026) --- ## Overview Natural Language Processing has largely focused on high-resource languages, leaving African languages underrepresented despite their vast diversity. This project tackles **multilabel emotion classification** for African languages using: - **BRIGHTER dataset** (28 languages) - **EthioEmo dataset** (4 Ethiopian languages) We explore how to improve performance in **low-resource settings** through: - Transfer learning (multilingual transformers) - Data augmentation (back-translation & paraphrasing) - Explainability (attention visualization) --- ## Objectives - Build strong baselines for African language emotion classification - Evaluate augmentation techniques for low-resource data - Analyse model behaviour using attention mechanisms --- ## Research Questions - **RQ1:** Which model performs best? (mBERT vs XLM-R vs AfroXLMR) - **RQ2:** How does data augmentation impact performance? - **RQ3:** What linguistic patterns do models learn across languages? --- ## Datasets - **BRIGHTER**: ~100k samples, 28 languages - **EthioEmo**: 23k samples, 4 languages - Optional: **AfriSenti** (auxiliary sentiment data) ### Emotion Labels (Multilabel) - Anger - Fear - Surprise - Sadness - Happiness - Disgust --- ## Methodology ### 1. Baseline Models We fine-tune: - mBERT - XLM-RoBERTa - AfroXLMR **Setup:** - Loss: Binary Cross-Entropy (weighted) - LR: 2e-5 - Batch size: 16 - Epochs: 5 (early stopping on F1) --- ### 2. Data Augmentation - **Back-translation** (via MarianMT / Google Translate) - **Paraphrasing** (multilingual T5) **Filtering:** - Cosine similarity > 0.85 (semantic consistency) --- ### 3. Explainability - Attention visualization using **BERTViz** - Cross-lingual comparison of token …