Olamieee/yoruba-emotion

Fine-tuning AfroXLMR on SemEval-2025 Task 11 for multilabel emotion detection in Yorùbá, a low-resource Niger-Congo language spoken by over 50 million people across Nigeria and West Africa. # Yoruba Multilabel Emotion Detection Fine-tuning AfroXLMR on SemEval-2025 Task 11 for multilabel emotion detection in Yorùbá, a low-resource Niger-Congo language spoken by over 50 million people across Nigeria and West Africa. --- ## Overview This project fine-tunes AfroXLMR-base — an Africa-centric multilingual transformer — on the Yorùbá split of the SemEval-2025 Task 11 shared task on text-based emotion detection. The model predicts six emotions simultaneously from Yorùbá text: **anger, disgust, fear, joy, sadness, and surprise**. Emotion detection in Yorùbá is a challenging task due to: - Severe class imbalance (sadness has 10x more samples than disgust or fear) - Morphological complexity and tonal diacritics (e.g. è, ó, ẹ) that standard NLP tools strip incorrectly - Limited annotated data compared to high-resource languages --- ## Results Evaluated on the official SemEval-2025 Task 11 Yorùbá test set (3,000 samples). | Emotion | Precision | Recall | F1 | |---------|-----------|--------|----| | Anger | 0.40 | 0.39 | 0.39 | | Disgust | 0.30 | 0.33 | 0.31 | | Fear | 0.39 | 0.27 | 0.32 | | Joy | 0.37 | 0.46 | 0.41 | | Sadness | 0.59 | 0.77 | 0.66 | | Surprise | 0.27 | 0.31 | 0.29 | | **Macro avg** | **0.39** | **0.42** | **0.40** | **Comparison to SemEval-2025 Task 11 baselines (Yorùbá Track A):** | Model | Macro F1 | |-------|----------| | Majority class baseline | 0.165 | | RoBERTa baseline | 0.463 | | **AfroXLMR-base (this work)** | **0.40** | | Best SemEval team (ensemble) | 0.657 | Our single-model AfroXLMR approach achieves 0.40 macro F1, significantly above the majority baseline and competitive with the RoBERTa baseline, without any ensembling, data augmentation, or external data. --- ## Model The fine-tuned model and tokenizer are publicly available on HuggingFace: - Model: Olamieee/yoruba-emotion-model - Tokenizer: Olamieee/yoruba-emotion-tokenizer ### Load and run inference ```python from transformers import AutoModelForSequenceClassifi …