Multilabel Emotion Detection in Yoruba: Fine-tuning AfroXLMR on SemEval-2025 Task 11
Domaine:
natural language processingType de record:
Abstract
Emotion detection in low-resource African languages remains a significant open challenge in natural language
processing. In this paper, we present a study on multilabel emotion detection in Yoruba, a Niger-Congo language
spoken by over 50 million people, using the SemEval-2025 Task 11 shared task dataset. We fine-tune
AfroXLMR-base, an Africa-centric multilingual transformer model, on the official training split and evaluate on
the held-out test set of 3,000 samples. To address severe class imbalance across six emotion categories, we apply
weighted binary cross-entropy loss with a capped positive weight strategy. Through systematic threshold tuning
we achieve a macro F1 of 0.40 on the official test set, compared to the published RoBERTa baseline of 0.463
and a majority class baseline of 0.165. Our analysis reveals that sadness, the most frequent emotion class,
achieves the highest F1 of 0.66, while disgust and fear remain challenging due to data scarcity. We release our
fine-tuned model publicly on HuggingFace and discuss implications for single-model approaches to lowresource African language emotion detection.
Keywords: emotion detection, multilabel classification, Yoruba NLP, AfroXLMR, low-resource NLP,
SemEval-2025, African languages