Abstract
Detecting emotions in low-resource languages remains a significant challenge for modern NLP. While Large Language Models (LLMs) dominate many tasks, their computational cost and data requirements often hinder performance in multilingual settings. This paper presents the UoB-NLP system for SemEval-2025 Task 11, which utilizes parameter-efficient adapters and a novel "cycling" training strategy. Our results demonstrate that small, modular components can outperform massive LLMs in emotion detection for low-resource languages like Amharic and Tigrinya.
Introduction
Emotion detection involves identifying nuanced psychological states such as joy, anger, or disgust, which are expressed differently across the world's 7,000+ languages. Most state-of-the-art AI models are biased toward high-resource languages like English, leaving a significant "resource gap" for languages spoken in Africa and other regions. Furthermore, the computational cost of fine-tuning massive "frontier" models for every individual language is prohibitively high.
The UoB-NLP team addressed these challenges by investigating parameter-efficient ways to transfer emotional knowledge from high-resource to low-resource languages.
Methodology: The Power of Adapters
The core of the proposed system is the use of Adapters, a form of Parameter-Efficient Fine-Tuning (PEFT). Instead of retraining an entire model with billions of parameters, the researchers inserted small, trainable modules between the layers of a "frozen" pre-trained backbone, such as XLM-RoBERTa or Afro-XLMR.
Four distinct adapter strategies were evaluated:
- Standard Adapters: A task-specific module trained on top of a single language module.
- Task-Only Adapters: Training for emotion detection without specific language modules.
- Target-Language-Ready (TLR) Adapters: A "cycling" approach where the emotion module is trained by rotating through all available language adapters. This forces the model to learn language-independent emotion features.
- Language-Family TLR: This strategy cycles only through languages within the same family (e.g., Afro-Asiatic or Niger-Congo), leveraging structural and lexical similarities.
Data and Training
The study utilized several key benchmarks, including the BRIGHTER dataset (28 languages) and EthioEmo for Ethiopian languages. Underlying Language Adapters (LAs) were trained using OSCAR and Wikipedia data for 27 languages. To "boost" initial training, additional Spanish data from SemEval-2018 Task 1 was incorporated.
Key Insights and Results
The "Target-Language-Ready" (TLR) approach emerged as the most effective configuration. By exposing the task adapter to multiple languages during training, the system became significantly more robust at "zero-shot" transfer—predicting emotions in languages it had not been specifically tuned for.
The UoB-NLP system achieved exceptional rankings in the SemEval-2025 competition, particularly in Track C (Cross-lingual detection):
- 3rd Place for Amharic.
- 4th Place for Oromo, Tigrinya, Kinyarwanda, Hausa, and Igbo.
Notably, these adapter-based models (approx. 125M parameters) outperformed LLMs in 11 different languages and matched them in 4 others. This suggests that for specialized tasks in low-resource settings, parameter efficiency and targeted transfer learning can be more effective than sheer model scale.
Conclusion
The success of the TLR-TA configuration, especially with the afro-xlmr-base backbone, highlights the potential of modular AI. While adapters still face challenges in matching full fine-tuning for high-resource languages, they offer a sustainable and high-performing alternative for the multilingual long tail. Future work will explore combining the broad knowledge of LLMs with the efficiency of adapters through methods like Prompt Tuning.
Summary updated