Multi-Label Emotion Recognition in Low-Resource Dialects: A Case Study on Algerian Arabic with Large Language Models
Hanane Boutouta, Ferial Senator, Zakia Lakab, Chahrazed Mediani, Abdelaziz Lakhfif · Procedia Computer Science · 2026
Emotion Recognition (ER) is a significant challenge in Natural Language Processing (NLP), as it aims to categorize text based on basic human emotions like joy, sadness, and anger. Due to the limited available linguistic resources and the distinctive linguistic features of Arabic dialects, this task becomes more difficult, particularly for the Algerian dialect. This study offers an in-depth assessment of both general-purpose and Arabic-specific Large Language Models (LLMs) for the task of multi-label emotion detection in the Algerian dialect. Despite the limited number of pre-trained LLMs specifically tailored for the Algerian dialect, we evaluate the performance of six well-known LLMs: three Arabic-specific open-source models (DziriBERT, AraBERT-Algerian, MARBERTv2) and three general-purpose models (ChatGPT, Gemini, DeepSeek). The Algerian Arabic subset of the SemEval-2025 Task 11 dataset, which is based on literary texts and has been manually annotated by native speakers, is used for the experiments. Evaluation against established metrics demonstrates promising results. The Gemini model in a zero-shot setting achieved a competitive macro F1-score of 66.53%, which is very close to the performance of the top-ranked system in SemEval-2025 Task 11. With a macro F1-score of 60.21%, the refined AraBERT-Algerian model outperformed SemEval-2025 baselines by a significant margin. These results show that zero-shot inference and fine-tuning techniques can both be successful in low-resource environments, highlighting the potential of both general-purpose and dialect-specific LLMs for multi-label emotion detection in the Algerian dialect.