LLM-driven Text Normalization for Improved Meta-Learning in Arabic Sentiment Analysis
Year: 2026
Type: article
Abstract: Arabic texts posted online are written in different dialects and styles. This makes the Arabic content very sparse, especially when the texts come from different domains such as reviews of hotels, books and others. These characteristics cause low accuracy when a sentiment classifier is trained on one domain and used on another. Few-shot classification, as implemented in meta-learning, is a promising solution. The problem with these algorithms is that they require a minimum level of similarity between domains, and to our knowledge no Arabic multi-domain dataset meets this condition. We propose a new approach that uses an LLM (Gemini 2.5 Pro) to apply normalization, preprocessing and conversion to MSA to existing Arabic datasets, which reduces the sparsity and variety of Arabic multi-domain data. We measured this reduction directly, as a drop in the average symmetric KL divergence between domains, and we tested the approach on three datasets that we composed: classification accuracy in the few-shot setting rose by up to 7 points for MAML and 5 points for Prototypical Networks. As the LLM processing is applied only once, before the training phase, its cost is low, and the approach can be applied with limited resources and budget.
Authors: Shadi Balloul, Yasser Rahal, Kadan Aljoumaa
Institution: Higher Institute for Applied Sciences and Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7225810718