Enhancing Large Language Models for Arabic Dialects Using Knowledge-Based Rethinking and Contrastive Learning
Alharbi, Abdulsalam Jameel, Shoaib Suleiman, Basem Razzak, Imran
Alharbi, Abdulsalam
Jameel, Shoaib
Suleiman, Basem
Razzak, Imran
Supervisor
Department
Computational Biology
Embargo End Date
Type
Conference proceeding
Date
License
Language
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Large Language Models (LLMs) have demonstrated remarkable capabilities across several natural language processing tasks. Nonetheless, their efficacy in multilingual and dialectally diverse contexts such as Arabic remains limited, particularly when addressing sensitive topics such as health-related information. This study presents a novel framework that combines knowledge-based Rethinking with dual Contrastive Learning (KBRCL) to improve the factual accuracy and dialectal consistency of generated responses. To assess our framework, we create a dialect-aware benchmark of 3,823 annotated health-related claims across several Arabic dialects. This benchmark enables a comprehensive assessment of both factual correctness and dialect alignment. Experimental findings indicate significant improvements in accuracy (from 46.66% to 88.94%), factual verification (89.80%) and dialect claim alignment (DCAS: 42.35%). These findings underscore the significance of integrating external knowledge sources with dialect-aware learning algorithms to provide more accurate and culturally relevant replies in Arabic NLP applications.
Citation
A. Alharbi, S. Jameel, B. Suleiman, I. Razzak, "Enhancing Large Language Models for Arabic Dialects Using Knowledge-Based Rethinking and Contrastive Learning," 2026, pp. 120-129.
Source
Lecture Notes in Computer Science
Conference
International Conference on Advances in Social Networks Analysis and Mining
Keywords
46 Information and Computing Sciences, 4605 Data Management and Data Science
Subjects
Source
International Conference on Advances in Social Networks Analysis and Mining
Publisher
Springer Nature