Bias detection and mitigation in arabic large language model datasets / Nour El Houda Chenaf

Public ISBD Titre : Bias detection and mitigation in arabic large language model datasets Type de document : document multimédia Auteurs : Nour El Houda Chenaf, Auteur ; Nour Alchams Telli, Auteur ; Mohamed El Habib Maicha, Directeur de thèse Editeur : Laghouat : Université Amar Telidji - Département d'informatique Année de publication : 2026 Importance : 85 p. Accompagnement : 1 disque optique numérique (CD-ROM) Note générale : Option : Artificial intelligence and data science Langues : Anglais (eng) Mots-clés : Bias detection  Bias mitigation  Arabic NLP  Large language models  Fairness  CrowS-Pairs  MADAR  AraWEAT Résumé : Bias in Arabic Large Language Models (LLMs) remains an insufficiently explored research area, particularly because of the language’s diglossia, dialectal diversity, and morphological richness, which are not adequately reflected in evaluation benchmarks originally developed for English. Most existing Arabic studies also rely on limited datasets and focus on either bias detection or mitigation separately, leaving the relationship between fairness and model performance insufficiently understood. This work presents a reusable and culturally-aware framework for detecting and mitigating bias in Arabic language models. The framework evaluates three Arabic-trained language models (AraBERT, MARBERT, and XLM-RoBERTa (XLM-R)) through an experimental pipeline that includes auditing pretrained models, fine-tuning them for sentiment analysis, re-evaluating them after adaptation, and applying several bias mitigation techniques. The evaluation relies on multiple benchmarks to measure gender and dialect bias in contextual embeddings and masked language modelling behaviour. The results show that pretrained models differ considerably in their bias levels and that finetuning can alter bias behaviour even when using a relatively simple task such as sentiment analysis. The proposed mitigation techniques successfully reduced both gender and dialect bias by more than 25% while causing only minimal impact on downstream performance. These findings demonstrate that building fairer Arabic language systems is both practically feasible and empirically measurable, providing a foundation for future Arabic Natural Language Processing (NLP) technologies that better account for fairness and reliability considerations. note de thèses : Mémoire de master en informatique Bias detection and mitigation in arabic large language model datasets [document multimédia] / Nour El Houda Chenaf, Auteur ; Nour Alchams Telli, Auteur ; Mohamed El Habib Maicha, Directeur de thèse . - Laghouat : Université Amar Telidji - Département d'informatique, 2026 . - 85 p. + 1 disque optique numérique (CD-ROM). Option : Artificial intelligence and data science Langues : Anglais (eng) Mots-clés : Bias detection  Bias mitigation  Arabic NLP  Large language models  Fairness  CrowS-Pairs  MADAR  AraWEAT Résumé : Bias in Arabic Large Language Models (LLMs) remains an insufficiently explored research area, particularly because of the language’s diglossia, dialectal diversity, and morphological richness, which are not adequately reflected in evaluation benchmarks originally developed for English. Most existing Arabic studies also rely on limited datasets and focus on either bias detection or mitigation separately, leaving the relationship between fairness and model performance insufficiently understood. This work presents a reusable and culturally-aware framework for detecting and mitigating bias in Arabic language models. The framework evaluates three Arabic-trained language models (AraBERT, MARBERT, and XLM-RoBERTa (XLM-R)) through an experimental pipeline that includes auditing pretrained models, fine-tuning them for sentiment analysis, re-evaluating them after adaptation, and applying several bias mitigation techniques. The evaluation relies on multiple benchmarks to measure gender and dialect bias in contextual embeddings and masked language modelling behaviour. The results show that pretrained models differ considerably in their bias levels and that finetuning can alter bias behaviour even when using a relatively simple task such as sentiment analysis. The proposed mitigation techniques successfully reduced both gender and dialect bias by more than 25% while causing only minimal impact on downstream performance. These findings demonstrate that building fairer Arabic language systems is both practically feasible and empirically measurable, providing a foundation for future Arabic Natural Language Processing (NLP) technologies that better account for fairness and reliability considerations. note de thèses : Mémoire de master en informatique

---

1 (1 - 1 / 1) Par page : 25 50 100 200