FHFouzi Harrag

Fouzi Harrag

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Fouzi Harrag has a total of 35 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    GraphSkinUQ: A Hybrid Transfer Learning Approach for Feature Extraction and Graph Transformer-Based Uncertainty Quantification for Robust Skin Cancer Classification

    by Ouissem Touameur, Fouzi Harrag, Mohamed Deriche, Ahmed Bouridane

    The early and reliable diagnosis of skin cancer is critical to mitigating its metastatic potential and progression. However, even highly accurate artificial intelligence (AI) classifiers tend to produce overconfident predictions on ambiguous lesions, increasing the risk of misdiagnosis. Uncertainty quantification (UQ) addresses this challenge by evaluating model confidence and distinguishing definitive classifications from cases requiring clinician review—a cornerstone for safely integrating AI into clinical practice. We propose GraphSkinUQ, a hybrid framework that integrates convolutional neural network (CNN)-based feature extraction and graph-based relational reasoning to model lesion contextual relationships, together with Bayesian Monte Carlo dropout, to simultaneously quantify epistemic uncertainty (model limitations) and aleatoric uncertainty (data noise). Through iterative sampling of the predictive posterior distribution, GraphSkinUQ generates well-calibrated confidence scores, evaluated using average predictive entropy (0.10 bits), Brier score (0.1382), and expected calibration error (ECE). Experimental results show that GraphSkinUQ achieves a Brier score of 0.1382, Receiver Operating Characteristic (ROC) area under the curve (AUC) of 0.954, and average predictive entropy of 0.10, outperforming conventional classifiers in calibration. The framework effectively identifies high-uncertainty cases, allowing clinicians to prioritize ambiguous lesions for additional scrutiny. By bridging AI-driven diagnostics with clinically interpretable confidence metrics, GraphSkinUQ enhances the safety and transparency of automated skin cancer screening, fostering trustworthy and robust human–AI collaboration in clinical decision-making, and aligning machine-driven insights with the nuanced demands of real-world oncology practice.

  • Open Access

    ADL-KG: Diacritic-Aware Knowledge Graph Prompting for Arabic LLM Question Answering

    by Narimene Ayat, Fouzi Harrag, Nassir Harrag, Khaled Shaalan

    Arabic’s complex morphological system and the optional use of short vowels (tashkīl) introduce substantial lexical ambiguity, posing significant challenges for Large Language Models (LLMs). While diacritics enhance linguistic precision, LLMs trained predominantly on undiacritized corpora often exhibit performance degradation when processing fully diacritized inputs due to representation shifts and tokenization inconsistencies. To address this limitation, we propose the Arabic Diacritic Lexical Knowledge Graph (ADL-KG), a structured framework that links diacritized and undiacritized forms through integrated lexical, morphological, and semantic knowledge. Building upon this resource, we introduce Diacritic-Aware Knowledge Graph Prompting (DA-KGP), a prompt augmentation strategy that injects explicit linguistic features into LLM inputs to facilitate robust interpretation of diacritized Arabic text. The framework is evaluated on the Arabic Reading Comprehension Dataset under zero-shot and few-shot question answering across AraGPT2-base, BLOOMZ-560M, SILMA-v1, and LLaMA 3.1-8B. Performance is assessed using Exact Match, BLEU, ROUGE-1, and BERTScore-F1. Experimental results show that fully diacritized prompts significantly degrade baseline performance, whereas DA-KGP consistently mitigates this effect by improving semantic alignment across diverse architectures. For AraGPT2-base, KG augmentation improves average BERTScore-F1 by +5.96 points. SILMA-v1 achieves the strongest lexical improvements, reaching 21.57 BLEU and 81.31% BERTScore-F1 in the KG-enhanced two-shot configuration. LLaMA 3.1-8B achieves the highest overall semantic performance with 82.54% BERTScore-F1 under KG-enhanced prompting, while BLOOMZ-560M also demonstrates statistically significant semantic gains through structured augmentation. These findings demonstrate that morphologically informed prompting and structured lexical grounding provide an effective and parameter-efficient strategy for improving the robustness and semantic fidelity of Arabic LLMs under fully diacritized input conditions.

  • Trustworthy Arabic LLMs: A Multidimensional Survey of Risks, Realities, and Resilience

    by Narimene Ayat, Fouzi Harrag, Nassir Harrag, Mohamed Deriche, Yahya Mohamed Elhadj

  • Open Access

    Towards Unified Multimodal Search and Generation System: An AI-Driven Text–Image Framework for Arabic

    by Fouzi Harrag, Narimene Ayat, Aymen Abderrahim Redjechta, Aya Lamria Rouabhi

    Recent advances in multimodal learning have significantly enhanced natural language understanding (NLU) and information retrieval across various languages; however, Arabic remains substantially underrepresented in this domain. In this work, we propose an innovative deep learning framework that draws inspiration from the Unified Multimodal Search and Generation System paradigm, combined with OpenAI’s CLIP and DALL⋅E architectures, to enhance Arabic-language image retrieval and generation. Given the highly inflected nature of Arabic and the limited application of modern deep learning methods in Arabic NLP, we fine-tune these models on an Arabic dataset to enable effective text–image alignment as well as the generation of semantically coherent images from Arabic queries. Our findings demonstrate the potential of the proposed framework to advance Arabic multimodal processing and contribute to narrowing the gap between Arabic NLP and recent developments in multimodal learning.

  • Open Access

    A Dual-Adversarial Learning Framework with Explicit Constraints for Two-Sided Fairness in Recommendation System

    by Sabrina Karboua, Fouzi Harrag, Farid Meziane

    Recommendation systems frequently exhibit two fundamental types of bias: user-side bias, where recommendations discriminate based on protected attributes such as gender, and item-side bias, where popular items dominate exposure irrespective of their actual relevance or quality. To address both simultaneously, we propose a dual-adversarial framework with two discriminators: a user-gender discriminator designed to mitigate gender bias by removing sensitive information from user embeddings, and an item-popularity discriminator aimed at reducing popularity bias among items. While adversarial learning encourages the model to generate bias-free embeddings by fooling discriminators, residual sensitive information may persist due to imperfect discriminator inference and distributional discrepancies. To further reduce this residual bias and enforce balanced, fair recommendations, we enhance the dual adversarial learning framework with two explicit fairness constraints that operate on the learned debiased user and item embeddings: an individual user fairness constraint that guarantees equitable recommendations for users with similar preferences, and a popularity regularization term that balances exposure across popular and less popular item groups. Experiments on three real-world datasets (MovieLens 1M, MovieLens 100K, and Book-Crossing) demonstrate significant gender predictability drops and popularity skew decreases, while maintaining competitive recommendation quality.

  • Leveraging AraBERT for COVID-19 event monitoring on Arabic Twitter

    by Fouzi Harrag, Ouissem Touameur, Maroua Zermani

    Purpose: The outbreak of COVID-19 has posed a significant public health threat, prompting the need for effective prediction and tracking of the virus’s spread. This study focuses on leveraging social media data, particularly Arabic Twitter posts, to predict and monitor COVID-19-related incidents. We aim to explore how natural language processing (NLP) techniques can be applied to extract meaningful information from Arabic tweets, offering valuable insights for health scientists and policymakers.Design/methodology/approach: Given the challenges associated with Arabic NLP, including its morphological richness and ambiguity, traditional word embedding models often fail to capture the context accurately. To address this issue, we propose a deep learning-based approach utilizing AraBERT, the state-of-the-art language representation model for Arabic. By fine-tuning AraBERT, we aim to extract named entities and identify key events related to COVID-19, enabling better understanding and tracking of the disease’s spread through social media. The data are sourced from Arabic-language Twitter posts, focusing on the COVID-19 pandemic, and processed using deep learning techniques.Findings: Our approach demonstrates that AraBERT, when fine-tuned for COVID-19-related tweets, can effectively capture context-rich entities and events, providing a reliable framework for real-time monitoring of the pandemic. This model outperforms traditional NLP techniques, showcasing its ability to handle the complexities of the Arabic language and improve the accuracy of event extraction in the context of social media.Originality/value: This research introduces a novel method for monitoring COVID-19 using Arabic social media, highlighting the potential of deep learning models, particularly AraBERT, in enhancing information extraction from complex, context-heavy languages. The study contributes valuable insights into how social media can be used as a tool for epidemics.

  • Hybrid NLP model for automating fashion product descriptions: integrating transformers and word embeddings

    by Fouzi Harrag, Ouissem Touameur, Hamza Behilil

    Purpose: This study aims to address the challenge of generating accurate and engaging product descriptions for e-commerce platforms, particularly in the fashion domain. It seeks to alleviate the labor-intensive and time-consuming process of manual description writing by leveraging advanced natural language processing (NLP) techniques.Design/methodology/approach: The proposed solution integrates GPT-Neo, a transformer model, with the word-embedding model word2vec to automate product description generation. A dataset comprising 14,000 product titles and descriptions was sourced from Noon, a prominent Arabic e-commerce platform, and used to fine-tune the models for specific fashion categories.Findings: The results demonstrate that the developed system effectively generates product descriptions based on product titles, achieving a recall rate of 67% and a precision of 72%. These findings validate the system’s potential to reduce manual effort while maintaining description quality.Originality/value: This research offers a novel approach to automating product description generation for Arabic e-commerce platforms. It combines state-of-the-art NLP techniques to address a significant bottleneck in the e-commerce industry, contributing to enhanced operational efficiency and scalability. The study’s outcomes also pave the way for further advancements in multilingual NLP applications.


  • Top Cited Papers
    Open Access

    Human vs. Machine: A Comparative Study on the Detection of AI-Generated Content

    by Amal Boutadjine, Fouzi Harrag, Khaled Shaalan

    The surge in advancements in large language models (LLMs) has expedited the generation of synthetic text imitating human writing styles. This, however, raises concerns about the potential misuse of synthetic textual data, which could compromise trust in online content. Against this backdrop, the present research aims to address the key challenges of detecting LLMs-generated texts. In this study, we used ChatGPT (v 3.5) because of its widespread and capability to comprehend and keep conversational context, allowing it to produce meaningful and contextually suitable responses. The problem revolves around the task of discerning between authentic and artificially generated textual content. To tackle this problem, we first created a dataset containing both real and DeepFake text. Subsequently, we employed transfer-learning (TL) and conducted DeepFake-detection utilizing SOTA large pre-trained LLMs. Furthermore, we conducted validation using benchmark datasets comprising unseen data samples to ensure that the model's performance reflects its ability to generalize to new data. Finally, we discussed this study's theoretical contributions, practical implications, limitations and potential avenues for future research, aiming to formulate strategies for identifying and detecting large-generative-models’ produced texts. The results were promising, with accuracy ranging from 94% to 99%. The comparison between automatic detection and the human ability to detect DeepFake text revealed a significant gap in the human capacity for its identification, emphasizing an increasing need for sophisticated automated detectors. The investigation into AI-generated content detection holds central importance in the age of LLMs and technology convergence. This study is both timely and adds value to the ongoing discussion regarding the challenges associated with the pertinent theme of "DeepFake text detection", with a special focus on examining the boundaries of human detection.