Amani-Bara/Deepfake-Text-DetectionSystems-for-Arabic-andEnglish-Using-BERT-and-ELECTRA-Based-Models

Transformer-based NLP system for detecting AI-generated text in Arabic and English using models such as AraBERT, AraELECTRA, BERT, and DeBERTa. # Deepfake Text Detection for Arabic and English Using BERT and ELECTRA-Based Models A research project investigating how effectively fine-tuned transformer models can distinguish between human-written and AI-generated text, across both Arabic and English. **Authors:** Amani Albarazi, Raghad Alamoudi, Maram Alhusami **Supervised by:** Dr. Naila Marir **Course:** CS4083 - Natural Language Processing & Text Mining, Effat University **Date:** December 2025 ## Overview The rapid rise of large language models has made AI-generated text nearly indistinguishable from human writing, raising concerns around academic integrity, misinformation, and authorship authenticity. This project fine-tunes and evaluates transformer-based models for binary classification of human vs. AI-generated text, with dedicated pipelines for both Arabic and English. ## Datasets - **ALHD** (A Large-Scale and Multigenre Benchmark Dataset for Arabic LLM-Generated Text Detection) — 20,268 Arabic samples (a balanced 10% subset of the full corpus), spanning news, tweets, reviews, health content, and conversational text. AI-generated portion produced by GPT-3.5-T. - **balanced_ai_human_prompts** — 2,750 English samples, evenly split between human-written and GPT-3/GPT-3.5-generated prompts. ## Models **Arabic:** - AraBERT-Base - AraBERT-Large - AraELECTRA **English:** - DistilBERT - BERT-base (EngBERT) - DeBERTa-v3-base ## Methodology 1. Data cleaning and preprocessing (URL removal, lowercasing, whitespace normalization for English; tokenizer-based normalization for Arabic) 2. Tokenization with truncation/padding (max length 128 for Arabic, 256 for English) 3. Fine-tuning via HuggingFace `Trainer` (Arabic) and a custom PyTorch training loop (English) 4. Evaluation using accuracy and F1-score 5. Unified inference function for real-time prediction across all models ## Results **Arabic models:** | Model | Accuracy | F1-Score | Generalization | |---|---|---|---| | AraBERT-Large | 92.97% | 0.9327 …