fatima-alzahra-mohamed/tunisian-hate-speech-rnn

This project focuses on hate speech detection in the Tunisian Arabic dialect, a labeled datasets for NLP research. This repository demonstrates how various Recurrent Neural Network (RNN) architectures can be applied to this dataset to automatically classify hate speech. # HateTune: Hate Speech Detection in Tunisian Dialect Using RNN-Based Deep Learning Models ## 📌 Overview This repository presents a complete deep learning pipeline for **hate speech detection in the Tunisian Arabic dialect**, an under-resourced language variant with limited NLP tools and datasets. To address this gap, **we collected, annotated, and published HateTune — the largest publicly available Tunisian dialect hate speech dataset to date — and published a peer-reviewed research paper based on it.** Using this dataset, we implement and compare multiple **Recurrent Neural Network (RNN)** architectures for text classification, evaluating their ability to detect hate speech in social media content. --- ## 🧠 Key Features ✅ Largest Tunisian hate speech dataset (HateTune) ✅ Dataset and research paper published by the dataset authors ✅ End-to-end preprocessing and modeling pipeline ✅ Arabic-specific text cleaning using Unicode ranges ✅ Tokenization and padding with Keras ✅ Embedding layer construction ✅ Multiple RNN-based models implemented and compared: - SimpleRNN (baseline) - LSTM - GRU - Bidirectional LSTM (BiLSTM) ✅ Evaluation using Accuracy and AUC ✅ Final comparison and conclusion included --- ## 📊 Dataset: HateTune (Created and Published by Us) **Name:** HateTune – Hate Speech Detection Dataset in Tunisian Dialect **Authors:** *Our research team* **Size:** Largest available dataset for this dialect **Labels:** Hate speech vs. non-hate speech **Format:** CSV 📎 **Dataset on Kaggle:** kaggle.com 📎 **Research Paper (Springer):** link.springer.com --- ## 🧪 Models Implemented | Model | Purpose | |---------------|----------------------------------| | SimpleRNN | Baseline recurrent model | | LSTM | Captures long-term dependencies | | GRU | Efficient LSTM alternativ …