ismailateam/-SudaHate-Sudanese-Hate-Speech-and-Abusive-Dataset

A dataset of Sudanese Arabic text labeled for hate, abusive, and normal content, part of a multi-dialect study. # SudaHate – Sudanese Hate Speech and Abusive Dataset ## Overview **SudaHate** is The first Sudanese Arabic Hate Speech and Abusive Language Dataset designed to support research on toxicity detection in Sudanese dialect Arabic. With the rapid expansion of social media usage in Sudan, online platforms have experienced increasing levels of toxic discourse including abusive and hate speech. SudaHate consists of 6908 Sudanese comments labeled as: - Normal - Abusive - Hate --- ## 📊 Dataset Statistics - Total instances: 6908 - Normal: 4859 - Abusive: 1063 - Hate: 986 - Train/Test split:70% / 30% (4836 / 2072) --- ## 🗂 Data Collection The SudaHate dataset was constructed from two complementary sources to ensure both diversity and dialectal authenticity. ### 1️⃣ Lîsan Corpus (Sudanese Subset) Part of the SudaHate dataset incorporates the Sudanese subset of the Lîsan corpus: Jarrar, M. et al. (2023). *Lîsan: A corpus, dictionary, and POS tagger for dialectal Arabic processing*. Language Resources and Evaluation, 57, 1025–1067. Official dataset page: sina.birzeit.edu The Lîsan corpus is publicly available and permits redistribution under its original license. In accordance with those terms, we include the Sudanese subset (3,000 instances) in this repository. Since the original Lîsan corpus was not annotated for hate or abusive language, all reused instances were fully re-annotated in this study using our unified three-class annotation scheme (Normal, Abusive, Hate). The original Lîsan license terms remain applicable to the redistributed portion included in this repository. ### 2️⃣ Facebook Data Collection To enhance coverage of contemporary Sudanese online discourse, we collected publicly accessible posts and comments from Sudanese Facebook pages using the Facebook Graph API via the Facepager tool. Data collection was conducted in compliance with Facebook’s publicly available data access policies at the time of collection. Only …