fgaim/TiALD

A Multi-task Benchmark for Abusive Language Detection in Tigrinya Social Media # Tigrinya Abusive Language Detection (TiALD) Dataset **Tigrinya Abusive Language Dataset (TiALD)** is a large-scale, multi-task benchmark dataset for abusive language detection in the Tigrinya language. It consists of **13,717 YouTube comments** annotated for **abusiveness**, **sentiment**, and **topic** tasks. The dataset includes comments written in both the **Ge’ez script** and prevalent non-standard Latin-based **transliterations** to mirror real-world usage. The dataset also includes contextual metadata such as video titles and VLM-generated and LLM-enhanced descriptions of the corresponding video content, enabling context-aware modeling. ⚠️ **The dataset contains explicit, obscene, and hateful language. It should be used for research purposes only.** ⚠️ This work accompanies the paper "A Multi-Task Benchmark for Abusive Language Detection in Low-Resource Settings", accepted at the **NeurIPS 2025** conference in the Datasets & Benchmarks Track, San Diego, December (2-7), 2025. **Outline:** - Dataset Overview - Tasks and Annotation Schema - How to access the TiALD Dataset - Baseline Models and Results - Trained Baseline Models - Code for Baseline Models - 1. Main Results: Performance of Fine-tuned and Prompted Models - 2. Performance of Models with Video Title as Context - 3. Performance of LLMs on Abusiveness Detection with Cross-Modality Context - Baseline Models Prediction Files - Dataset Details - Dataset Statistics - Dataset Features - Inter-Annotator Agreement (IAA) - Croissant Metadata for TiALD Dataset - Intended Usage of TiALD Dataset - Ethical Considerations - Evaluation and Computing Metrics - Model Predictions File Format - Computing Metrics - Citation - License ## Dataset Overview - **Data Source**: YouTube comments from 51 popular channels in the Tigrinya-speaking community. - **Scope**: 13,717 human-annotated comments from 7,373 videos with over 1.2 billion cumulative views at the time of collection. - **Sampling**: Comments selected usin …