BanglaHateLens-3D: A three-dimensional Bengali hate speech dataset with type, severity and target annotations
Description
BanglaHateLens-3D is a comprehensive, structured Bengali-language hate speech dataset comprising 50,746 annotated comments, designed to support research in natural language processing (NLP), hate speech detection, offensive language analysis, and multidimensional text classification. The dataset incorporates three complementary annotation dimensions: Type of Hate, Severity of Hate, and Target of Hate, enabling systematic analysis of harmful language, its intensity, and its intended targets. The dataset includes six hate-type classes (None, Abusive, Political Hate, Profane, Religious Hate, and Sexism), three severity levels (Little to None, Mild, and Severe), and five target classes (None, Individual, Organization, Community, and Society). Furthermore, the comments are organized into 19 topical categories and 120 subcategories, facilitating detailed analysis of linguistic patterns and harmful discourse across diverse contexts. BanglaHateLens-3D consists of 35,522 training instances, 5,024 development instances, and 10,200 testing instances. The dataset is provided in UTF-8 encoded JSON and CSV formats, ensuring accessibility and compatibility with conventional machine learning algorithms, deep learning architectures, and transformer-based language models. The repository is organized into two components: (1) Dataset, containing annotated Bengali comments and predefined training, development, and testing partitions; (2) Metadata & Others Validation Data, providing comprehensive documentation, data dictionaries, annotation specifications, class distribution statistics, structural validation reports, duplicate analysis, and SHA-256 checksum verification. To support data quality and reproducibility, the repository includes validation resources for assessing dataset structure, annotation consistency, data completeness, and file integrity. The structured organization and accompanying documentation facilitate transparent experimental workflows, systematic evaluation, and comparative benchmarking. BanglaHateLens-3D serves as a valuable resource for advancing Bengali-language hate speech research, particularly in low-resource NLP, multidimensional classification, severity assessment, target identification, automated content moderation, and responsible artificial intelligence. Its multidimensional annotation framework supports the development and evaluation of robust computational models for identifying and analyzing harmful online communication.
Files
Institutions
- Daffodil International UniversityDhaka Division, Dhaka