TriDox-BN: A Three-Class Bengali Dataset for Doxxing Detection and Potential Doxxing Identification
Description
TriDox-BN is a Bengali-language social media dataset developed to support research on automated doxxing detection, privacy-risk identification, and context-aware text classification. The dataset contains 1,755 labeled samples across three categories: Doxxing (600), Potential Doxxing (574), and Non-Doxxing (581). Each sample provides conversational context, target text, and a corresponding classification label, enabling the investigation of privacy-related language and contextual information in Bengali social media. The release contains four CSV files: 1. TriDox_cleaned_dataset.csv The main dataset contains 1,755 records with four attributes: - sample_id: Unique identifier for each sample. - context_text: Surrounding conversational or social media context. - target_text: Bengali text to be classified for doxxing-related behavior. - label: Classification label (Doxxing, Potential Doxxing, or Non-Doxxing). 2. BPDL_lexicon_7_core.csv Contains 220 Bengali privacy-related lexical patterns organized into 7 categories. The file includes three attributes: category, pattern_type, and pattern. These patterns support structured lexical feature extraction for identifying privacy-related expressions in Bengali text. 3. BPDL_lexicon_16_core.csv Contains the same 220 lexical patterns organized into 16 fine-grained privacy-related categories. It follows the same three-column structure, enabling the investigation of detailed lexical characteristics and comparison of different category granularities in doxxing detection. 4. Tridox_BN_Split.csv Provides predefined training, validation, and testing assignments for all 1,755 samples. The file contains two attributes: - sample_id: Unique identifier linking each record to the main dataset. - split: Assigned partition (train, validation, or test). The dataset is divided into 1,225 training samples, 264 validation samples, and 266 testing samples. These partitions were established using a group-aware splitting strategy to reduce the risk of related samples appearing across different partitions and to facilitate reproducible experimental evaluation. Research Applications TriDox-BN is intended to facilitate research on Bengali doxxing detection, contextual privacy-risk classification, lexicon-enhanced machine learning, Transformer-based text classification, and explainable artificial intelligence. The accompanying privacy lexicons and predefined data partitions provide a structured resource for developing, evaluating, and comparing privacy-aware Bengali NLP models.
Files
Steps to reproduce
1. Dataset Preparation: The Bengali text samples were prepared through Unicode normalization, whitespace standardization, and duplicate detection. Exact and near-duplicate samples were identified using character-level TF-IDF cosine similarity and reviewed before removal. The finalized dataset contains 1,755 labeled samples across three categories: Doxxing, Potential Doxxing, and Non-Doxxing. 2. Dataset Access: Download TriDox_cleaned_dataset.csv, which contains four columns: sample_id, context_text, target_text, and label. The context and target text can be used together or independently for text classification experiments. 3. Lexicon-Based Feature Extraction: Use BPDL_lexicon_7_core.csv and BPDL_lexicon_16_core.csv, containing 220 lexical patterns organized into 7 and 16 categories, respectively. Apply the patterns to Bengali text using the matching method specified in pattern_type to extract category-wise lexical features. 4. Data Partitioning: Load Tridox_BN_Split.csv and match its sample_id column with the main dataset. Use the predefined group-aware partitions containing 1,225 training, 264 validation, and 266 testing samples. Maintain these assignments to ensure consistent experimental comparisons. 5. Model Development: Train Bengali text classification models using the training partition. Researchers can compare target-only, context-aware, and lexicon-enhanced approaches. Use the validation partition for model selection and hyperparameter tuning. 6. Performance Evaluation: Evaluate the selected models on the predefined test partition using macro-F1 score, accuracy, macro-recall, and confusion matrices. Compare results across different input configurations and lexical feature representations. 7. Reproducibility: Preserve the provided sample identifiers, class labels, lexical patterns, and split assignments throughout experimentation to ensure consistent and comparable results. Note: The four released CSV files support reproduction of downstream classification experiments. Exact reconstruction of the original data collection and annotation process requires access to the original source records and annotation materials, which are not included in this release.
Institutions
- Chittagong University of Engineering & TechnologyChittagong, Chittagong