A Multi-Platform Cyberbullying and Toxic Span Detection Dataset for Advanced NLP
Description
This dataset features 75,000 carefully curated and fully balanced social media textual records designed for cyberbullying classification and sequence-labeling toxic span detection. Collected and mapped across four distinct digital ecosystems—Facebook (FB), Instagram (IG), Reddit, and online News Portals—the corpus spans seven operational categories: Racism/Xenophobia, Misogyny/Sexism, Homophobia/LGBTQphobia, Religious Hate, Harassment/Stalking, Profanity/Insult, and a baseline of Non-Toxic control samples. A standout feature of this dataset is the precise character-level localization tags (toxic_span_start and toxic_span_end), which mark the exact spans of offensive words within toxic strings. This structural architecture makes it ideal for training multi-task learning models, sequence tagging frameworks, and deep learning models (such as BERT, RoBERTa, and BiLSTM-CRF) aimed at explainable AI and targeted content moderation.
Files
Steps to reproduce
To replicate or reconstruct this cyberbullying and toxic span sequence dataset, follow these sequential engineering steps: 1. Data Sourcing and Platform Sampling: - Aggregated text corpuses across four major online ecosystems: Facebook (FB), Instagram (IG), Reddit forums, and mainstream digital News Portal comment sections. 2. Annotation Pipeline & Categorization: - Establish a taxonomy mapping textual communications into seven targeted behavioral categories: 'Racism_Xenophobia', 'Misogyny_Sexism', 'Homophobia_LGBTQphobia', 'Religious_Hate', 'Harassment_Stalking', 'Profanity_Insult', and 'Non_Toxic'. - Perform string analysis to ensure each category contains a uniform distribution of approximately 10,714 to 10,715 records, totaling 75,000 records. 3. Character-Level Toxic Span Labeling: - For all toxic-classified strings, programmatically evaluate and isolate the exact substring containing the active slurs, insults, or targeted hate words. - Map the precise python string indices for the starting character ('toxic_span_start') and the ending character ('toxic_span_end') of the offending token. - Assign index values of -1 for both 'toxic_span_start' and 'toxic_span_end' across all 10,714 baseline 'Non_Toxic' records to preserve sequence uniformity. 4. Structuring and Deduplication: - Filter and sanitize structural rows to guarantee text authenticity and alignment. - Ensure text sequences remain intact with original linguistic choices, hash-tags, and markers to best serve deep learning sequence labelers (e.g., BERT-CRF, RoBERTa token classifiers). 5. Final Integration: - Formulate the verified 6-column structure ('id', 'text', 'label', 'platform', 'toxic_span_start', 'toxic_span_end'). - Export the structured data frame to a comma-separated artifact titled "Cyberbullying_Toxic_Span_Research_Dataset.csv".