BADD: A Large-Scale Dataset for Arrogance Detection in the Bengali Language
Description
This dataset contains 46,128 labeled Bengali comments curated for the task of arrogance detection. While existing datasets focus heavily on hate speech or cyberbullying, this dataset addresses the subtle linguistic nuances of "arrogance", characterized by overbearing pride, lack of empathy, and social superiority, which is often expressed without overt toxicity. The data was compiled to support research in Bengali NLP. It serves as the primary resource for training the high-performing BanglaBERT model (96% accuracy) described in the accompanying research paper. Dataset Structure The dataset is provided in a single .csv file with the following columns: comment: The raw Bengali text. source: The origin of the comment (online or AI). weak_label: Initial label assigned by heuristic functions. snorkel_label: Refined label produced by the Snorkel framework. final_label: The target label for classification. 1: Arrogant 0:Non-arrogant **Further an automaited English translated dataset is attached as test_translated_data.csv
Files
Steps to reproduce
Sources: The data includes 44,257 comments scraped from social media platforms (Facebook, YouTube) and online news portals, alongside 1,871 AI-generated samples (using Gemini 1.5 Pro and ChatGPT-4) to ensure diversity in arrogant patterns. Labeling Strategy: Due to the scale of the data, a Weak Supervision framework was employed. Initial labels were generated using keyword-based heuristics (focusing on self-referential pronouns and possessive words) and subsequently refined with the Snorkel labeling model to resolve conflicts and enhance label confidence. Preprocessing: The text is provided in its raw form. According to the research findings, stopwords were retained because they contain critical stylistic markers (e.g., 'আমি', 'আমার') essential for detecting an arrogant tone.