POLITIX-BN: A dataset of Bengali political discourse annotated for sentiment and toxicity
Description
POLITIX-BN is a Bengali-language corpus of 5,786 political texts from Bangladesh, annotated on two separate dimensions: the sentiment (tone) of a text and its toxicity/discourse type (harmfulness). This lets users tell a harsh but civil criticism apart from an abusive attack. The texts were collected from ten source categories: YouTube, Facebook, Twitter, Instagram, TV interviews, talk shows, newspapers, Reddit, online portals and community forums. They were gathered using election and politics related keywords between March 13 and June 22 2026. Only the comment text was collected; usernames and other personal identifiers were not included. Every Bengali text comes with an English representation, produced with Google Translate and manually checked against the original. Each record is one row in an Excel file (POLITIX-BN.xlsx) and contains: • ID: record identifier • political_post_bn: Bengali text • political_post_en: English representation • source: source category • Gender, Occupation: metadata fields • sentiment_score / sentiment_label: 1 = Hostile, 2 = Negative, 3 = Neutral, 4 = Positive, 5 = Enthusiastic • toxicity_score / toxicity_label: 1 = Predatory/Threat, 2 = Hate Speech, 3 = Degrading/Insult, 4 = Antagonistic/Troll, 5 = Objective/Matter-of-fact, 6 = Constructive/Civil, 7 = Affirmative/Supportive • content_theme: dominant political theme (such as election and voting, party and leadership, price rise and economy, law and order, religion and identity, foreign influence, national projects and progress) Labels were assigned manually by native Bengali speakers following a written guideline, with a worst-first decision order for toxicity. Reliability was checked in two stages with an independent second annotator: a 500-instance pilot, then a validation of 150 unseen instances. Cohen’s κ was 0.321 for sentiment and 0.409 for toxicity/discourse. The label distribution is imbalanced. Neutral makes up 57.2% of sentiment labels. Objective/Matter-of-fact (51.5%) and Constructive/Civil (28.2%) dominate toxicity/discourse, while Predatory/Threat (12) and Hate Speech (11) are very rare. Baseline results from 17 machine-learning, deep-learning and transformer models on original and preprocessed Bengali and English text are reported in the accompanying article. The dataset can support research on Bengali sentiment analysis, toxicity and hate speech detection, content moderation, political communication and low-resource NLP. It contains abusive, threatening and hateful language, included only for developing and evaluating moderation tools. This language does not reflect the views of the authors. The labels describe text and should not be used to profile individual speakers.
Files
Institutions
- American International University-BangladeshDhaka Division, Dhaka