BanTH: A Multi-Label Hate Speech Detection Dataset for Transliterated Bangla

BanTH is a dataset of 37.3 thousand transliterated Bangla text samples for hate speech detection and target identification. The data was sourced from 26 YouTube channels across News & Politics, People & Blogs, and Entertainment categories. It was created by aplycaebous and last updated on March 25, 2025.

Use Cases

• Train binary hate speech classifiers based on the described hate/non-hate labels.
• Develop multi-label target identification models based on the seven target classes (Political, Religious, Gender, etc.).
• Analyze hate speech patterns in transliterated Bangla content sourced from YouTube.
• Benchmark NLP models on a dataset with a specific class imbalance (~38% hate samples).

Strengths

• Contains 37.3 thousand annotated text samples, providing a substantial corpus.
• Includes multi-label annotations across seven distinct target classes for granular analysis.
• Sourced from a diverse set of 26 YouTube channels across three content categories.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect geographic, temporal, or source bias inherent to the specific YouTube channels sampled.

Provenance

Source

YouTube channels (26 channels across News & Politics, People & Blogs, and Entertainment).

Collection Method

Sourced from YouTube channel content; each sample labeled by three annotators.

Freshness

Last updated 2025-03-25 15:50:04.