Aspect & Target-Aware Contextual Bengali Abusive Content Dataset

This dataset represents a context-aware Bengali abusive comment corpus, compiled from publicly available YouTube comments on Bangladeshi news channels such as JamunaTV, SomoyTV, ATN News and RTv. It consists of 1,500 manually collected and annotated comments, including both abusive and non-abusive instances, to support research in abusive language detection, sentiment analysis, target identification and computational social linguistics in low-resource languages.

Each record in the dataset contains the following fields:
ID, post_title, comment, parent, sequence, aspect, target, and abusive (Yes/No). The inclusion of parent and sequence fields enables contextual linkage between comments and their replies, supporting hierarchical analysis of online interactions.

Comments were categorized across 11 aspect classes (e.g., education, economy, sports political, social and so on) and 4 target types (e.g., individual, group, society, others). To clarify this schema, we provide two representative labeled comments: An abusive comment: “এরা ছাত্র না, এ রাজাকারের বাচ্চা।” → “They are not students; they are the children of the collaborators.” (labeled abusive(yes/no) = yes; aspect: politics; target: group). A non-abusive comment: “বাংলাদেশ দল আজ দারুণ খেলেছে, গর্বিত বোধ করছি।” → “The Bangladesh team played really well today; I feel proud.” (labeled abusive(yes/no) = no; aspect: sports; target: group). These paired examples demonstrate how we map Bengali comments to binary abuse labels while capturing aspect and target for richer analysis.