Bangla-ToCo: A context-aware dataset for Bengali toxic comment detection

In today's online world, abusive and toxic comments on social media have become a growing concern, particularly for languages like Bengali, where there are currently few reliable automatic moderation systems. In this study, we introduce a context-aware dataset for identifying toxic comments in Bengali by considering how a comment's meaning changes when we read it alongside the relevant information. We present a dataset of 1004 Bengali news comments obtained from the official Facebook pages of two widely read Bangladeshi news portals such as ProthomAlo and News24. Each data record includes six attributes: the news title, metadata of the article, the target comment, its predecessor comment, its successor comment, and the commenter's username. Unlike existing Bengali corpora that present isolated text, this dataset preserves conversational context, enabling more accurate interpretation of ambiguous or short comments. The dataset has been annotated into two categories: Toxic and Non-Toxic. Annotation was performed independently by three human annotators, and the final label for each comment was assigned through majority voting to ensure consistency and reduce individual bias. This balanced class distribution makes the dataset particularly suitable for supervised learning and benchmarking. To the best of our knowledge, this is the first context-aware Bengali toxic comment dataset. Although modest in size, it provides one of the first systematically annotated resources for Bengali, a low-resource language in NLP. It has potential applications in abusive content detection, sentiment and stance classification, conversational modelling, and social discourse analysis, making it a valuable contribution for both computational linguistics and social sciences.