Code-mixed Chaos : Multi-labeled Banglish & Bangla Corpus for Toxicity analysis

The dataset addresses a crucial gap in toxicity detection for Banglish—a code-mixed form of Bengali and English written in Roman script—which is often undervalued in NLP research. To mitigate this, we present a manually collected, multi-labeled dataset comprising 10,234 Banglish social media comments, annotated across 10 classes with toxic and non-toxic categories. The toxic comments are categorized into nine types: (1) Vulgar-based, (2) Religious-Hostility, (3) Troll-based, (4) Insult-based, (5) Loathe-based, (6) Threat-based, (7) Race-based, (8) Sexual-based, and (9) Political-Chaos. And a single Non-toxic category representing comments that do not have any form of toxicity. It is equally divided between toxic (5,117) and non-toxic (5,117) entries. Each sample was sourced from platforms such as Facebook, YouTube, Instagram, and X (formerly Twitter). To balance the dataset, it is enriched by selectively adding non-toxic texts from a publicly available corpus: "Bengali & Banglish: A Monolingual Dataset for Emotion Detection in Linguistically Diverse Contexts". Additionally, we provided a Bangla-translated version of the dataset to support the script-based comparative analysis in toxicity detection.

本数据集针对孟加拉英语混合语（Banglish，一种以罗马字母书写的孟加拉语与英语代码混合形式）的毒性检测领域长期存在的关键空白——此类语言在自然语言处理（Natural Language Processing，NLP）研究中常被忽视。为弥补这一不足，我们构建了人工采集标注的多标签数据集，包含10234条孟加拉英语混合语社交媒体评论，共设置10个分类标签，涵盖有毒与无毒两大范畴。其中有毒评论细分为9个类型：(1) 低俗言论类、(2) 宗教敌意类、(3) 引战类、(4) 侮辱类、(5) 憎恶类、(6) 威胁类、(7) 种族歧视类、(8) 性相关类、(9) 政治乱象类；另有1个无毒类别，用于标注无任何毒性的评论。该数据集在有毒与无毒样本间实现了均衡分布，两类样本各5117条。所有样本均采集自Facebook、YouTube、Instagram及X（原Twitter）等社交平台。为平衡数据集分布，我们从公开语料库"Bengali & Banglish: A Monolingual Dataset for Emotion Detection in Linguistically Diverse Contexts"中选择性添加无毒文本以扩充数据集。此外，我们还提供了该数据集的孟加拉语译本，以支持基于书写脚本的毒性检测对比分析。