🤖 AI Summary
This study addresses the long-standing limitation in Urdu toxic text detection, which has predominantly focused on sentence-level classification while neglecting fine-grained identification of toxic spans. To overcome challenges posed by scarce annotated data, code-mixing, and rich morphological variation, the authors present the first token-level annotated dataset for toxic span detection in Urdu. They propose a sequence labeling framework that integrates XLM-RoBERTa with a Conditional Random Field (CRF) layer to enable context-aware, fine-grained toxicity detection across multiple domains. Evaluated on social media posts, news articles, and YouTube comments, the approach achieves a token-level F1 score of 60%, establishing the first supervised baseline for Urdu toxic content analysis. The method demonstrates both strong performance and interpretability, offering a significant advancement in multilingual online safety research.
📝 Abstract
Urdu toxic span detection remains limited because most existing systems rely on sentence-level classification and fail to identify the specific toxic spans within those text. It is further exacerbated by the multiple factors i.e. lack of token-level annotated resources, linguistic complexity of Urdu, frequent code-switching, informal expressions, and rich morphological variations. In this research, we propose MUTEX: a multilingual transformer combined with conditional random fields (CRF) for Urdu toxic span detection framework that uses manually annotated token-level toxic span dataset to improve performance and interpretability. MUTEX uses XLM RoBERTa with CRF layer to perform sequence labeling and is tested on multi-domain data extracted from social media, online news, and YouTube reviews using token-level F1 to evaluate fine-grained span detection. The results indicate that MUTEX achieves 60% token-level F1 score that is the first supervised baseline for Urdu toxic span detection. Further examination reveals that transformer-based models are more effective at implicitly capturing the contextual toxicity and are able to address the issues of code-switching and morphological variation than other models.