Abstract: As the online communication grows exponentially, the issue of toxic comments, varying from hate speech and cyberbullying to offensive and abusive content, has emerged as a pressing issue for social media sites, online forums, and news portals. Although a lot of headway has been achieved in the detection and moderation of toxic content in English, the task becomes more challenging in multilingual environments because of the varying linguistic frameworks, cultural environments, and the lack of sufficiently annotated datasets in most languages.
This survey article delves into the recent research in multilingual toxic comment classification, emphasizing datasets, methods, and challenges used in this process. We detail a comprehensive critique of different strategies, such as rule-based and lexicon-based methods, old-school machine learning models, and state-of-the-art deep models. Particular emphasis is on the performance of transformer-based architectures like multilingual BERT (mBERT) and XLM- RoBERTa, based on large-scale pretraining that facilitates cross-lingual competency. In addition, we address the use of cross-lingual transfer learning in overcoming low-resource language issues and the effect of code-switching and transliteration on toxicity detection.
There are still some challenges that exist despite progress, such as model and dataset biases, the absence of contextual awareness in some languages, and the dynamic nature of toxic language on the internet.
Keywords: Multilingual Toxic Comment Classification, Hate Speech Detection, Offensive Language Identification, Cross-Lingual Transfer Learning, Transformer Models, Natural Language Processing (NLP), Content Moderation.
|
DOI:
10.17148/IJARCCE.2025.14564
