Multilingual Toxic Comment Classification using Deep Learning

Toxic comments can have a profoundly negative impact on online communities, fostering environments that feel unsafe, unwelcoming, and discouraging for participation. The task of toxic comment classification focuses on identifying and categorizing harmful or disruptive comments within online discussions. Despite extensive research in this area, several challenges persist, such as Multilingual Toxicity Detection, Data Imbalance, and Interpretability. Existing studies have categorized toxic comments into six distinct labels: toxic, severely toxic, obscene, threat, insult, and identity hate. Traditional machine learning models like SVM have achieved an accuracy of 87%, while more advanced models like BERT have reached up to 98.3%. The goal of this proposed work is to improve the identification of toxic comments across multiple social media platforms, particularly in datasets containing entries in multiple languages. To address this, deep learning models such as XLM-RoBERTa and BERT are employed to detect the language of the comment and classify it as either toxic or non-toxic. Each model’s performance will be evaluated primarily using Accuracy metrics, providing a basis for comparison to determine the most effective approach for multilingual toxic comment detection. High accuracy can be achieved through careful dataset preprocessing, advanced model architectures, and hyperparameter tuning. Performance can be improved by fine-tuning model hyperparameters and employing better feature extraction techniques. More clarity is required in explaining the rationale behind selecting particular models.