Unmasking Harmful Comments: An Approach to Text Toxicity Classification Using Machine Learning in Native Language

In the modern digital era, social media plays a central role in our everyday lives, bringing both advantages and obstacles. While it facilitates global communication and trade, it also brings significant negative impacts, such as the spread of toxic comments that can lead to severe consequences, including mental distress and, in extreme cases, suicide. This study centers on classifying toxic comments by utilizing three different machine learning models: the Stochastic Gradient Descent (SGD) Classifier, the baseline Support Vector Machine (SVM), and an Optimized SVM model. A dataset comprising text comments in both Bangla and English, collected from various social media platforms, was used to train these models. The initial SVM model, while powerful, showed limitations, particularly in handling class imbalance, leading to suboptimal performance. The SGD Classifier provided a balanced approach but fell short in recall for non-harmful comments. To address these challenges, an Optimized SVM model was developed by fine-tuning key hyperparameters, significantly enhancing the model's performance. The optimized model achieved superior accuracy, precision, recall, and F1-scores across both harmful and non-harmful comment classifications. Our results demonstrate that the Optimized SVM model is the most reliable and effective among the three, making it the preferred choice for deployment in real-world scenarios where accurate classification of toxic comments is critical.