Multilingual Toxic Comment Classification Using Bidirectional LSTM

The growth of social networking sites and online platforms has brought about an unprecedented surge in user-generated content. However, along with the immense benefits of increased communication and information sharing, there has been an alarming growth in toxic and offensive comments. Detecting and moderating such comments is crucial to maintain a healthy and safe online environment. In this research, we propose a multilingual toxic comment classification system that leverages the power of Bidirectional Long Short-Term Memory (BiLSTM) neural networks. We use a comprehensive dataset which contains a diverse range of toxic comments in multiple languages. We employ a BiLSTM architecture because it is effective at detecting both contextual and sequential dependencies in text data. We train our model by combining word embeddings with character level embeddings in order to capture the semantic and morphological information found in the comments. Multiple cutting-edge methods are used to compare the model’s performance, including RNN and LSTM. The experimental findings show that the suggested model performs competitively in classifying multilingual toxic comments, surpassing other approaches with an accuracy of 94.21%.