Exploring BERT and Bi-LSTM for Toxic Comment Classification: A Comparative Analysis

This study analyzes on the classification of toxic comments in online conversations using advanced natural language processing (NLP) techniques. Leveraging advanced natural language processing (NLP) techniques and classification models, including BERT and Bi-LSTM models to classify comments into 6 types of toxicity: toxic, obscene, threat, insult, severe toxic and identity hate. The study achieves competitive performance. Specifically, fine-tuning BERT using TensorFlow and Hugging Face Transformers resulted in an AUC ROC rate of $98.23 \%$, while LSTM yielded a binary accuracy of $96.07 \%$. The results demonstrate the effectiveness of using transformer-based models like BERT for toxicity classification in text data. The study discusses the methodology, model architectures, and evaluation metrics, highlighting the effectiveness of each approach in identifying and classifying toxic language. Additionally, the paper discusses the implementation of a userfriendly interface for real-time toxic comment detection, leveraging the trained models for efficient moderation of online content.