Comparative Study Between Traditional and Hybrid LSTM Models for Toxic Comment Classification

The widespread use of social media platforms such as X, Instagram, and Threads has led to a surge in user generated content, where freedom of expression and speech is often exercised. However, such openness results in an increased number of toxic comments that may cause personal hurt. Being able to identify and classify such comments provides the key for keeping online communities healthy. This paper compares and develops deep learning models on the classification of toxic comments for recognizing malicious comment. The approaches include Long Short-Term Memory (LSTM) and a hybrid approach that combines LSTM with CNN. For improving performance, we incorporate several preprocessing techniques into the model, while using Word2Vec embeddings for transfer learning. In this regard, our results of the implementation of this approach depict that the LSTM with CNN model provides the highest accuracy of 95.12%, outperforming the traditional approach. These findings indicate that LSTM with CNN based architecture, applied together with strong preprocessing and word embeddings, might lead to higher performance in toxic comment classification. The implications of this study are of use for social media platforms and developers willing to implement more effective moderation tools.