Paper Title: Enhanced toxic comment detection model through Deep Learning models using Word embeddings and transformer architectures
Authors: S.Sushma, Sasmita Kumari Nayak, M. Vamsi Krishna
Corresponding Author: Sushma S (sushma.cse2@gmail.com)/India
Abstract
The proliferation of harmful and toxic comments on social media platforms necessitates the development of robust methods for automatically detecting and classifying such content. This paper investigates the application of natural language processing (NLP) and ML techniques for toxic comment classification using the Jigsaw Toxic Comment Dataset. Several deep learning models, including recurrent neural networks (RNN, LSTM, and GRU), are evaluated in combination with feature extraction methods such as TF-IDF, Word2Vec, and BERT embeddings. The text data is pre-processed using both Word2Vec and TF-IDF techniques for feature extraction. Rather than implementing a combined ensemble output, the study conducts a comparative evaluation of model-embedding combinations to determine the most effective pairings. Results indicate that integrating BERT with traditional models (RNN+BERT, LSTM+BERT, GRU+BERT) leads to significant improvements in classification accuracy, precision, recall, and F1-score, demonstrating the effectiveness of BERT embeddings in capturing nuanced text features. Among all configurations, LSTM combined with Word2Vec and LSTM with BERT yielded the highest performance. This comparative approach highlights the potential of combining classical recurrent models with transformer-based embeddings as a promising direction for detecting toxic comments. The findings of this work provide valuable insights into leveraging deep learning techniques for toxic comment detection, suggesting future directions for refining such models in real-world applications.
Keywords
Toxic comment classification,Word embeddings, Ensemble modeling
