Study of Deep Learning Techniques for Real-Time Online censorship using Comment Toxicity Detection

In an era where our online interactions are integral to our daily lives, addressing and mitigating online toxicity is crucial. This study, titled "Comment Toxicity Detection," aims to significantly advance ongoing efforts in this area. Our objective is to protect user welfare and encourage positive online interactions. This research addresses the urgent issue of online toxicity and its negative impact on user interactions and well-being. We explore various data sources and employ advanced feature engineering techniques such as BERT and its derivatives like m-BERT and MURIL BERT. By integrating these methods with various deep learning classifiers, we enhance our collective efforts to curb the spread of harmful content in digital spaces. We utilized an English dataset extracted from a larger multilingual corpus and preprocessed it using count vectorization for classification tasks. The evaluation metrics indicated a recall of 0.7103 and a precision of 0.7997. Our classification model used a Deep LSTM architecture with a batch size of 17. We are currently investigating the potential improvements in our data representation by incorporating a BERT-based model. This investigation is essential to optimize our classification performance, and the results should clarify the best practices for using advanced language representations to improve classification accuracy.