Enhancing Toxic Comment Classification with Multi-Label Capabilities: Leveraging BERT and RoBERTa Models

In recent years, the proliferation of multi-labeled sentiment text classification techniques, driven by machine learning, has been remarkable. This paper delves into a comprehensive comparative analysis of two cutting-edge models, BERT and RoBERTa, with a primary emphasis on evaluating their respective accuracies. The study aims to provide a conclusive assessment based on the research findings. The accuracy achieved by both BERT and RoBERTa, as discovered through rigorous evaluation, consistently falls within the range of 85% to 95%. The dataset is a combination of diverse categories, some of which are toxicity, insults, and obscenities, serving as the basis for classifying comments fetched from the Twitter source. The dataset consists of total of 34,000 tweets, ensuring the diversity and quantity required for comprehensive analysis of comments. The findings indicate distinctive characteristics of each model. RoBERTa emerges as the preferred choice for datasets of high volume, especially when backed by robust computational resources, yielding exceptionally robust results. Also, BERT seems to give better results with smaller datasets, consistently delivering optimal and highly accurate outcomes. These results obtained can provide insights and necessary recommendations for researchers and practitioners, offering a deeper understanding of when to leverage BERT or RoBERTa.