Multi-Label Toxic Comment Classification Using Machine Learning: An In-Depth Study

Sammanfattning: The classification of toxic comments is a well-researched area with many techniques available. However, effectively managing multi-label categorization still requires a considerable amount of work. In this thesis, we performed a classification experiment on over 200 thousand comments from the Jigsaw toxic comment competition data available on Kaggle. We aimed to optimize a model to identify six different categories of hate speech. Initially, we implemented a baseline model using a simple vectorization technique and logistic regression. Subsequently, we compared this model with more advanced approaches that employed elaborate vectorization techniques in conjunction with recurrent neural networks and transformers. After thorough analysis, we found that a fine-tuned transformer-based model called RoBERTa yielded the best performance, achieving a mean macro average F1-score of 0.808. This model surpassed the previous state-of-the-art set by van Aken et al. (2018), which achieved an F1 score of 0.791. Finally, we integrated the optimized model in a web application to visualize the toxicity of messages.