Wikipedia toxic comment classification via deep neural networks

With the technological developments and global Internet usage, online information sharing has increased considerably. Especially on platforms such as online forums and social media, people can share information with each other and comment freely on topics. Thus, it has become much easier to access information. However, insulting comments may be shared on these platforms apart from information sharing. As a negative result of this condition, it seems that some of who make useful shares, begin avoiding online sharing. The data used within the research has been collected from Wikipedia toxic comment data which was shared for a competition on Kaggle website. Insulting, hateful, offending or harassing comments have been classified under 6 titles. These titles are "toxic", "severe toxic", "obscene", "insult", "threat", "identity hate". Within the concept of this study, Keras library was used for word embeddings, Adam optimizer algorithm was used for optimizing weights in the models; LSTM, Bi-LSTM and GRU algorithms were used for classifying toxic comments. Performance measurements were carried out by accuracy and F1 score metrics. Each method was processed 30 times and when compared statistically on collected accuracy and F1 score metrics, it has been observed that there is no difference between the methods.