An integrated text-based hybrid RNN-CNN model for toxic comment classification.

Abstract

Forum and social media especially Facebook, Instagram and Twitter play a vital role in the communication and become the pivotal role in expressing feelings and opinions toward certain posts and news through leaving a comment and review. The comment and review can be positive, neutral, or negative. There have been several longitudinal studies involving online comments showed that one of five comments, 20 percent of comments are toxic comments. Toxic comments are those negative comments that can greatly affect a person emotionally which might cause people to suffer from mental problems. Hence, toxic comments detection is fast becoming a key instrument in protecting social media platforms' users from cyberbullying. This paper seeks to remedy this issue by proposing and comparing different artificial intelligence models to classify the toxic online comments. Thus, six different models with various algorithms which are CNN, RNN, Hybrid-NN, Linear SVC, multinomial naive bayes and logistic regression will be trained with a pre-processed dataset and evaluated with cross validation and ROC AUC. These models are then compared by observing the accuracy of each model and result shows that Hybrid-Neural network, combination between RNN and CNN score the top of the chart with accuracy of 0.9778 which suggest the best model to classify the online comments.