Classifying Speech into Offensive and Hate Categories along with Targeted Communities using Machine Learning

In the modern-day age, the problem of hate speech and offensive speech has increased due to the internet being widely used and technical resources being available to most people. If not moderated, it could lead to severe riots and hate-mongering against minorities. Such behavior is difficult to moderate so an efficient approach with be to filter out hate and offensive speech. Therefore, using one vs rest classification this research study has introduced a system to classify a comment as being normal, hateful, or offensive, and the communities targeted by it; totally 18 labels are used, one for the classification of comment and the other 17 being the target communities. In addition to the global accuracy, this research study has also provided individual accuracy for each community being targeted in the One Vs Rest model. The proposed research study exhibits the following global test accuracies for the different models for comparative analysis - One vs Rest - 90.97%, Binary Relevance - 28.7%, Label Powerset - 39.1%, Multi KNN - 19.6%.