Verbal Abuse Classification Using Multiple Deep Neural Networks

People can be exposed to verbal abuse practically anywhere. It is considered to be one of serious issues in society. In this paper, we describe a method to classify verbal abuse into five lasses by adding a convolutional neural network (CNN), a long short-term memory, and a dense layer on top of bidirectional encoder representations from transformers (BERT). The data are collected from Korean drama, movies, and YouTube. Due to data imbalance, weighted random sampler and data augmentation are used to train the models to be generalized. Experiments show that BERT with CNN after data augmentation performs the highest accuracy among all the compared methods.