Offensive Sentence Classification Using Character-Level CNN and Transfer Learning with Fake Sentences

Abstract

There are two difficulties in classifying offensive sentences: One is the modifiability of offensive terms, and the other is the class imbalance which appears in general offensive corpus. Solving these problems, we propose a method of pre-training fake sentences generated as character-level to convolution layers preventing under-fitting from data shortage, and dealing with the data imbalance. We insert the offensive words to half of the randomly generated sentences, and train the convolution neural networks (CNN) with theses sentences and the labels of whether offensive word is included. We use the trained filter of CNN for training new CNN given original data, resulting in the increase of the amount of training data. We get higher F1-score with the proposed method than that without pre-training in three dataset of insult from kaggle, Bullying trace, and formspring.