Classification of abusive Thai messages in social networks using deep learning

Abstract

Social media has improved on traditional news sources by allowing increased access to information. However, the anonymity social media provides can lead to abusive and hateful speech without detection or repercussion from individuals with malicious intentions. This research develops a binomial and a multinomial classification model for classifying Thai social media text for five categories of abusive content detection in social media that include Rude, Figurative, Dirty, Offensive and Non-Abusive. The experiments demonstrated that DistilBERT achieved the highest F1 score with 0.8510 for the binomial model and 0.9067 for the multinomial model. BiLSTM performed second best with an F1 score of 0.8403 and 0.8969 for the binomial and multinomial models, respectively. Both deep learning models outperformed the traditional machine learning classifiers’ highest F1 score of 0.7452 and 0.8090 for the binomial and multinomial models, respectively. The deep learning architectures allow for better contextual representations of the words with the DistilBERT, enabling better modeling of long-range dependencies between words.