SIREN! Detecting Burmese Hate Speech Comments on Social Media

Abstract

Hate Speech on Social Media is definitely an evolving threat for every nation, especially for countries like Myanmar. Lack of media and digital literacy is playing a huge role in making people insult to each other or misallocating their stresses to others without physical encounter. Moreover, disingenuous politicians fuel online hate speech campaigns backstage of the elections by targeting different religions in the regard of heretics and using racialism. To emphasize this matter, we scraped over 16,000 social media comments from the most popular social media platform in Myanmar and performed hate-speech research using those samples. With the precise definition of a hate speech labelling guideline, annotation on the sample dataset was done systematically and efficiently. Experiments and evaluations were conducted using different linear and non-linear deep-learning classification models. Performances of the models are at the peak in Logistic Regression among linear models with 0.8974 AUC score and XLM-RoBERTa among deep learning models with 0.8958 AUC score on the test dataset. We observed that it is more advantageous to use linear models on our dataset since they achieved comparable results to the deep learning models and have much lower computational cost.