Fine-tuning XLM-R for the Detection of Sinhala Hate Speech Content on Twitter and Youtube

Hate speech disseminated through social media platforms triggered multiple intense incidents in Sri Lanka during the past few years. Regardless of the gravity of these incidents, existing policies and automated algorithms used by social media platforms failed to detect hate speech content in the Sinhala language. Further, only a few studies had been conducted to detect Sinhala hate speech due to the limited datasets available. In this paper, a new hate speech dataset consisting of 4168 annotated text data samples extracted from Twitter and Youtube platforms has been introduced. Further, five XLM-Rbase based fine-tuning strategies have been evaluated for Sinhala hate speech classification using the newly introduced Twitter-Youtube dataset and publicly available Kaggle dataset. XLMRbase model fine-tuned with bidirectional RNN layer has outperformed all the other fine-tuning strategies for both datasets. Further, XLM-Rbase-RNN models have achieved a 91.2 F1-score for the Kaggle dataset and an 88.3 F1-score TwitterYoutube dataset.