An Innovative Novel Transformer Model and Datasets for Safeguarding Religious Sensitivities in Online Social Platforms

Abstract

Nowadays, blasphemy, inter-religious hatred, and intolerance toward other people's faith become increasing problems in different social media. To effectively tackle these problems, it is necessary to develop efficient techniques for automated detection. With the increasing concern about hate speech on the internet, this research study focuses on creating and testing advanced models to identify and categorize texts related to religious hatred. Leveraging two meticulously curated datasets, the first comprising binary classification tasks to discern hate speech directed towards religion and the second featuring a nuanced five-class categorization including Blasphemy, Ethnic Hatred, Ethnic Group Hatred, Undetermined, and Non-Hatred. The main focus of the research is examining the effectiveness of different established machine learning techniques, deep learning methods, and pre-trained transformer models. Notably, the study introduces hatebnBERT, a novel BERT-based model tailored to the Bangla language's specific nuances and linguistic intricacies specifically built for hate speech detection. Through rigorous evaluation and comparative analysis, the results demonstrate the superior accuracy and precision of hatebnBERT, achieving an accuracy of 98.8% in the binary-class dataset and 98.6% in the multiclass dataset, in effectively identifying and classifying various forms of hate speech within the realm of religious discourse. The outcomes underscore the significance of leveraging specialized language models, such as hatebnBERT, to combat the proliferation of abusive speech and promote a more courteous and inclusive online social media environment within the Bangla-speaking community, considering the potentially devastating consequences of unchecked religious hatred propagation. Insert here your abstract text.