CROSS-PLATFORM HATE SPEECH DETECTION BY TARGET CATEGORY: EVALUATING TRADITIONAL AND TRANSFORMER MODELS ENHANCED WITH SMOTE

Abstract

Societal and technological challenges are significant when it comes to cyberbullying, which is an ubiquitous issue in the social media space including Twitter, Facebook, YouTube, and Instagram. This paper will focus on the identification of hate speech that targets particular individuals, especially in the context of the data in Hindi language. It tries to fill the gap between the overall abusive language and the hatred directed at specific people or populations. To this end, an annotated and carefully edited compilation was created where hate speech was divided into the following categories: racial/ethnic, religious, sexual orientation, and political. In order to address the problem of the imbalance of the classes, both Synthetic Minority Over-sampling Technique (SMOTE) and Grouped SMOTE were applied to the model to enhance its efficiency. The traditional machine learning frameworks (Support Vector Machines (SVM), Convolutional Neural Networks (CNN), Long Short-Term Memory networks (LSTM), and the transformer-based model Bidirectional Encoder Representations from Transformers (BERT) have been explored and the results show that BERT outperforms the traditional models (the F1-score of BERT is 89.2), which supports the initial hypothesis. The second hypothesis was also supported by the use of SMOTE which increased accuracy and precision. It was found that there was a strong correlation between the frequency of hate speech and the different demographic attributes particularly with regard to racial and political biases which support the third hypothesis. Also, the document underlines the fact that goal-based classification is a better approach to binary classification models, thus validating the fourth hypothesis. Further analysis reveals a difference in hate speech trends on different platforms: on Twitter, politically charged hate is quite widespread, and on Facebook, hate speech is mostly based on religious topics. The outlined findings emphasise the importance of creating detection systems of hate speech that are platform-adaptable, demographically aware, and that are also built upon sentiment analysis. This study contributes to the field of context-sensitive content control and enhances fairness of hate speech detection through natural languages processing tools and is of great interests to the researcher of artificial intelligence, decision-makers, and social networks.