Transformer-Based Hate Speech Detection in Online Content

Abstract

Detection of hate speech is now an important area of study in light of the intensive expansion of social media and the proliferation of offensive and damaging content online. Hate speech detection is a difficult task since it often encompasses subtle, context-based, and slang language and evolving terms. Classical machine learning methods including Decision Trees and ensemble classifiers have been observed with encouraging results through features such as TF-IDF, Bag of Words, and tweet length. The models tend to perform poorly when it comes to grasping the underlying context and semantics of words. Modern breakthroughs in deep learning, particularly the application of models such as CNNs, Bi-Directional LSTMs with attention, and Transformer-based models such as BERT and XLM-R, have greatly enhanced the performance of hate speech detection systems. These models use contextual embeddings to capture more effectively the subtleties of hate speech. Experimental results on numerous datasets indicate that Transformer-based models are more accurate, F1-score, and robust compared to conventional methods. Even with these improvements, issues like false positives, dataset bias, and the requirement of real-time detection persist. This research demonstrates the efficiency of deep learning in detecting hate speech and underscores the need for continued study in order to create fair, reliable, and flexible detection systems.