Enhanced Hate Speech Detection using Balanced Datasets and AI Techniques

Abstract

Detecting hate speech is an essential factor in preventing toxicity from ruining communication online while letting them operate in safer digital spheres. For this purpose, here we scale up hybrid CNN-RNN architectures through augmentation and carefully engineered dataset balancing. While previous works had a constant struggle against imbalanced datasets, our work constructs balanced datasets using synthetic oversampling such as SMOTE, under sampling, and modern augmentation schemes for a fairer and more robust distribution for training. We also employ other state-of-the-art Machine Learning and Deep Learning systems, such as pre-trained language models. Explainable AI techniques are to ensure transparency and offer insights into the flagged content for interpretation and trust. This research attains improved accuracy (up to 0.908) and F1 scores (up to 0.914) with computational efficiency capable of real-time deployment. Ethical concerns such as bias mitigation and opportunities for adaptation within torture communities keep strengthening the framework. This project builds into large-scale, multi-linguous, big-impact hate-speech detection systems within the interest of society at large