High-Performance Hate Speech Detection with Hybrid Attention
Vineet Kaundal, Naveen Chauhan · 2024
Online hate speech is a serious threat that endangers people and communities, impacting inclusivity. In this paper, we aim to design a novel deep learning system that performs more effectively in identifying hate speech. Our approach uti lizes hierarchical attention to capture cues for hate speech in sentences, multi-head attention to capture some pattern, and a BiLSTM to understand the context. We have also applied specific preprocessing techniques to address the issue of typos and collocation of words into colloquial language commonly found in hate speech. Moreover, we fine-tuned the model's hyperparameters to perfection to improve performance, custom made for our dataset on hate speech. The accuracy was a huge 93% under test conditions for a specific type of dataset, almost matching performance with a huge discrepancy from the baseline of the usual BiLSTM. We have demonstrated that our approach, incorporating customized preprocessing and attention processes, provides a practical way of mitigating the damages from hate speech online. This research will further assist in making online platforms more friendly, inclusive, and safe.