Deep Learning-Based Systems for Detecting Hate Speech and Offensive Language in Texts
Tanjim Mahmud, Md. Faisal Bin Abdul Aziz, Avishek Majumder, Sha Md Farid, Tahmina Akter
Abstract
The critical work of recognizing hate speech and objectionable language on social media is essential to ensuring the safety of the online community. The usefulness of deep learning systems in recognizing such content is investigated in this work. Using regular expression replacement, lowercase conversion, stopword removal, stemming, and tokenization, we pre-processed a dataset of 24,783 tweets that were categorized as hate speech, offensive language, or neither. For training and testing, we divide the data using word embedding techniques. Numerous DL models, including CNN-LSTM, CNN-BiLSTM, LSTM, and BiLSTM fusion models, were used in the study. The BiLSTM model obtains a maximum accuracy of 90.54% in comparison to other models. A comparison examination with multiple data splits validates the robustness and practicality of our method. Our study’s findings show that BiLSTM models can reliably distinguish between hate speech and provocative language. This is advantageous for social media networks that use automatic content control.