Hate speech detector based on hybridized BERT-attention mechanism and context analyzer

Abstract

Aim: The research aims to create a new hate-speech detection model by utilizing a hybridized method that captures complex contextual linkages within textual data. Hate speech remains a threat to the peaceful coexistence of humans in societies especially via open social networks in this current age, presenting grave obstacles to online safety, and promoting inclusive environments. Methods: This is achieved by combining the advantages of bidirectional encoder representations from transformers (BERTs) attention processes with a context analyzer. Careful data augmentation was carried out utilizing back translation, which is made possible by the deep-translator library, enhancing the dataset’s diversity and quantity to guarantee a comprehensive and reliable dataset. Results: The training of the frozen BERT layer out of the two layers of the model produced a total accuracy of 0.99 on the 20th epoch by identifying the multi-labeled classes of hate speech using the Adam optimizer and softmax. Promising performance is shown by the trained model’s assessment metrics, which include a macro precision of 0.79875, a macro recall of 0.71587, and a macro F1-score of 0.74825. Conclusion: By utilizing the hybridized BERT model, damaging information can be understood holistically as it can identify not only explicit hate speech but also subtle sensitivities and underlying meanings.