Abusive language detection using customised BERT. (16th February 2022)

• Record Type: Journal Article
• Authors: Chandu, Burre; Rohitha, Kaza Phani; Nihal, Nampally; Bindu, K. Hima
• Is Part Of: International journal of swarm intelligence. Volume 7:Number 1(2022)
• Page Start: 25
• Page End: 38
• Publication Date: 2022-02-16
• Subjects: abusive language detection -- offensive language detection -- hate speech -- language models -- transformers -- attention -- BERT -- natural language processing -- NLP -- fine-tuning -- attention -- LSTM; Swarm intelligence -- Periodicals; 006.3
• Languages: English
• ISSNs: 2049-4041

Abstract

Freedom of expression and speech is widely misused in social media today. Unfortunately, the content shared on these platforms contains abusive content which is in multiple forms. This humongous amount of text requires automated detection of abusive language. This is a challenging problem because of a lot of noise, large vocabulary, context-dependency, and multilingualism. Deep learning (DL) models are being used due to the inefficiency of regular expressions, blacklists, and machine learning approaches. Hence, we have used neural language models for extracting high-quality features from the text. This paper demonstrates the usage of the natural language processing (NLP) model, bidirectional encoder representations from transformers commonly called BERT to perform various classification techniques. We have presented the results of fine-tuning the BERT pre-trained model for abusive language detection. The empirical analysis of offensive language identification and toxic comment classification datasets showed that these architectures achieved better results than existing models.