Token and part-of-speech fusion for pretraining of transformers with application in automatic cyberbullying detection

Cyberbullying detection remains a significant challenge in the context of expanding internet and social media usage. This study proposes a novel pretraining methodology for transformer models, integrating Part-of-Speech (POS) information with a unique way of tokenization. The proposed model, based on the ELECTRA architecture, undergoes pretraining and fine-tuning and is referred to as ELECTRA_POS. By leveraging linguistic structures, this approach improves understanding of context and subtle meaning in the text. Through evaluation using the GLUE benchmark and a dedicated cyberbullying detection dataset, ELECTRA_POS consistently delivers enhanced performance compared to conventional transformer models. Key contributions include the introduction of POS-token fusion techniques and their application to improve cyberbullying detection, as well as insights into how linguistic features influence transformer-based models. The result highlights how integrating POS information into the transformer model improves the detection of harmful online behavior while benefiting other natural language processing (NLP) tasks.