Early discovery of cyberbullying incidents in multiparty Chinese-English code-mixed colloquial dialogue: A generative AI approach

Various machine learning models have been employed to detect aggressive speech and cyberbullying incidents on social media platforms. Previous studies have not addressed the applicability of handling a lengthy social media session at a time due to the short content window and technical limitations imposed on early NLP models. In addition, some studies have focused on content classification without addressing the identification of related perpetrators and victims, which are important components for determining cyberbullying severity levels. This paper introduces a victim-perpetrator-category association modeling approach and investigates the practical use of recently developed long context length large language models to handle early cyberbullying incident detection tasks. Tailored prompt templates with proper stop phrases were designed to explicitly model the associations between victims and perpetrators, enabling a more accurate prediction of the occurrence of cyberbullying incidents. Two large pretrained models, Llama 3.1 8B and Qwen 2.5 14B, were finetuned using the low-rank adaptation fine-tuning technique with 1685 manually annotated multiparty Chinese-English Cantonese colloquial dialogue sessions, resulting in a total of 14,257 tweets. It is shown that the holistic use of dialogue session messages at a time provides significant performance advantages over handling messages individually in the victim identification task. The explicit victim-perpetrator-category association is empirically shown to improve early cyberbullying detection performance in terms of the ERDE and F-latency across all customizable severity levels gauged by different thresholds of insult frequency, number of perpetrators and power imbalance. • Explicit victim-perpetrator-category association for early bullying detection. • Identify cases by multi-thresholds: frequency, participants and power imbalance. • Demonstrate practical prompt templates for advanced generative LLMs. • Efficient fine-tuning architecture with rank stabilized low-rank adaptation method. • Annotate 1685 sessions of Chinese-English code-mixed colloquial language dialogues.