Cyberbullying victimization identification and large language model-assisted assessment: a study of cyberbullying victimization lexicon construction and validation
Xingyun Liu, Yuehan Liao, Fan Feng, Yiming Tu, Xin Kang, Miao Liu, Nuo Han · Frontiers in Psychiatry · 2026
Cyberbullying, as a global psychological health threat, faces challenges in accurate identification due to the biases of traditional self-report methods and obstacles to help-seeking. This study constructs a Chinese cyberbullying victim dictionary based on social media big data, psychological feature dictionaries, and cyberbullying questionnaires to enhance the accuracy of victim identification, providing the possibility for timely intervention. The dictionary's validity was confirmed through a correlation analysis between word frequency statistics of 500 Weibo posts and expert ratings (n=3). To assess whether DeepSeek-R1 and GPT-4o could aid in the dictionary compilation process, the study replicated the manual creation process and compared their performance to expert evaluations using Kappa consistency coefficients, intraclass correlation coefficients (ICC), recall rates, and precision rates under both simple and complex instructions.Results showed that: (1) The dictionary, consisting of 442 words across three dimensions—cyberbullying methods, harm perception, and coping strategies—demonstrated high validity in identifying cyberbullying victimization expressions (r = 0.870, p < 0.001) and assessing victimization severity (r = 0.533, p < 0.001); (2) In text classification, DeepSeek-R1 performed well on small samples (Kappa = 0.775–0.781), but its performance dropped with a large dataset of 12,600 posts. Both models showed discrepancies in vocabulary selection and weight assignment (lowest Kappa = -1); however, DeepSeek-R1 showed moderate consistency in dimensional categorization (Kappa = 0.659–0.660) and outperformed GPT-4o. This study is the first to develop a Chinese cyberbullying victim dictionary, highlighting AI's potential in dictionary construction and suggesting human-AI collaboration for large-scale data tasks.