AI & Society
BERT achieved the best performance among five large language models, with 95% accuracy
A 2025 study found that BERT identified cyberbullying on social media with 95 percent accuracy, precision, recall, and F1 score. Researchers at University of Science and Technology Beijing, Blekinge Institute of Technology, and Stevens Institute of Technology compared BERT with RoBERTa, XLNet, DistilBERT, and GPT-2.0. The comparison, published in IEEE Transactions on Information Forensics and Security, found BERT achieved optimal performance with 95 percent accuracy, precision, recall, and F1 score, and a 5 percent error rate.
What did the study aim to do?
The paper assessed text classification methods for detecting cyberbullying in real time on social media. It framed cyberbullying as a serious mental health problem. The researchers tested large language models for how well they could identify harmful content.
How did BERT perform?
BERT achieved 95 percent accuracy, precision, recall, and F1 score. Its error rate was 5 percent. The paper also reports that BERT needed 0.053 seconds of inference time, used 35.28 MB of RAM, used 0.4 percent CPU/GPU utilization, and consumed 0.000263 kWh of energy.
What does the paper conclude?
The authors note that generative AI models can be powerful, but fine-tuned models often outperform them when adapted to a specific dataset and task. In this comparison, BERT achieved optimal performance among the models tested.
Limits and what's next
The study is one benchmark comparison. It tested text classification models for cyberbullying detection on social media, but the abstract does not name the specific platforms or dataset. The abstract reports no independent replication. It also notes that current detection systems face challenges in performance, dataset quality, time efficiency, and computational cost. The abstract does not say what future work the authors propose.
Source: Adamu Gaston Philipo et al., “Assessing Text Classification Methods for Cyberbullying Detection on Social Media Platforms,” IEEE Transactions on Information Forensics and Security, 2025.