Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Synthetic vs. Gold: The Role of LLM Generated Labels and Data in Cyberbullying Detection

Created by
  • Haebom

Author

Arefeh Kazemi, Sri Balaaji Natarajan Kalaivendan, Joachim Wagner, Hamza Qadeer, Kanishk Verma, Brian Davis

Outline

This paper addresses the challenges of developing a cyberbullying (CB) detection system for online users, including children. Specifically, we propose a method for generating synthetic data and labels using a large-scale language model (LLM) to address the lack of labeled data reflecting children's language and communication styles. Experimental results show that a BERT-based CB classifier trained on synthetic data generated via LLM achieves comparable performance (75.8% accuracy vs. 81.5% accuracy) to a classifier trained on real data. Furthermore, LLM is also effective for labeling real-world data, with the BERT classifier achieving comparable performance (79.1% accuracy vs. 81.5% accuracy). This suggests that LLM can be a scalable, ethical, and cost-effective solution for generating cyberbullying detection data.

Takeaways, Limitations

•
Takeaways:
◦
We demonstrate that LLM can effectively address the data generation and labeling challenges for cyberbullying detection systems.
◦
We offer practical solutions to the challenge of obtaining cyberbullying data on children, which is hampered by ethical, legal, and technical constraints.
◦
Leveraging LLM-based synthetic data enables the construction of cost-effective and scalable cyberbullying detection systems.
•
Limitations:
◦
The performance of the model using synthetic data was slightly lower than that of the model using real data (75.8% vs. 81.5%). Further research is needed to reduce the performance gap.
◦
Further validation of the quality and diversity of the data generated by LLM is needed.
◦
There is a need to evaluate how accurately the data generated by the LLM reflect the language use patterns of actual children.
👍
Made with Slashpage