A Domain-Adaptive Soft Prompting Framework for Multi-Type Bias Detection in News

Chengjun Zhang, Benjamin M. Ampel, Sagar Samtani
Last updated on Jun 9, 2026
PDF DOI
In Plain Terms

This paper introduces a method to automatically detect six different kinds of bias in news articles using large language models. Rather than relying on large hand-labeled datasets, the approach first adapts the model to news-style language and then uses “soft prompting” that needs only a small amount of labeled data. Tested on over 400,000 New York Times articles, it detects bias more accurately than standard prompting methods.

Artifacts

PDF DOI

Related Papers

Cite this

APA Zhang, C., Ampel, B. M., & Samtani, S. (2026). A Domain-Adaptive Soft Prompting Framework for Multi-Type Bias Detection in News. In Proceedings of the 59th Hawaii International Conference on System Sciences (HICSS). https://doi.org/10.24251/HICSS.2026.218
BibTeX
@inproceedings{zhang2026a,
  title={A Domain-Adaptive Soft Prompting Framework for Multi-Type Bias Detection in News},
  author={Chengjun Zhang and Benjamin M. Ampel and Sagar Samtani},
  booktitle={Proceedings of the 59th Hawaii International Conference on System Sciences (HICSS)},
  year={2026},
  doi={10.24251/HICSS.2026.218}
}
Bias Detection Media Bias Large Language Models Soft Prompting Domain-Adaptive Pretraining Natural Language Processing
Benjamin M. Ampel

Benjamin M. Ampel

Assistant Professor in Computer Information Systems and Director, Center for CyberAI Research (CCAIR)

My research focuses on AI-enabled Cybersecurity, including Cyber Threat Intelligence, Large Language Models, and Phishing Detection.