Political Bias in Neural Retrieval: Model-Agnostic Measurement on Canadian Social Media

Zeynep Pehlivan

Keywords: Political Bias, Neural Information Retrieval, Dense Retrieval, Bias Measurement

Abstract:

Representational bias inside embedding spaces is a structural property: geometric distortions
that persist across tasks and prompts. Because the same vector geometry underlies ranking, classification,
clustering, and generation, intrinsic bias can reappear unpredictably at deployment. We study
model-agnostic, no-retraining ways to measure and ultimately mitigate such bias in real systems.

Submission Number: 34