A Note on Bias to Complete

TL;DR

This paper reframes bias as a context-dependent and relative phenomenon, proposing eight bias classes and eight corresponding hypotheses to model how dynamic environments shape bias in LLMs. It outlines a novel framework that embraces bias through tensor-based representations, contrastive context, and multilingual data generation, paired with five methods aimed at definition, quantification, data augmentation, mitigation, and mis/disinformation analysis. The key contributions include a formalized eight-hypothesis paradigm and a practical strategy for incorporating diverse biases to improve debiasing outcomes, along with a plan to validate these ideas using embedding-space geometry and deep learning. The work aims to deliver tangible paths toward reducing biased interpretations and misinformation propagation in AI-assisted communication, with potential impact on policy, education, and cross-cultural discourse.

Abstract

Minimizing social bias strengthens societal bonds, promoting shared understanding and better decision-making. We revisit the definition of bias by discovering new bias types (e.g., societal status) in dynamic environments and describe them relative to context, such as culture, region, time, and personal background. Our framework includes eight hypotheses about bias and a minimizing bias strategy for each assumption as well as five methods as proposed solutions in LLM. The realization of the framework is yet to be completed.