MEDICINE: Towards Multiple Dimensional Bias Mitigation via Causal Inference

ACL ARR 2026 January Submission6077 Authors

Keywords: Bias Mitigation, Causal Inference, Social Biases

Abstract: Language models significantly enhance the capabilities of natural language processing systems, yet they often inadvertently encode harmful biases that undermine societal fairness. To address this issue, we propose a causal inference framework to simultaneously mitigate multi-dimensional biases through a unified debiasing process. Our causal effect estimation framework enables systematic separation of genuine semantic influences from bias-induced spurious correlations during language model inference. Extensive experimental results demonstrate three key advantages of our approach: (1) it addresses multiple-dimensional biases in a unified framework without antagonistic effects, (2) the debiasing algorithm maintains task performance without negative impacts, (3) it requires no additional external corpus and operates with high efficiency under low resource demands.

Paper Type: Long

Research Area: Ethics, Bias, and Fairness

Research Area Keywords: model bias/unfairness mitigation, model bias/fairness evaluation

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 6077