A Robust Bias Mitigation Procedure Based on the Stereotype Content Model
This work addresses bias mitigation in language models for AI fairness, presenting a prototype procedure that is incremental by building on existing methods to reduce stereotypes without prior knowledge of specific biases.
The paper tackled bias in language models by adapting the Stereotype Content Model to contextualized word embeddings and evaluating a fine-tuning procedure to reduce stereotypes. They found that SCM terms better captured bias than demographic-agnostic terms, and the procedure reduced stereotypes without harming downstream performance, requiring minimal resources.
The Stereotype Content model (SCM) states that we tend to perceive minority groups as cold, incompetent or both. In this paper we adapt existing work to demonstrate that the Stereotype Content model holds for contextualised word embeddings, then use these results to evaluate a fine-tuning process designed to drive a language model away from stereotyped portrayals of minority groups. We find the SCM terms are better able to capture bias than demographic agnostic terms related to pleasantness. Further, we were able to reduce the presence of stereotypes in the model through a simple fine-tuning procedure that required minimal human and computer resources, without harming downstream performance. We present this work as a prototype of a debiasing procedure that aims to remove the need for a priori knowledge of the specifics of bias in the model.