Selection Collider Bias in Large Language Models

TL;DR

This paper investigates how selection collider bias affects Large Language Models, causing them to learn spurious dependencies, and proposes a method to leverage this bias for better uncertainty estimation in gender pronoun tasks.

Contribution

It explains the causal mechanisms of selection collider bias in LLMs and introduces an uncertainty metric that aligns with human judgment in gender pronoun tasks.

Findings

Uncertainty metric matches human uncertainty in gender pronoun tasks.

Selection collider bias can be exploited to assess model uncertainty.

Method demonstrated on extended Winogender Schemas.

Abstract

In this paper we motivate the causal mechanisms behind sample selection induced collider bias (selection collider bias) that can cause Large Language Models (LLMs) to learn unconditional dependence between entities that are unconditionally independent in the real world. We show that selection collider bias can become amplified in underspecified learning tasks, and although difficult to overcome, we describe a method to exploit the resulting spurious correlations for determination of when a model may be uncertain about its prediction. We demonstrate an uncertainty metric that matches human uncertainty in tasks with gender pronoun underspecification on an extended version of the Winogender Schemas evaluation set, and we provide an online demo where users can apply our uncertainty metric to their own texts and models.