Toward Deconfounding the Influence of Entity Demographics for Question Answering Accuracy

TL;DR

This paper investigates how demographic biases in QA datasets affect model accuracy, highlighting the need for more diverse datasets to uncover potential biases related to gender, nationality, and profession.

Contribution

It introduces a framework to deconfound demographic influences in QA datasets, emphasizing the importance of balanced representation for fairer model evaluation.

Findings

Model accuracy shows little bias based on gender or nationality.

Profession-related question topics cause more accuracy variation.

Current datasets lack sufficient demographic diversity.

Abstract

The goal of question answering (QA) is to answer any question. However, major QA datasets have skewed distributions over gender, profession, and nationality. Despite that skew, model accuracy analysis reveals little evidence that accuracy is lower for people based on gender or nationality; instead, there is more variation on professions (question topic). But QA's lack of representation could itself hide evidence of bias, necessitating QA datasets that better represent global diversity.