Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mitigating Bias for Question Answering Models by Tracking Bias Influence

Mingyu Derek Ma, Jiun-Yu Kao, Arpit Gupta, Yu-Hsiang Lin, Wenbo Zhao, Tagyoung Chung, Wei Wang, Kai-Wei Chang, Nanyun Peng

TL;DR

This work proposes BMBI, an approach to mitigate the bias of multiple-choice QA models by measuring the bias level of a query instance by observing its influence on another instance and introducing a new bias evaluation metric to quantify bias in a comprehensive and sensitive way.

Abstract

Models of various NLP tasks have been shown to exhibit stereotypes, and the bias in the question answering (QA) models is especially harmful as the output answers might be directly consumed by the end users. There have been datasets to evaluate bias in QA models, while bias mitigation technique for the QA models is still under-explored. In this work, we propose BMBI, an approach to mitigate the bias of multiple-choice QA models. Based on the intuition that a model would lean to be more biased if it learns from a biased example, we measure the bias level of a query instance by observing its influence on another instance. If the influenced instance is more biased, we derive that the query instance is biased. We then use the bias level detected as an optimization objective to form a multi-task learning setting in addition to the original QA task. We further introduce a new bias evaluation metric to quantify bias in a comprehensive and sensitive way. We show that our method could be applied to multiple QA formulations across multiple bias categories. It can significantly reduce the bias level in all 9 bias categories in the BBQ dataset while maintaining comparable QA accuracy.

Mitigating Bias for Question Answering Models by Tracking Bias Influence

TL;DR

This work proposes BMBI, an approach to mitigate the bias of multiple-choice QA models by measuring the bias level of a query instance by observing its influence on another instance and introducing a new bias evaluation metric to quantify bias in a comprehensive and sensitive way.

Abstract

Models of various NLP tasks have been shown to exhibit stereotypes, and the bias in the question answering (QA) models is especially harmful as the output answers might be directly consumed by the end users. There have been datasets to evaluate bias in QA models, while bias mitigation technique for the QA models is still under-explored. In this work, we propose BMBI, an approach to mitigate the bias of multiple-choice QA models. Based on the intuition that a model would lean to be more biased if it learns from a biased example, we measure the bias level of a query instance by observing its influence on another instance. If the influenced instance is more biased, we derive that the query instance is biased. We then use the bias level detected as an optimization objective to form a multi-task learning setting in addition to the original QA task. We further introduce a new bias evaluation metric to quantify bias in a comprehensive and sensitive way. We show that our method could be applied to multiple QA formulations across multiple bias categories. It can significantly reduce the bias level in all 9 bias categories in the BBQ dataset while maintaining comparable QA accuracy.
Paper Structure (76 sections, 5 equations, 1 figure, 8 tables)

This paper contains 76 sections, 5 equations, 1 figure, 8 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Bias detection and mitigation for NLP tasks.
  4. Social biases in question answering.
  5. Preliminaries
  6. Bias Definition, Category and Axis
  7. Bias definition.
  8. Bias category and axis.
  9. Task Definition and Base Models for QA
  10. The QA task.
  11. Base models.
  12. Task Formulation and Reference Instances for Bias Mitigation
  13. The Bmbi Bias Mitigation Method
  14. Bias Detection by In-Context Bias Influence Tracing
  15. Reference instances and bias detection axis.
  16. ...and 61 more sections

Figures (1)

  • Figure 1: Model design of Bmbi. The example illustrates the bias mitigation process of a query instance in terms of the Gender identity bias category. The output space of the ruler instance defines the bias axis. Since the common societal bias about emotional closedness is negative towards males (represented by the answer candidate "Kenneth"), a positive bias level indicates the query instance contains a negative bias towards the protected group of males. The output of the QA task and bias detection module will be used to calculate losses respectively.