Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Bias in News Summarization: Measures, Pitfalls and Corpora

Julius Steen, Katja Markert

TL;DR

The paper addresses bias in text summarization by defining three bias types—inclusion, hallucination, and representation—and proposing a controlled input mutation framework that disentangles input-driven from summarizer-driven effects. It systematically evaluates gender and race biases across-purpose built and chat summarizers using OntoNotes-derived inputs with locally/global gender balancing, and novel alignment and gender-detection methods for hallucinations. The findings show minimal bias in content selection but clear gender-related hallucinations, with race biases largely similar; the authors validate their measures through quality checks and induced-bias tests, and extend the approach to racial bias to demonstrate generality. The work provides practical, rigorous tools for bias measurement in summarization and highlights faithfulness as a key mitigation target, while acknowledging dataset and task limitations and outlining paths for broader bias coverage.

Abstract

Summarization is an important application of large language models (LLMs). Most previous evaluation of summarization models has focused on their content selection, faithfulness, grammaticality and coherence. However, it is well known that LLMs can reproduce and reinforce harmful social biases. This raises the question: Do biases affect model outputs in a constrained setting like summarization? To help answer this question, we first motivate and introduce a number of definitions for biased behaviours in summarization models, along with practical operationalizations. Since we find that biases inherent to input documents can confound bias analysis in summaries, we propose a method to generate input documents with carefully controlled demographic attributes. This allows us to study summarizer behavior in a controlled setting, while still working with realistic input documents. We measure gender bias in English summaries generated by both purpose-built summarization models and general purpose chat models as a case study. We find content selection in single document summarization to be largely unaffected by gender bias, while hallucinations exhibit evidence of bias. To demonstrate the generality of our approach, we additionally investigate racial bias, including intersectional settings.

Bias in News Summarization: Measures, Pitfalls and Corpora

TL;DR

The paper addresses bias in text summarization by defining three bias types—inclusion, hallucination, and representation—and proposing a controlled input mutation framework that disentangles input-driven from summarizer-driven effects. It systematically evaluates gender and race biases across-purpose built and chat summarizers using OntoNotes-derived inputs with locally/global gender balancing, and novel alignment and gender-detection methods for hallucinations. The findings show minimal bias in content selection but clear gender-related hallucinations, with race biases largely similar; the authors validate their measures through quality checks and induced-bias tests, and extend the approach to racial bias to demonstrate generality. The work provides practical, rigorous tools for bias measurement in summarization and highlights faithfulness as a key mitigation target, while acknowledging dataset and task limitations and outlining paths for broader bias coverage.

Abstract

Summarization is an important application of large language models (LLMs). Most previous evaluation of summarization models has focused on their content selection, faithfulness, grammaticality and coherence. However, it is well known that LLMs can reproduce and reinforce harmful social biases. This raises the question: Do biases affect model outputs in a constrained setting like summarization? To help answer this question, we first motivate and introduce a number of definitions for biased behaviours in summarization models, along with practical operationalizations. Since we find that biases inherent to input documents can confound bias analysis in summaries, we propose a method to generate input documents with carefully controlled demographic attributes. This allows us to study summarizer behavior in a controlled setting, while still working with realistic input documents. We measure gender bias in English summaries generated by both purpose-built summarization models and general purpose chat models as a case study. We find content selection in single document summarization to be largely unaffected by gender bias, while hallucinations exhibit evidence of bias. To demonstrate the generality of our approach, we additionally investigate racial bias, including intersectional settings.
Paper Structure (39 sections, 4 equations, 1 figure, 20 tables)

This paper contains 39 sections, 4 equations, 1 figure, 20 tables.

Table of Contents

  1. Introduction
  2. Bias in Text Summarization
  3. Bias Measures
  4. Inclusion: Word Lists
  5. Inclusion: Entity Inclusion Bias
  6. Hallucination: Entity Hallucination Bias
  7. Representation: Distinguishability
  8. Input Documents are Already Biased
  9. Gender Bias Experiments
  10. Dataset
  11. Entity Alignment
  12. Identifying Gender of Hallucinated Entities
  13. Summarizers
  14. Gender Bias Results
  15. Validating our Measures
  16. ...and 24 more sections

Figures (1)

  • Figure 1: Schematic overview of our approach for summary gender bias evaluation with an example generated by BART XSum lewis_bart_2020. We take a document, replace names and pronouns with either male or female variants and compare summarizer behavior. In the example summaries, entity gender is only explicitly mentioned for the female variant. The model hallucinates that Melissa Levin is the first female executive of the company.