PaperlayerView exact revision on arXiv

Source-linked AI summary

First Women, Second Sex: Gender Bias in Wikipedia

Eduardo Graells-Garrido, Mounia Lalmas, Filippo Menczer

arXiv:1502.02341v2cs.SI

TL;DR

Women are underrepresented among Wikipedia editors, and whether gender bias also shapes their characterization in biographies remains a question. This paper compares biography metadata, language, and network structure. It finds differences across all three dimensions, attributable both to the offline world and to bias in Wikipedia content.

  • Problem

    Women make up approximately 16% of Wikipedia editors, and the paper asks whether gender bias also appears in how men and women are characterized in biographies.

    Source: §1 · p. 1

  • Method

    The paper compares men’s and women’s biographies across metadata, language, and network structure.

    Source: §1 · pp. 1–2

  • Results

    The study finds differences in biography metadata, language, and network structure, attributable both to offline conditions and to bias endogenous to Wikipedia content.

    Source: §7.2 · pp. 8–9

  • Takeaways & Limitations

    The authors suggest Wikipedia may consider revising its guidelines to address women’s non-findability and encourage less biased language.

    Source: §7.2 · p. 9

    The study focuses on English Wikipedia, which is biased toward Western cultures.

    Source: §7.2 · p. 9

Abstract

from arXiv · show

Contributing to history has never been as easy as it is today. Anyone with access to the Web is able to play a part on Wikipedia, an open and free encyclopedia. Wikipedia, available in many languages, is one of the most visited websites in the world and arguably one of the primary sources of knowledge on the Web. However, not everyone is contributing to Wikipedia from a diversity point of view; several groups are severely underrepresented. One of those groups is women, who make up approximately 16% of the current contributor community, meaning that most of the content is written by men. In addition, although there are specific guidelines of verifiability, notability, and neutral point of view that must be adhered by Wikipedia content, these guidelines are supervised and enforced by men. In this paper, we propose that gender bias is not about participation and representation only, but also about characterization of women. We approach the analysis of gender bias by defining a methodology for comparing the characterizations of men and women in biographies in three aspects: meta-data, language, and network structure. Our results show that, indeed, there are differences in characterization and structure. Some of these differences are reflected from the off-line world documented by Wikipedia, but other differences can be attributed to gender bias in Wikipedia content. We contextualize these differences in feminist theory and discuss their implications for Wikipedia policy.

1. INTRODUCTION

Although Wikipedia’s policies are intended to support unbiased content, women are underrepresented among contributors and may also be characterized differently in its biographies. This study compares those characterizations across metadata, language, and network structure, finding differences that reflect both the offline world and bias endogenous to Wikipedia content.

1. INTRODUCTION

  • Women make up only 16% of Wikipedia editors, prompting questions about whether gender bias appears in biographies as well as participation.The study asks how such bias can be identified and quantified, and how it can be explained through social theory.Source: §1 · p. 1
  • The study compares men’s and women’s biographies across metadata, language, and network structure to assess how they are characterized.It focuses on English Wikipedia and uses Western feminist theories to contextualize its analysis.Source: §1 · pp. 1–2
  • Women’s biographies contain more marriage-related and sex-related content, while men’s biographies emphasize cognition and receive greater network centrality.The authors attribute differences both to offline conditions represented in Wikipedia and to bias in Wikipedia content.Source: §1 · p. 2

2. BACKGROUND

Prior work has studied Wikipedia’s contributor gender gap and cultural bias in content, but large-scale content-bias measurement had not focused on gender. This paper builds on that work to compare biography language and metadata and assess gendered linking patterns.

2. BACKGROUND

  • Research on Wikipedia contributors has documented a gender gap, differences in communication and revision, and women’s contributions being reverted more often.Women have also reportedly contributed less because of aggressive behavior toward them.Source: §2 · p. 2
  • Large-scale studies of Wikipedia content bias had measured culture rather than gender, while prior biography studies found gender differences in marriage and divorce coverage.Other work found Wikipedia’s coverage of notable women remained low and biased toward men, and that women’s biographies more often included marriage or divorce events.Source: §2 · p. 2
  • The paper motivates studying biography differences because quality assessments vary with the portrayed person’s gender, and readers may value different topics than contributors.The authors’ hypothesis is that biographies are written differently, inspired by prior work on language-based characterization of women.Source: §2 · p. 2
  • The analysis compares gendered word-category frequencies using LIWC and measures network bias by comparing biography PageRank with null models.It also uses burstiness to estimate how important words are within individual biographies.Source: §2 · p. 2

3. DATASET

The study combines structured Wikipedia metadata, an English Wikipedia dump, and inferred gender labels. DBpedia supplies normalized article metadata from infoboxes, while the biography text is analyzed in overview and full-text forms.

3. DATASET

  • The study uses DBpedia 2014, the October 2014 English Wikipedia dump, and inferred biography gender from Bamman and Smith.The inferred-gender dataset classifies biographies as male or female.Source: §3 · pp. 2–3
  • DBpedia derives structured person metadata from Wikipedia infoboxes, assigning biographies to the Person class and, where applicable, a subclass such as Artist.It normalizes article URIs, links, and, when possible, attribute keys and values; the Simone de Beauvoir infobox is one example.Source: §3 · pp. 2–3, Figure 1 · p. 3
  • The analysis uses biography overviews for full-vocabulary analysis and full texts for LIWC-word analysis, removing template markup in both cases.Gender labels are inferred from grammatically gendered words; the method’s reported test on 500 biographies had 100% precision and 97.6% recall, and the study considers only male and female labels.Source: §3 · p. 3

4. META-DATA PROPERTIES

The metadata analysis finds that women are a small share of biographies with gender data, with representation varying by birth year and infobox attributes. Several attribute differences align with the gender composition of occupational groups or with biography conventions.

Presence and Proportion According to Class

  • 15.5% of the 893,380 biographies with gender metadata are about women; mean article lengths are 5,955 characters for men and 6,013 for women, while mean out-degrees are 42.1 and 39.4, respectively.Both length and out-degree differences were statistically significant, with small effect sizes; the authors focus on global differences in the Person class.Source: §4 · p. 3

Distribution According to Date of Birth

  • Since 1943, the yearly fraction of women’s biographies has consistently exceeded the global 0.155 share, and the representation trend became positive in 1750–1800.Of biographies with a recorded birth year, 53% are from 1943–2000; the birth-year analysis covers 1600–2000.Source: §4 · pp. 3–4, Figure 2 · p. 3, Figure 3 · p. 4

Infobox Attributes

  • Only 3.53% of 340 infobox attributes differ significantly by gender, and all identified attributes have large effect sizes (Cohen’s w > 0.5).Sports-related attributes are more frequent in men’s biographies, while birthName, occupation, title, and spouse are more frequent in women’s biographies.Source: §4 · p. 4, Table 2 · p. 4
  • Men’s greater frequency of sports-related attributes is linked to their prominence in sports classes, while sport biography templates omit occupation fields because the templates indicate occupations.The spouse attribute is more common in women’s biographies; Person accounts for 45% of biographies containing that attribute.Source: §4 · p. 4

5. LANGUAGE PROPERTIES

Men’s and women’s biographies use broadly similar vocabularies, but differ in gender-associated words and in a small number of semantic categories. The reported category differences favor men in cognitive-process frequency and burstiness, and women in sexual-process frequency.

Overview of Results

  • Word frequencies were similar across genders (rank correlation ρ = 0.65, p < 0.001); both distributions followed Zipf patterns with α ≈ 1.8.The paper’s overview describes the section’s overall finding as more similarities than differences.Source: §5 · p. 4, Gender Differences in Semantic… · p. 5

Associativity of Words with Gender

  • PMI-associated words linked women chiefly to arts, gender, and family roles, while men’s were especially sports-related, particularly football.The paper relates women’s family-role terms to the spouse attribute findings in its meta-data analysis.Source: Associativity of Words with Gender · p. 5, Figure 5 · p. 5

Gender Differences in Semantic Categories of Words

  • The analysis compares overview word frequencies and full-text burstiness across cleaned LIWC categories; repeated occurrences approximate a word’s importance within a biography.Words were counted once per biography for relative frequencies, and the cleaned category dictionary contained 2,877 words.Source: §5 · p. 4, Gender Differences in Semantic… · p. 5

Differences in Frequency

Differences in Burstiness

  • 60% (cogmech burstiness ES), 71% (cause), and 64% (work) favored men in the three significant full-biography burstiness categories.The categories are cognitive processes, causal processes, and work concerns, respectively.Source: Gender Differences in Semantic… · p. 5, Table 4 · p. 6

6. NETWORK PROPERTIES

Women are systematically underrepresented among highly ranked biographies in the observed network, a pattern not reproduced by the tested null models. The observed network also has gendered linking patterns and a higher Self-Focus Ratio than expected.

Empirical Network and Null Models

Gender, Link Proportions and Self-Focus Ratio

  • The observed Self-Focus Ratio was 6.55 versus an expected 5.41; women’s links to women exceeded expectation (χ² = 40.54, p < 0.001).The null models had Self-Focus Ratios similar to the expected value.Source: Empirical Network and Null Models · p. 7

Biography Importance

  • The study ranks biographies by PageRank and compares the women’s share among top-ranked articles with null-model distributions.The analysis considers the top r biographies for r from 10 to 700,706.Source: Biography Importance · p. 7, Figure 6 · p. 7
  • Women reached the expected 15% share among top biographies only when the full dataset was considered; observed results differed from every null model (p < 0.001).The null models stabilized around the expected share by r ≤ 10^4.Source: Biography Importance · p. 7, Figure 6 · p. 7

7. DISCUSSION

The discussion distinguishes offline reflections from gender bias endogenous to Wikipedia: metadata differences often reflect profile classes and recency, while language and network patterns point to gendered characterization. It connects these findings to Wikipedia policy and notes the study’s English-language and binary-gender scope.

7. DISCUSSION

  • Most metadata differences reflect profile fields and recency; article-length and out-degree effects are small and class-dependent, while spouse frequency may reflect gender roles or arts-related marriage patterns.The generic person class was the most common, complicating the alternative explanation that arts biographies are more likely than sports biographies to feature a notable spouse.Source: §7 · p. 7
  • Men are associated more with sports, while women are associated more with arts, gender, and family; “her husband” and “first woman” may reinforce stereotypes about women.The discussion notes that “first woman” can signal notability but may also foreground gender over achievement, as the Finkbeiner test cautions.Source: §7 · pp. 7–8
  • Men’s biographies more often foreground cognitive processes, while women’s more often foreground sexuality; cognitive-process and work-concern language is also burstier in men’s biographies.“Burstier” means those aspects are more important than others at the individual level.Source: §7 · p. 8
  • Men’s articles are disproportionately central by PageRank, while women’s greater tendency to link to other women may reflect offline networks rather than Wikipedia bias.The authors caution that some highly central women do not negate the broader pattern of male centrality.Source: §7 · p. 8

7.1 Implications

  • The authors argue that objectifying language and reduced findability warrant editor tools and an explicit gender-bias addition to neutral-point-of-view guidance.They note Wikipedia’s educational use as one reason characterization matters to readers.Source: §7.1 · p. 8
  • The authors suggest affirmative action for women, especially those from before 1943, to improve inclusion and discoverability, while cautioning that relaxed notability rules risk original research.Source: §7.1 · p. 8

7.2 Summary, Limitations and Future Work

  • Significant differences in metadata, language, and network structure reflect both the offline world and gender bias endogenous to Wikipedia content.The authors contextualize these differences in feminist theory and discuss their implications for Wikipedia policy.Source: §7.2 · pp. 8–9
  • The study is limited to English Wikipedia and a binary view of gender; future work could develop bias-detection tools, examine contributor differences, and test finer-grained categories and article maturity.The authors also propose examining whether bias varies by region, religion, or article quality.Source: §7.2 · p. 9
Loading 1502.02341v2…