Towards Understanding Gender Bias in Relation Extraction

1 Introduction

• Because NRE outputs extend knowledge bases used by question-answering systems, gender-skewed extraction may leave those knowledge bases with less information about women and affect downstream predictions.The authors note that such bias may also contribute to allocative harms and reinforce gender-stereotypical beliefs.Source: §1 · p. 1
• The paper presents an evaluation framework for gender bias in NRE and examines whether mitigation techniques can reduce bias while maintaining performance.The evaluation covers models trained on relation-extraction data and existing mitigation techniques.Source: §1 · pp. 1–2
• Experiments found gender-related effects for spouse and hypernym predictions, while birthDate and birthPlace served as relations not intuitively related to gender.The authors report that results confirm this conjecture.Source: §1 · p. 2
• WikiGenderBias is a distantly supervised dataset of 45,000 Wikipedia sentences covering spouse, hypernym, birthDate, and birthPlace relations.The dataset was created because existing NRE datasets lack a reliable way to identify the gender of entities in input sentences.Source: §1 · p. 2
• The paper's stated contributions include creating WikiGenderBias, demonstrating gender bias in NRE outputs, and testing existing mitigation approaches.Source: §1 · p. 2

2 Related Work

• Prior NLP studies measured gender bias through male–female performance differences and fairness metrics such as Equality of Opportunity; this paper uses both kinds of metrics for NRE.Source: §2 · p. 2
• Bias mitigation has been applied during training, in embeddings, or at inference; this paper tests data augmentation and Hard Debiasing for NRE.The cited approaches include data augmentation and word-embedding debiasing.Source: §2 · pp. 2–3
• Relation extraction research includes supervised, semi-supervised, and unsupervised methods, with distant supervision used when annotation is insufficient or difficult to obtain.Prior NRE work also sought to reduce noise introduced by distant supervision.Source: §2 · p. 3
• The authors report that, to their knowledge, no prior studies had examined bias or ethics in NRE.Source: §2 · p. 3

3 WikiGenderBias

• Gender bias in NRE is defined as a difference in model performance on sentences from male versus female articles.The dataset was created because existing NRE datasets lack reliable entity gender information.Source: §3 · p. 3

3.1 Dataset Creation

• WikiGenderBias uses Wikipedia and DBPedia with distant supervision, treating a sentence mentioning the other entity in an article as expressing their known relation.It covers spouse, hypernym, birthDate, and birthPlace, plus NA examples from parents, deathDate, and almaMater, yielding a five-class task.Source: §3.1 · pp. 3–4
• Head entities are kept separate across train, development, and test sets; male instances are down-sampled in training and development to address gender imbalance.This split prevents reuse of sentences from a head entity’s article, though head entities may appear as tails elsewhere.Source: §3.1 · p. 4

3.2 Test Sets

• The test set balances male and female instances and uses three-worker relation annotations to check noisy distant-supervision labels; Fleiss’ κ is 0.44, indicating moderate agreement.A majority “no” response labels a sentence as NA.Source: §3.2 · p. 4

3.3 Data Analysis

• In sampled articles, spouse mentions are relatively more frequent for female entities, while hypernym mentions are relatively more frequent for male entities.These are proportions relative to the other target relations, not raw counts.Source: Figure 2 · p. 4, §3.3 · p. 5

4 Gender Bias in NRE

• The experiments use OpenNRE with a PCNN and Selective Attention, training encoder-selector combinations on WikiGenderBias and its gender-equalized version with WikiGenderBias-trained Word2Vec embeddings.Source: §4 · p. 5

4.1 Performance Parity Score

• Performance Parity Score (PPS) combines macro-average F1 with Disparity Score, subtracting the latter per relation with equal weighting to balance performance and inter-group parity.Dividing by the number of relations keeps PPS within (−1, 1], though the authors say it will generally fall between [0, 1].Source: §4.1 · pp. 5–6

4.2 Measuring Performance Differences

• Gender differences are measured by comparing performance on male and female datapoints, with Macro-F1 as the performance metric and F1Gap used to calculate Disparity Score.Source: §4.2 · p. 6

4.3 Equality of Opportunity Evaluation

• For each relation, Equality of Opportunity compares recall for male and female instances with that relation as the true label; here, this is equivalent to per-relation recall by gender.Source: §4.3 · p. 6

4.4 Result

• Across OpenNRE configurations, encoder and selector changes had no conclusive effect on bias, while average selectors improved aggregate performance; spouse predictions favored male instances, with smaller gaps on other relations.Female instances made up a higher share of spouse cases in WikiGenderBias, and the authors leave writing-style differences as a question for future work.Source: §4.4 · pp. 6–7, Table 3 · p. 7
• The observed gaps do not necessarily indicate model-internal bias because Wikipedia articles about males and females may be written differently; the authors nevertheless caution that Wikipedia-trained NRE can propagate latent training-data biases into knowledge bases.Source: §4.4 · p. 7

5 Bias Mitigation

• The mitigation study evaluates data augmentation and Hard-Debiasing as approaches to reducing gender bias in NRE.Source: §5 · p. 7

5.1 Bias Mitigation Techniques

• Equalizing the training distribution involves downsampling male instances until male and female counts are approximately equal.Source: §5.1 · p. 7
• Data augmentation swaps masculine and feminine words in sentences, then trains on the union of the original and augmented corpora.The authors suggest that augmentation may introduce artifacts and biases because NRE relies heavily on surface-level cues.Source: §5.1 · p. 7, §5.2 · p. 8
• Hard-Debiasing is applied to Word2Vec embeddings trained on WikiGenderBias sentences, with embeddings retrained on both corpora when paired with augmentation.Source: §5.1 · p. 7

5.2 Effectiveness of Bias Mitigation

• Downsampling male training instances brings male–female performance differences close to zero for every relation except hypernym, with a relatively small aggregate performance drop.Training on the equalized data is also more efficient because it uses fewer instances.Source: §5.2 · p. 8, Figure 4 · p. 8
• Most tested debiasing techniques cause significant performance drops, and none reduces the gender performance gap; embedding debiasing slightly improves aggregate F1.Combining mitigation approaches is also ineffective, lowering Macro F1, raising Disparity Score, or both.Source: §5.2 · p. 8

• WikiGenderBias is released as the first dataset aimed at evaluating bias in NRE models.Source: §6 · p. 8
• The study finds a difference in spouse-relation F1 scores between predictions on male and female sentences, while naive data augmentation causes a significant performance drop.Source: §6 · p. 8
• Building unbiased neural relation extractors remains an open and difficult research question.The authors suggest that some noise-adding mitigation methods may encourage spurious correlations and unwanted biases.Source: §6 · p. 8
• The study is preliminary, considers only binary gender, and calls for future work on non-binary genders and closer examination of prediction sources.Source: §6 · p. 8