Are Gender-Neutral Queries Really Gender-Neutral? Mitigating Gender Bias in Image Search

1 Introduction

• The paper defines a framework for gender-neutral English queries and diagnoses bias in specialized retrieval models and general-purpose pretrained representations.Gender-balanced results are expected for these queries.Source: §1 · p. 1
• FairSample addresses imbalanced training data, while feature clipping prunes representation features highly correlated with gender using mutual information.The clipping method requires no training and is compatible with various pretrained models.Source: §1 · pp. 1–2
• On MS-COCO and Flickr30K, FairSample and feature clipping significantly reduce gender bias in SCAN and CLIP, respectively.For CLIP, post-processing also mitigates gender-group similarity discrepancies in internet occupation searches.Source: §1 · p. 2

2 Gender Bias in Image Search

• For gender-neutral queries, approximately equal proportions of images depicting men and women are expected; gender-specific queries fall outside this study’s scope.Source: §2 · p. 2

2.1 Problem Statement

• Image search matches a text query to images, and evaluation assumes that test queries use gender-neutral language.Models predict image–text similarity scores, with cosine similarity given as an example.Source: §2.1 · pp. 2–3

2.2 Measuring Gender Bias

• Image gender is labeled male when only men appear, female when only women appear, and neutral otherwise; bias compares masculine and feminine result proportions.The metric preserves skew direction: positive per-query differences indicate a male skew, and the test-set measure averages differences across queries.Source: §2.2 · p. 3

3 Mitigating Gender Bias in Image Search

• The two approaches target different model families: fair sampling requires in-domain training, whereas feature clipping works with pre-trained models without repeating training.The clipping method is motivated by reducing dependence between representations and gender attributes.Source: §3 · p. 3, §3.2 · p. 4

3.1 In-processing Debiasing: Fair Sampling

• Fair sampling balances negative-image sampling between male and female images for gender-neutral queries, while retaining the original sampling strategy for gender-specific queries.The approach modifies contrastive learning for in-domain training.Source: §3.1 · p. 4
• The fair-sampling image-search loss weights masculine and feminine negative-image terms for gender-neutral pairs; the sentence-retrieval loss remains unchanged.The authors report that applying fair sampling thoroughly reduces recall too much, so they combine objectives with a weight α.Source: §3.1 · p. 4

3.2 Post-processing Debiasing: Feature Clipping based on Mutual Information

• Feature clipping greedily removes the m image-representation dimensions with highest mutual information with gender and applies the same index set to text representations.The remaining image and text representations are compared using cosine similarity for ranking.Source: §3.2 · pp. 4–5

4.1 Datasets

• MS-COCO uses 113,287 training, 5,000 validation, and 5,000 test images; results use five 1K test folds or the full 5K test set, while Flickr30K uses 1,000 validation and 1,000 test images.Flickr30K contains 31,000 images, with the remaining images used for training.Source: §4.1 · p. 5

4.2 Models

• SCAN and FairSample are trained and tested on in-domain datasets, while CLIP and CLIP-clip use the pre-trained model without fine-tuning on dataset training sets.CLIP-clip retains 400 dimensions from CLIP's original 500-dimensional image and text vectors, using m = 100.Source: §4.2 · p. 5

4.3 Evaluation

• Evaluation neutralizes gender-specific test captions by removing or replacing gender-specific words, and measures Bias@K and Recall@K among the top-K images.Bias@K measures gender bias; Recall@K is the fraction of queries whose correct image appears among the top K.Source: §4.3 · p. 6

5.1 Main Results on MS-COCO & Flickr30K

• FairSample and CLIP-clip consistently reduce test-set gender bias; CLIP-clip achieves the best Bias@10 on all three test sets, while FairSample retains recall comparable to SCAN.Examples include a 10.7% Bias@10 reduction for SCAN with FairSample on Flickr30K and reductions of 22.3%, 24.4%, and 17.7% for CLIP-clip on MS-COCO 1K, MS-COCO 5K, and Flickr30K, respectively.Source: §5.1 · p. 6, §5.2 · p. 7, Table 2

5.3 Tradeoff between Recall and Bias

• On Flickr30K, α = 1 reduces Bias@10 by 22.5% but Recall@10 by 10.9%; the final FairSample model uses α = 0.4 for better recall retention.Across the FairSample curve, higher recall often comes with higher gender bias.Source: §5.3 · p. 7, Figure 3 · p. 7
• On MS-COCO 1K, clipping dimensions from m = 100 to 400 can drop Recall@1 from 46.1% to 11.3%, Recall@5 from 75.2% to 25.4%, and Recall@10 from 86.0% to 34.2%; m = 100 is selected.The paper also reports high standard deviation in gender bias when too many covariates are clipped.Source: §5.3 · p. 7, Figure 4 · p. 7

5.4 Evaluation on Internet Image Search

• The MS-COCO and Flickr30K evaluations are limited by gender labels extracted from human captions; the occupation evaluation instead uses crowd-annotated gender attributes.The occupation dataset contains the top 100 Google Image Search results for each gender-neutral occupation term.Source: §5.4 · pp. 7–8
• On 100 crowd-labeled occupation searches, CLIP-clip has 42.3% lower mean absolute similarity bias than CLIP (0.0064 vs. 0.0111), though doctor and police-officer discrepancies rise but remain below 0.01.Figure 5 visualizes absolute similarity bias for 18 occupations; the broader mean comparison is over 100 occupations.Source: §5.4 · p. 8, Figure 5 · p. 8

Fairness in Machine Learning

• Fair-machine-learning approaches include preprocessing, in-processing, and post-processing; the paper positions FairSample and clip in the latter two regimes.Preprocessing methods reweight or repair training data, in-processing methods modify objectives, and post-processing applies corrections to trained models.Source: §6 · p. 9

Social Bias in Multi-modality

• Prior multimodal-bias research covers image captioning, visual question answering, face recognition, and representation learning.This work complements that literature by examining gender bias in gender-neutral image-search language.Source: §6 · p. 9

Gender Bias in Online Search Systems

• Online-search studies document gender inequality and its effects on perceptions, but the cited works do not attempt to mitigate bias in search algorithms.This work extends that line toward understanding bias in search algorithms and mitigating it in two image-search model families.Source: §6 · p. 9

7 Conclusion

• FairSample targets training a fairer image-search model from scratch, while clip offers lightweight deployment for pre-trained representation models when gender information is accessible.The conclusion frames both approaches as ways to mitigate gender bias for gender-neutral image-search queries.Source: §7 · p. 9

• Unequal male–female inclusiveness in gender-neutral image-search results can shape and exaggerate gender stereotypes.The authors connect search systems’ influence on access to information with this potential societal impact.Source: Broader Impact · pp. 9–10
• The authors propose their mitigation methods as potential contributions to responsible image search and encourage study of other biases, including racial bias.They also suggest exploring whether their methodology generalizes to other bias measures.Source: Broader Impact · p. 10
• Gender-bias measurement and mitigation depend on gender labels inferred from human-annotated captions, whose coverage and language inclusiveness affect accuracy.The authors warn that missing gendered words or preprocessing errors may introduce unforeseen bias, and that English word lists may not transfer across languages or cultures.Source: Broader Impact · p. 10

• Table 3 presents the gender word lists used for identifying caption attributes.Source: A Gender Word Lists · p. 13, Table 3 · p. 13
• Caption gender attributes are identified from the occurrence of gender-specific words.Source: Table 3 · p. 13
• The word-list procedure uses words appearing in sentences as cues for caption gender attributes.Source: Table 3 · p. 13

B.1 Computing Infrastructure

• Training and evaluation used a GPU server with four NVIDIA RTX 2080 Ti GPUs.Source: §B.1 · p. 13

B.2 Computational Time Costs

• SCAN and SCAN with fair sampling took about 20 hours to train for 30 epochs on MS-COCO, while testing on the 1K test set took 8–10 minutes.Pre-trained CLIP and CLIP-clip evaluation on the MS-COCO 1K test set took under one minute.Source: §B.2 · p. 13

• For “a person riding a bike,” FairSample placed images depicting females at a higher rank, while all models returned relevant images.Source: C Qualitative Examples · p. 13
• Figure 6 compares the top five retrieved images for SCAN, SCAN+FairSample, CLIP, and CLIP-clip.Source: Figure 6 · p. 14