PaperlayerView exact revision on arXiv

Source-linked AI summary

Handling Bias in Toxic Speech Detection: A Survey

Tanmay Garg, Sarah Masud, Tharun Suresh, Tanmoy Chakraborty

arXiv:2202.00126v3cs.SIcs.CYcs.LG

TL;DR

Toxicity judgments depend on context, complicating detection. This survey organizes bias by its sources and targets of harm and synthesizes findings on bias in toxicity detection. Its takeaways identify sources of lexical bias and note that in-domain mitigation may not reduce disparity out of domain.

  • Problem

    The conditions for flagging content as toxic vary, complicating toxicity detection.

    Source: §1 · p. 2

  • Method

    The survey develops a taxonomy of bias organized by its sources and targets of harm.

    Source: §1 · p. 3, §7 · p. 19

  • Results

    Lexical bias may result from sampling bias, annotation bias, or both, while identity attacks can further promote disparity among social groups.

    Source: §5 · p. 16

  • Takeaways & Limitations

    The discussion treats language change, including word reclamation, as a future-research issue.

    Source: §6 · p. 18

    Mitigation evaluated on in-domain samples failed to reduce disparity on out-of-domain datasets.

    Source: §6 · p. 18

Abstract

from arXiv · show

Detecting online toxicity has always been a challenge due to its inherent subjectivity. Factors such as the context, geography, socio-political climate, and background of the producers and consumers of the posts play a crucial role in determining if the content can be flagged as toxic. Adoption of automated toxicity detection models in production can thus lead to a sidelining of the various groups they aim to help in the first place. It has piqued researchers' interest in examining unintended biases and their mitigation. Due to the nascent and multi-faceted nature of the work, complete literature is chaotic in its terminologies, techniques, and findings. In this paper, we put together a systematic study of the limitations and challenges of existing methods for mitigating bias in toxicity detection. We look closely at proposed methods for evaluating and mitigating bias in toxic speech detection. To examine the limitations of existing methods, we also conduct a case study to introduce the concept of bias shift due to knowledge-based bias mitigation. The survey concludes with an overview of the critical challenges, research gaps, and future directions. While reducing toxicity on online platforms continues to be an active area of research, a systematic study of various biases and their mitigation strategies will help the research community produce robust and fair models.

1 INTRODUCTION

Toxicity judgments depend on context, while datasets and annotations can carry biases, complicating fair automated detection. This survey organizes bias by source and target of harm and examines mitigation methods, including bias shift in knowledge-based methods.

Unique characteristics of online toxic speech

  • Toxicity judgments depend on context and vary across platforms and legal agencies, while datasets may reflect collection biases and unreliable annotations.These characteristics make toxic speech detection vulnerable to unintended bias.Source: §1 · p. 2

Scope of the survey

  • The survey focuses on methods for evaluating and mitigating bias in toxic speech detection, rather than surveying all toxicity-detection methods or NLP bias research.It also reviews popular datasets used to study bias and the reproducibility of debiasing methods.Source: §1 · p. 3
  • The survey classifies bias by sources and targets of harm, mapping sampling bias to sampling, annotation bias to annotation, and lexical bias to annotation and feature extraction.The paper notes that mitigation methods can apply at one or more data-transformation stages.Source: §1 · pp. 2–3, Figure 1 · p. 2
  • The survey’s case study introduces bias shift as a phenomenon in knowledge generalization-based mitigation methods.The case study is described in Appendix B.Source: §1 · p. 3

Survey methodology

  • The authors searched Google Scholar and relevant conference, journal, and workshop proceedings, then followed citations from seed papers to identify studies.Their shortlisting took place between September and November 2021 and focused mainly on research published from 2016 onward.Source: §1 · p. 3

2 CATEGORIES OF BIAS

The section organizes toxicity-detection bias by its sources in the data pipeline and by the groups or attributes harmed. It also relates where bias originates to where it becomes visible, and notes that psychographic bias remains underexplored.

2 CATEGORIES OF BIAS

  • Bias can enter during data collection or training and become visible in evaluation and in-the-wild testing.The figure links collection to sampling and lexical biases, training to co-occurrence and identity biases, and downstream use to their manifestation.Source: Figure 2 · p. 4
  • The section defines identity bias as prejudice against demographic, psychographic, or other identifiable physical or mental characteristics, including prejudice associated with spurious lexical correlations.It calls bias exhibited by a model unintended bias and uses “bias” to mean unintended bias in the paper.Source: §2 · pp. 4–5
  • The taxonomy groups bias by source—sampling, lexical, and annotation—and by harm target—racial, gender, and psychographic, including political affiliation.The source categories follow data-collection transformations; the target categories concern groups affected by downstream harm.Source: §2 · p. 5
  • Study of psychographic bias, concerning individuals’ beliefs and interests, has yet to gain popularity, and the survey calls for further exploration.The survey highlights bias categories tied to psychographic attributes as an area for future work.Source: §2 · p. 5
  • The survey’s taxonomy approaches proposed bias-mitigation methodologies from an application point of view.The section contrasts this framing with categorizations that segment papers by their motivation for handling bias.Source: §2 · p. 5

3 BIASES BASED ON SOURCE OF HARM

This section examines how sampling, annotation, and lexical biases arise in toxicity data and affect what models learn or judge as toxic. It reviews approaches to representing annotator disagreement and mitigating lexical bias through changes to data and models.

3 BIASES BASED ON SOURCE OF HARM

  • Speaker background can help annotators recognize offensive wording as non-toxic, but can also trigger latent bias and erroneous harmful labels that suppress minority voices.Source: §3 · p. 6

3.1 Sampling Bias

  • Text source affected toxic-dataset quality more than sampling strategy, while boosted-random datasets had less overall toxicity but more explicit abuse, whose higher ratio aided downstream performance.Source: §3.1 · p. 7
  • Models trained on topic-sampled datasets did not generalize across datasets; B1 measures topical similarity to hate lexicons, B2 topic-level keyword likelihood, and robust B1 scores indicate topic bias.Source: §3.1 · p. 7
  • Proxy keyword sampling produced only minimally more-toxic samples and did not validate its claim of lower bias.The keywords came from Reddit posts with roughly equal upvotes and downvotes; the filtered sample was compared with an equal-sized unfiltered tweet sample.Source: §3.1 · p. 7
  • Removing top toxic-label PMI terms from W&H sharply reduced performance, while limiting comments per user and augmenting with Davidson data improved unseen Hateval generalization [21], with topical homogeneity a caveat.W&H had more than 70% of sexist tweets from two authors and about 99% of racist tweets from one author; train-test topical homogeneity could explain the generalization gain.Source: §3.1 · pp. 7–8

3.2 Annotation Bias

  • Toxicity labels are subjective and sometimes ambiguous, and annotation bias cannot be mitigated through modeling alone; the survey calls for annotator and societal sensitization.The paper notes that annotation guidelines and annotator prejudice affect label quality.Source: §3.2 · p. 8
  • Generic Jigsaw2 guidance enabled more annotators to agree on toxicity, guided and unguided labels were highly correlated [81], and Fortuna et al. suggested hierarchical multi-class annotation.Source: §3.2 · p. 8
  • Experts and amateurs showed low agreement within their groups across all W&H labels.Source: §3.2 · p. 8
  • Models trained on native English speakers’ annotations outperformed those trained on non-native speakers’ annotations in F1 and sensitivity, with similar age and education differences but none between genders.The study also notes that collecting annotator demographics can be difficult and only covers demographic strata represented during dataset creation.Source: §3.2 · p. 8
  • Training on annotations from low-agreement groups yielded low performance on other groups’ test sets, and a high-agreement but distinct group also showed combined annotation bias.The study did not report multiple runs to assess variation from random model initialization.Source: §3.2 · pp. 8–9
  • Ignoring annotator disagreement causes a larger performance drop on subjective toxicity tasks than on objective tasks, while majority labels obscure the underlying distribution of human judgments.For five annotations split 3-to-2, majority voting gives [1,0], whereas the distribution is [0.6,0.4].Source: §3.2 · p. 9

3.3 Lexical Bias

  • Lexical bias arises when keyword-driven sampling and missing context lead models to associate bias-sensitive words with toxicity; length-sensitive upsampling improved pAUCED and reduced FPR without increasing FNR [24].These results were measured on SBET, whose static target-attribute list may not capture changing sample semantics after debiasing.Source: §3.3 · p. 9
  • AFLite and DataMaps reduced the correlation between toxicity labels and bias-sensitive words in Founta, while knowledge-based generalization replaces such words with WordNet ancestors but may introduce new biases [120].The survey’s case study addresses this potential bias shift after mitigation.Source: §3.3 · p. 10
  • Regularization reduced the model’s importance score for identity terms, suggesting decreased lexical bias; on the test split, multi-task learning improved AUC and GMB AUC and reduced false-positive bias without increasing false-negative bias.Source: §3.3 · p. 10
  • The ensemble method’s final model had lower FPR on Founta samples containing bias-sensitive words, but this came at the cost of accuracy and F1.It trained a naive model to learn biased features and encouraged a robust model to rely on other features before discarding the naive model.Source: §3.3 · p. 10

4 BIASES BASED ON TARGET OF HARM

This section reviews biases in toxicity and hate-speech detection based on targets of harm, including racial, gender, intersectional, cross-geographic, and political biases, along with mitigation approaches and gaps in their evaluation.

4 BIASES BASED ON TARGET OF HARM

  • Bias mitigation aims to shift models’ attention from specific terms toward the content’s overall context, but is limited by bias in the underlying language frameworks and embeddings.The section describes lexical correlations arising from data collection and annotation as a source of bias against identity groups.Source: §4 · p. 11

4.1 Racial Bias

  • Detecting dialect from text is contested and can proxy race, raising ethical concerns unless user background is considered; existing racial-bias work primarily focuses on AAE.Using Twitter-trained Blodgett LMs for dialect prediction risks carrying forward negative associations and increasing bias against non-white speakers [20].Source: §4.1 · p. 12
  • Blodgett LM-based dialect prediction may reproduce negative associations and stereotyping, so human-annotated dialect labels or methods without dialect ground truth are preferred [20].The survey notes that using such language models as black boxes could amplify bias against non-white speakers [20].Source: §4.1 · p. 12
  • Using tweet text alone to infer the author’s race, the study found n*gga and b*tch more significantly related to the NHW class, suggesting derogatory use of these terms.The Huang dataset was among the few works not relying on the Blodgett LM as a proxy for user race.Source: §4.1 · p. 12
  • Black-oriented tweets were often mislabeled as toxic, and an AAE-prevalent latent topic in Founta100k was more likely to be flagged as toxic [19].Sap et al. found a positive correlation between p(NHB|tweet) and toxic labels in Davidson and Founta100k; Davidson and Bhattacharya [19] identified the latent topic using structural topic modeling.Source: §4.1 · p. 12
  • Priming annotators with an author’s dialect or race made them more likely to label samples non-toxic, but priming can reinforce implicit bias.Dialect-aware label correction using GPT-3 [13] and paired Black-oriented/White-oriented tweets reduced false positives on Black-oriented test samples.Source: §4.1 · p. 13
  • Models trained on Davidson or Founta100k showed higher false-positive rates for Black-oriented comments and higher false-negative rates for White-oriented comments; disparities persisted after conditioning on identity terms.Sap et al. also reported higher positive rates for Black-oriented tweets, while Huang et al. [44] found higher FPED and FNED across languages.Source: §4.1 · p. 13
  • Reweighting training comments reduced bias against Black-oriented tweets across toxicity labels for models trained on W&H and Davidson datasets [64].Although reweighting was applied to the complete vocabulary, racial bias was the only bias evaluated [64].Source: §4.1 · p. 13
  • Racial debiasing showed mixed gains: one method reduced FPR and raised macro-F1, while ensemble debiasing lowered FPR on Black-oriented comments but reduced macro-F1 and barely changed out-of-domain disparity.Zhou et al. cautioned against real-world use of the first method because of GPT-3 limitations.Source: §4.1 · pp. 13–14

4.2 Gender Bias

  • Gender-bias research largely assumes binary identities, and non-binary gender bias is missing from toxicity-detection literature.Jigsaw is described as probably the only toxicity dataset with more inclusive, fine-grained gender and sexual-orientation labels.Source: §4.2 · p. 14
  • Pre-trained embeddings improved toxic-speech detection at the cost of greater equality-difference scores, with higher bias on W&H than Founta.Combining debiased embeddings and gender-swapping reduced FPED and FNED; length-sensitive upsampling achieved the best debiasing with a small AUC drop, while transfer learning had the highest AUC loss, attributed to incompatible labels.Source: §4.2 · p. 14

4.3 Intersectional bias

  • In Founta100k, Black-community tweets were more likely to be classified as abusive, while Black males were significantly more likely to be classified as hateful [51].Party was used as a control variable, and the authors call for more careful evaluation of intersectional bias.Source: §4.3 · pp. 14–15

4.4 Cross-geographic bias

  • Perspective API gave lower toxicity scores to terms toxic in Indian contexts, causing false negatives, but higher scores to “muslim” and “hindu” in non-toxic contexts, causing false positives [37].The survey notes that English-focused research concentrates attention on a few geographies, creating a knowledge gap that can lead to lexical bias.Source: §4.4 · p. 15
  • A two-step weakly supervised method analyzed tweets from seven countries, testing overrepresented terms in toxicity-varied templates; deletion, substitution, and balancing did not significantly reduce bias [37].The method separates term associations into descriptive (correlational or causal) and prescriptive (desirable or undesirable) axes.Source: §4.4 · p. 15

4.5 Political bias

  • A right-wing-biased dataset produced statistically lower F1 than the left-wing and neutral datasets, but the study did not test political bias in toxic labels.The survey questions treating politically inclined tweets as only non-toxic.Source: §4.5 · p. 15

5 MAJOR TAKEAWAYS: A SUMMARY

The survey’s takeaways emphasize that bias categories can overlap and that dataset construction, annotation, and social context shape toxicity-detection bias. Because one-time mitigation may not remain valid as systems evolve, bias should be evaluated throughout the workflow and at regular intervals.

Ongoing Evaluation

  • The survey frames bias mitigation as requiring ongoing external evaluation and criteria rather than expecting biases to disappear unaided.It likens continuous probing of the NLP pipeline to a regular supply of external energy.Source: §5 · p. 16
  • A one-time mitigation and evaluation in a controlled setting may not indicate bias reduction as a system evolves; the survey calls for regular checks throughout the workflow.The authors argue that mitigation in an evolving system of human interactions requires evaluating biases at every workflow step and analyzing them regularly.Source: §5 · p. 16

Categories of Bias

  • Bias categories based on sources of harm are not necessarily exclusive, and downstream-harm categories can arise from all three source categories.Lexical bias can result from sampling bias, annotation bias, or both.Source: §5 · p. 16

Sampling Bias

  • Text sources and sampled topics influence dataset bias more than sampling strategy, while platform-specific tolerance and policies also shape sampled data.Online social networks have varying tolerance towards toxic speech.Source: §5 · p. 16

Lexical Bias

  • Lexical correlations can obscure context, causing models to conflate identity disclosures such as “I am gay” with identity attacks such as “I hate all gays.”The survey notes that this conflation promotes disparity among social groups.Source: §5 · p. 16

Annotation Bias

  • Toxic-content annotation guidelines range from strict to loose, and annotator agreement varies without a fixed benchmark for labeling.The survey highlights examining annotator sensitivity to socio-cultural factors, including dialects and race.Source: §5 · p. 16

Racial Bias

  • Racial debiasing methods assume Black- and White-oriented samples share p(Y|X), but term acceptability can differ by speaker.Resources have aided racial-bias investigation [8] [44], but lack ground-truth labels for dialect identities.Source: §5 · p. 16

6 DIRECTION FOR FUTURE RESEARCH

The future-research agenda identifies open questions spanning interacting biases, data and annotation, social understanding, and the scope of toxicity systems. It calls for bias-aware practices across the pipeline and broader evaluation across domains, populations, languages, and modalities.

Proactive Bias Mitigation

  • The survey advocates starting mitigation early, such as during sampling and annotation, and maintaining bias awareness throughout the learning pipeline.Most surveyed papers demonstrated mitigation as a single-step solution [64].Source: §6 · pp. 17–18

Intersectional Bias and Social Understanding

  • How different biases interact in toxicity detection remains open, including the interplay of gender and race examined by Kim et al. [51].The survey also points to a gap between social and psychological research on harassment and its computational analysis.Source: §6 · pp. 17–18

Data Collection and Annotation

  • Data source and topic shape dataset characteristics, while random sampling skews toward non-hate content and topic-based priming can introduce unintended bias.Standardized annotation codebooks could reduce label variability across datasets.Source: §6 · p. 17

Gender and Side Effects

  • Toxicity research often treats gender as binary, while annotator awareness of gender fluidity can affect label consistency and models may flag identity statements as toxic.The survey notes that harms to non-binary people include erasure and misgendering.Source: §6 · p. 18

Language Is Not Static

  • Static hate lexicons and knowledge graphs cannot account for evolving language and socio-cultural context, so debiasing methods need to address language change.A learned hate-speech model intended to capture evolving content is proposed by Qian et al. [76].Source: §6 · p. 18

Out-of-Domain Evaluation

  • Encouraging in-domain mitigation results may fail to reduce disparity on out-of-domain data, leaving generalizability in question.The survey calls for diverse benchmarks to compare mitigation techniques.Source: §6 · p. 18

Collecting User Feedback

  • Where feasible, user-feedback infrastructure can collect data that better represents the target population.The Jigsaw team used user feedback as a source of bias mitigation for Perspective API.Source: §6 · p. 18

Beyond English Text

  • Bias research remains limited in non-English and code-mixed settings, and the survey identifies multimodal bias in memes and videos as an open research area.Toxicity judgments can vary across geographies, and the survey raises whether textual and image-based biases should be studied separately or jointly.Source: §6 · p. 19

7 CONCLUSION

The survey fills a gap in research focused on proposed bias-mitigation methods for toxic speech detection. It organizes bias by source and target of harm, discusses method limitations and future directions, and presents a case study of bias shift.

7 CONCLUSION

  • The survey notes that unintended bias has received substantial attention, but its mitigation still needs further exploration given online toxicity’s multifaceted and dynamic nature.Source: §7 · p. 19
  • The survey fills a gap in work focused on proposed methods for mitigating bias in toxic speech detection.It presents a taxonomy based on the source and target of harm to discuss methods, drawbacks, challenges, and future directions.Source: §7 · p. 19
  • The authors call for more attention to psychographic biases, including political-leaning discrimination in harmful-speech detection, which remains underexplored.Source: §7 · p. 19
  • The case study introduces bias shift from knowledge-based mitigation, with a shift from source terms toward generalized target terms that may also harm the non-toxic class.The authors describe it as a proof-of-concept case study and expect similar shifts from other knowledge-based generalization techniques.Source: §7 · p. 19

A EVALUATION METRICS

  • Figure 5 presents a taxonomy of bias-evaluation metrics used in toxic speech detection.Source: Figure 5 · p. 25

A.1 Background on bias evaluation

This background reviews psychological tests, embedding-association metrics, and bias-subspace approaches. It notes that these approaches have not found direct utility for evaluating bias in toxic speech classification.

A.1 Background on bias evaluation

  • IAT compares reaction times and accuracy across reversed word-category pairings, but its reliability is questionable and it is at best a weak indicator of bias.Source: §A.1 · p. 25
  • WEAT measures social bias through associations in non-contextual word embeddings, but its effectiveness depends heavily on the terms used to seed it [15].Source: §A.1 · p. 25
  • SEAT extends WEAT to sentence embeddings by inserting individual words into templates and applying WEAT to the resulting synthetic sentence embeddings [60].Source: §A.1 · p. 25
  • Bolukbasi et al. define a word’s gender bias through its embedding’s projection onto a gender subspace, derived using PCA on gender-differing word pairs [10] [97].Source: §A.1 · p. 25
  • Liang et al. extend subspace-based evaluation to contextual sentence embeddings, deriving the gender subspace from sentence pairs that differ in a gender-specific word [55].Source: §A.1 · p. 25
  • Neither psychological-test nor subspace-based evaluation has direct utility established for toxic-speech classification; analyzing fine-tuned embeddings or encoders is described as a rudimentary exploration.Source: §A.1 · p. 25

A.2 Unintended bias metrics

This section organizes classification-specific bias metrics used in toxic speech detection. It describes subgroup error and AUC measures alongside limitations involving thresholds, distributions, and evaluation-set construction.

A.2 Unintended bias metrics

  • The taxonomy covers bias-evaluation metrics developed specifically for classification models in toxic speech detection.Source: §A.2 · p. 26
  • FPED and FNED measure term-wise false-positive and false-negative rates against the corresponding rates on the full evaluation set; both depend on thresholds and binary labels.Source: §A.2 · p. 26
  • A low subAUC for a term indicates difficulty distinguishing toxic from non-toxic samples containing it, while a high value does not rule out high toxicity scores for non-toxic mentions.Source: §A.2 · p. 26
  • Pinned AUC evaluates a term or subgroup on a sampled auxiliary set, but its ability to reveal bias depends on label distribution; a balanced synthetic set can be tedious to generate.Source: §A.2 · pp. 26–27
  • BPSN and BNSP assess AUC on complementary term-containing and background examples; lower values indicate different confusion patterns for non-toxic versus toxic term mentions.Source: §A.2 · p. 27
  • GMBAUC combines subgroup AUC-based metrics across terms into a generalized mean, with overall AUC included to capture downstream performance.Source: §A.2 · p. 27
  • Pinned bias compares a term-only toxicity prediction probability with a pinned threshold, but its formulation uses one constant threshold even though real-world thresholds are dynamic.Source: §A.2 · p. 27
  • Pinned bias can replace or accompany pinned AUC and uses the same dataset, avoiding pinned AUC’s need for a balanced auxiliary dataset for each term.Source: §A.2 · p. 28

B CASE STUDY ON KNOWLEDGE-DRIFT

The case study examines WordNet-based lexical debiasing by replacing bias-sensitive words with ancestors. Its observations indicate a shift in bias toward generalized terms, while evaluation focuses on the original terms.

B CASE STUDY ON KNOWLEDGE-DRIFT

  • The case study replaces bias-sensitive words with WordNet ancestors, testing the knowledge-drift hypothesis that this shifts bias from the original words to their ancestors.Source: B CASE STUDY ON KNOWLEDGE-DRIFT · p. 28
  • After substitution, the study evaluates bias on the original terms but does not discuss evaluating it on the generalized terms that replace them.Source: B CASE STUDY ON KNOWLEDGE-DRIFT · p. 28
  • The table observations indicate bias shifts from source to target words, and because target terms are more likely in non-toxic comments, the shift may harm the non-toxic class.Source: B CASE STUDY ON KNOWLEDGE-DRIFT · p. 28, Table 4 · p. 28
  • The authors characterize the result as a proof of concept and expect other knowledge-based generalization techniques to suffer a similar shift rather than debiasing the dataset.Source: B CASE STUDY ON KNOWLEDGE-DRIFT · p. 28
  • The proposed next step is upsampling a generalized dataset with newly scraped comments, while accounting for the possibility that toxicity ratios differ for generalized terms.Source: B CASE STUDY ON KNOWLEDGE-DRIFT · p. 28

C REPRODUCIBILITY OF DEBIASING METHODS

The section reproduces selected debiasing methods to assess their engineering feasibility across datasets and evaluation settings. It extends prior studies with additional datasets and combinations of settings.

C REPRODUCIBILITY OF DEBIASING METHODS

  • The authors reproduce methods from Zhou et al. and Park et al. [70], adding W&H, Davidson [21], and Founta [34] datasets.The study aims to assess engineering feasibility as well as the techniques’ theoretical discussion.Source: C REPRODUCIBILITY OF DEBIASING METHODS · p. 28
  • The reproductions examine different dataset, embedding, and evaluation combinations for the debiasing methods.The authors describe these combinations as a way to give readers a glimpse of the methods’ settings.Source: C REPRODUCIBILITY OF DEBIASING METHODS · pp. 28–29
  • Table 5 presents reproduced racial-debiasing results for W&H [109], Davidson [21], and Founta [34].Its three panels correspond to those datasets.Source: Table 5 · p. 29

C.1 Racial Bias

The racial-bias reproductions find different outcomes across datasets and methods: Aflite’s FPR reduction is comparable on Founta, while LearnMixIn does not improve on the original composition for W&H and Davidson.

C.1 Racial Bias

  • On Founta [34], Aflite operates on only 33% of the data and provides a comparable FPR drop; LearnedMixIn has the smallest percentage of Black tweet samples wrongly labelled toxic.Source: §C.1 · p. 29
  • The authors did not test data filtering on W&H and Davidson because retaining 33% would leave about 5k and 8k tweets, respectively.They report LearnMixIn results for these small-scale datasets instead.Source: §C.1 · p. 29
  • On W&H and Davidson, LearnMixIn debiasing is not better than the original dataset composition.The authors assume this may reflect small sample volumes and other lexical or topical biases.Source: §C.1 · p. 29

C.2 Gender Bias

The gender-bias reproductions show trade-offs between predictive performance and fairness metrics, with no single setting best across datasets.

C.2 Gender Bias

  • Pre-trained word embeddings give higher FPED and FNED on average, while transfer learning raises F1 at the cost of increasing both fairness metrics.Table 6 reports embedding comparisons across W&H, Davidson, and Founta; Table 7 reports transfer-learning combinations on W&H and Davidson.Source: §C.2 · p. 29, Table 6 · p. 30, Table 7 · p. 30
  • Combining debiased word2vec, gender-swapping, and transfer learning offers the best tradeoff between reducing F1 and increasing FPED-FNED values, but no setting is best across datasets.The reproduced results cover W&H and Davidson in Table 7.Source: §C.2 · p. 29, Table 7 · p. 30

Related papers

Loading 2202.00126v3…