PaperlayerView exact revision on arXiv

Source-linked AI summary

Bias and Fairness in Large Language Models: A Survey

Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, Nesreen K. Ahmed

arXiv:2309.00770v3cs.CLcs.AIcs.CYcs.LG

TL;DR

LLMs can learn, perpetuate, and amplify harmful social biases, while existing research does not always precisely describe the harms or fairness concepts involved. This survey formalizes those concepts and synthesizes evaluation and mitigation research through three taxonomies. It provides a structured guide to methods, limitations, and open challenges, while acknowledging that technical taxonomies do not fully address broader power hierarchies.

  • Problem

    LLMs can reproduce or amplify social biases, but bias and fairness research often does not precisely characterize who is harmed, why, or how harms relate to social hierarchies.

  • Method

    The paper formalizes social bias and fairness for NLP and organizes metrics, datasets, and mitigation techniques into three taxonomies.

  • Results

    The survey consolidates a wide range of research, defines forms of language-related harm, proposes fairness desiderata, and structures metrics, datasets, and mitigation methods for comparison.

  • Takeaways & Limitations

    The taxonomies provide researchers and practitioners a reference for selecting metrics, datasets, and mitigations while understanding their tradeoffs and limitations.

  • Takeaways & Limitations

    The survey is limited by a technical-solutionist perspective because its taxonomies are organized by implementation details rather than downstream usage contexts.

Abstract

from arXiv · show

Rapid advancements of large language models (LLMs) have enabled the processing, understanding, and generation of human-like text, with increasing integration into systems that touch our social sphere. Despite this success, these models can learn, perpetuate, and amplify harmful social biases. In this paper, we present a comprehensive survey of bias evaluation and mitigation techniques for LLMs. We first consolidate, formalize, and expand notions of social bias and fairness in natural language processing, defining distinct facets of harm and introducing several desiderata to operationalize fairness for LLMs. We then unify the literature by proposing three intuitive taxonomies, two for bias evaluation, namely metrics and datasets, and one for mitigation. Our first taxonomy of metrics for bias evaluation disambiguates the relationship between metrics and evaluation datasets, and organizes metrics by the different levels at which they operate in a model: embeddings, probabilities, and generated text. Our second taxonomy of datasets for bias evaluation categorizes datasets by their structure as counterfactual inputs or prompts, and identifies the targeted harms and social groups; we also release a consolidation of publicly-available datasets for improved access. Our third taxonomy of techniques for bias mitigation classifies methods by their intervention during pre-processing, in-training, intra-processing, and post-processing, with granular subcategories that elucidate research trends. Finally, we identify open problems and challenges for future work. Synthesizing a wide range of recent research, we aim to provide a clear guide of the existing literature that empowers researchers and practitioners to better understand and prevent the propagation of bias in LLMs.

1. Introduction

LLMs can perpetuate and amplify social harms, motivating a survey that formalizes bias and fairness concepts and organizes evaluation and mitigation research into taxonomies. The paper also examines approach limitations and defines a focused scope for LLM bias research.

  • Motivation: LLMs trained on large, uncurated Internet data can inherit stereotypes, derogatory language, and other harms disproportionately affecting marginalized communities.The survey describes harms including negative sentiment, toxicity, stereotypical associations, and limited recognition of language dialects.
  • Contributions: The survey consolidates and expands social-bias and fairness concepts, distinguishing harms and introducing fairness desiderata for NLP.It draws on machine learning, NLP, and sociolinguistics to describe bias more precisely.
  • Taxonomies: Its evaluation taxonomies organize metrics by model level and clarify their relationship with datasets, while mitigation methods are classified by intervention stage.Metric levels include embeddings, probabilities, and generated text; mitigation stages include pre-processing, in-training, intra-processing, and post-processing.
  • Scope and synthesis: The survey synthesizes research across NLP tasks and applications, examines limitations of approach classes, and offers insights and recommendations for future work.It aims to provide broader and deeper coverage than surveys limited to general algorithmic fairness or narrower NLP axes.
  • Scope and synthesis: The review focuses on bias in LLMs, primarily for English, and restricts included works to novel closed-form metrics, datasets, or mitigation techniques.Additional languages are included for machine translation and multilingual models.

2. Formalizing Bias and Fairness for LLMs

This section formalizes LLMs, social bias, and fairness, then organizes bias evaluation and mitigation around task, dataset, metric, and intervention-stage distinctions.

  • 2.1 Preliminaries: LLMs include autoregressive, autoencoding, and encoder-decoder models trained on hundreds of millions to trillions of tokens.They can be adapted through fine-tuning or task-agnostic constraints for bias mitigation.
  • 2.1 Preliminaries: An evaluation dataset and metric are distinct, with dataset structure and model access determining which metrics are appropriate.The metric set for a dataset is represented as ψ(D) ⊆ Ψ.
  • 2.2 Defining Bias for LLMs: The survey frames bias and fairness through a taxonomy of harms and examines how bias manifests across the LLM development and deployment cycle.It recognizes that technical approaches often fail to precisely describe the harms of model behavior.
  • 2.2.1 Social Bias and Fairness.: Social bias encompasses disparate treatment or outcomes rooted in historical and structural power asymmetries, including representational and allocational harms.The survey notes that “debiased” does not imply bias has been completely removed and that “neutral” words are context-dependent.
  • 2.2.1 Social Bias and Fairness.: Group fairness can hold while subgroup constraints fail, so intersectional identities motivate subgroup fairness notions.The survey presents group fairness as useful but comparatively weak for capturing overlapping subsets of a population.
  • 2.2.2 Bias in NLP Tasks.: Bias may appear locally or globally in generation and through gendered defaults or stereotypes in translation, retrieval, and question-answering.Examples include differences in next-token likelihoods and masculine defaults for ambiguous translations or retrieved concepts.
  • 2.3 Fairness Desiderata for LLMs: Rather than impose one fairness constraint, the survey proposes multiple fairness desiderata for NLP tasks, extending ideas beyond conventional classification settings.Traditional definitions such as equalized odds or statistical parity may apply unchanged to classification, while generation requires broader desiderata.
  • 2.4 Overview of Taxonomies: The survey taxonomizes evaluation metrics by dataset structure and access, datasets by targeted harms and social groups, and mitigation methods by intervention stage.Mitigation stages are pre-processing, in-training, intra-processing, and post-processing.

3. Taxonomy of Metrics for Bias Evaluation

Bias evaluation metrics must be matched to dataset structure and task, while separating metrics from the datasets they use. The survey organizes evaluation around task, bias type, data structure, and model output, illustrating how substitution can reveal task-dependent gender bias.

  • Evaluation facets: Gender substitution produced different question-answering responses: data-source strategies for a male data engineer versus technical-skill advice for a female data engineer.The contrast implies an assumption that male data engineers already possess needed technical skills.
  • Evaluation facets: In text completion, substituting gender yielded the same generated text, making the gender signal more implicit than in question-answering.The figure contrasts explicit gender substitution in question-answering with an unchanged completion output.
  • Evaluation facets: Evaluation considers task, bias type, input data structure, and whether metrics use embeddings, probabilities, or generated text.Metrics and datasets are often task-specific, while the underlying dataset and model output determine what a metric can measure.
  • Evaluation facets: Appropriate evaluation metrics depend on the dataset’s structure and task, so metrics should be categorized separately from datasets.The survey represents the appropriate subset as ψ(D) ⊆ Ψ for dataset D.

3.2 Taxonomy of Metrics based on What They Use

The survey groups LLM bias evaluation metrics by the model information they use. This taxonomy distinguishes embedding-based, probability-based, and generated-text-based approaches.

  • Metric types: Embedding-based metrics use dense vector representations, typically contextual sentence embeddings, to measure bias.
  • Metric types: Probability-based metrics use model-assigned probabilities to estimate bias, including scoring text pairs or answering multiple-choice questions.
  • Metric types: Generated-text-based metrics use text conditioned on prompts to measure co-occurrence patterns or compare outputs from perturbed prompts.

3.3 Embedding-Based Metrics

Embedding-based metrics assess bias through distances or associations in contextualized representations, extending word-embedding methods to sentence-level LLM encoders. The section also highlights that representation-level bias may not reliably reflect downstream harms.

  • Embedding-based metric taxonomy: Embedding-based metrics measure bias by computing distances or associations between neutral words, such as professions, and identity-related words, such as gender pronouns.For sentence-level encoders, these metrics operate on contextualized sentence embeddings.
  • Word embedding metrics: WEAT measures associations between protected social-group concepts and neutral attributes using a test statistic and an effect size, with larger effect sizes indicating stronger bias.WEAT* instead uses definitionally masculine and feminine words to capture stronger gender associations.
  • Sentence embedding metrics: SEAT adapts WEAT to contextualized embeddings by generating template-based sentences and using their sentence representations to evaluate social-group and attribute associations.Unbleached templates can target more specific dimensions of bias.
  • Sentence embedding metrics: CEAT generates sentences from social-group and neutral-attribute combinations, samples embeddings, and estimates bias magnitude from a random-effects distribution of effect sizes.The method uses variance-derived weights in its aggregation.
  • Sentence embedding metrics: Sentence Bias Score sums word-level gender-direction similarities within a sentence, weighting each word by a semantic importance score derived from max pooling.The gender direction is obtained from feminine–masculine embedding differences reduced with PCA.
  • Discussion and limitations: Embedding-space bias has weak or inconsistent relationships with downstream-task bias, and cosine-distance metrics may capture only superficial rather than subtler forms of harm.The survey notes that sentence templates may not capture derogatory language, disparate performance, exclusionary norms, or toxicity.

3.4 Probability-Based Metrics

Probability-based metrics evaluate bias by comparing model-assigned probabilities for inputs with perturbed social attributes, using masked-token or pseudo-log-likelihood methods. The section describes several scoring variants while cautioning that these metrics are not sufficient standalone deployment checks.

  • Probability-based metric taxonomy: Probability-based metrics compare predicted token probabilities for template sentences whose protected attributes have been perturbed.The survey illustrates masked-token and pseudo-log-likelihood techniques as two main classes.
  • Masked-token methods: Masked-token methods score completions for templates and normalize probabilities to isolate bias attributable to a neutral attribute rather than prior social-group preferences.DisCo compares normalized probability scores for opposing social-group words.
  • Pseudo-log-likelihood methods: Pseudo-log-likelihood methods estimate token probabilities by masking one token at a time and predicting it from the remaining sentence tokens.CrowS-Pairs uses this procedure to compare stereotypical and less-stereotypical sentence pairs.
  • Pseudo-log-likelihood methods: CAT compares stereotype, anti-stereotype, and meaningless options, while iCAT combines meaningful-option preference with an idealized 50% stereotype preference.The language-modeling score targets meaningfulness, whereas the stereotype score compares stereotype and anti-stereotype choices.
  • Pseudo-log-likelihood methods: AUL scores every token in an unmasked sentence, and AULA additionally weights tokens by attention to account for differing token importance.Unmasked inputs can improve prediction accuracy and avoid selection bias from choosing which words to mask.
  • Pseudo-log-likelihood methods: For CPS, CAT, AUL, and AULA, an ideal model has an average bias score of 0.5 when comparing stereotypical with less- or anti-stereotypical sentences.The score uses an indicator function over sentence-pair preferences.
  • Discussion and limitations: Probability-based metrics may correlate weakly with downstream-task bias and should be paired with additional metrics that assess downstream tasks more directly.Template limitations, restricted target words, questionable stereotype assumptions, and binary groupings can reduce generalizability and fairness validity.

3.5 Generated Text-Based Metrics

Generated text-based metrics evaluate free-text outputs through distributional, classifier-based, and lexicon-based analyses, often comparing outputs across perturbed social groups. The survey also highlights substantial sensitivity to decoding choices and limitations in metric validity.

  • Overview: Generated text-based metrics evaluate free-text outputs, which is especially useful when model probabilities or embeddings are inaccessible.They can therefore support evaluation of black-box LLMs.
  • Distribution Metrics: Distribution metrics compare associations or demographic representation in generated text against another group or a reference distribution.Examples include co-occurrence bias, Demographic Representation, and Stereotypical Associations.
  • Classifier Metrics: Classifier-based metrics use auxiliary models to score toxicity, sentiment, regard, or style and compare scores across socially perturbed prompts.Examples include Expected Maximum Toxicity, Toxicity Probability, Toxic Fraction, Score Parity, Regard Score, and Full Gen Bias.
  • Lexicon Metrics: Lexicon-based metrics perform word-level analysis by matching generated words against precompiled harmful-word lists or assigned bias scores.HONEST counts hurtful completions containing terms from the HurtLex lexicon.
  • Limitations: Decoding parameters such as generated length, sampling temperature, and beam-search top-k can drastically change bias levels and even reverse conclusions across datasets.The survey recommends reporting prompting sets and decoding parameters for transparency.
  • Limitations: Distribution metrics may poorly proxy downstream disparities, while classifier metrics can inherit biases from toxicity or sentiment classifiers.Co-occurrence may conflate harmful use with mention, and classifiers may disproportionately flag African-American English or stigmatized groups.

3.6 Recommendations

The survey recommends evaluating bias with care because existing metrics can correlate weakly with downstream behavior and vary with model specifications. It calls for theoretically grounded metrics with broader group coverage and transparent reporting.

  • Recommendations: Embedding- and probability-based metrics should not be the sole bias measures because their relationships with downstream bias can be weak or unreliable.They should be accompanied by direct evaluation of the downstream task.
  • Recommendations: Researchers should report model specifications, including prompting and decoding choices, because these choices affect evaluation results.The supplied recommendation passage begins this reporting guidance but does not enumerate all specifications.
  • Future Metrics: New metrics should be explicitly and theoretically grounded in the sociolinguistic mechanisms of bias they seek to measure.The survey reiterates desiderata including generalization to unconsidered groups, social-science grounding, and exhaustive coverage.

4. Taxonomy of Datasets for Bias Evaluation

The survey organizes LLM bias-evaluation datasets by structure, targeted harms, and social groups, covering counterfactual inputs and prompting-based tasks. It also documents validity, generalizability, and fairness-scope limitations.

  • Taxonomy: The dataset taxonomy summarizes each resource by structure, bias issue, and targeted social groups, while consolidating publicly available datasets for access.Table 4 records instance counts alongside issues and groups.
  • Counterfactual Inputs: Counterfactual input datasets perturb social-group terms while preserving other words and semantic meaning to compare model outputs across groups.They include masked-token and unmasked-sentence formats, which can be adapted to one another.
  • Masked Tokens: Masked-token datasets use blanks with specified candidate words and are suited to masked-token probability or pseudo-log-likelihood metrics.Coreference datasets such as WinoBias evaluate stereotypical and anti-stereotypical gender-occupation associations.
  • Unmasked Sentences: Unmasked resources include conversational, participatory, and community-sourced datasets covering demographic, gender, race, religion, and LGBTQ+ stereotypes.Examples include RedditBias, HolisticBias, and WinoQueer.
  • Validity Limitations: Several counterfactual benchmarks contain ambiguous stereotype definitions, and small semantic-preserving edits can drastically change measured bias.These issues question whether some benchmark scores validly indicate real-world stereotypes.
  • Scope Limitations: Many datasets have limited generalizability because they rely on United States contexts, narrow templates, or restricted notions of fairness.Gender and occupational associations dominate some benchmarks, while pronoun-choice tasks may not measure spontaneous biased generation.
  • Question Answering: Question-answering datasets assess whether models rely on stereotypes across multiple social groups, but group labels can create ambiguity about whom a completion harms.The survey suggests situation-based prompts as an alternative to over-reliance on social-group labels.

5. Taxonomy of Techniques for Bias Mitigation

The survey organizes bias mitigation techniques by where they intervene in the LLM workflow, from inputs and training to inference behavior and outputs.

  • Mitigation methods are categorized into pre-processing, in-training, intra-processing, and post-processing stages.The taxonomy distinguishes modifying inputs, training, inference behavior, and black-box model outputs.
  • Pre-processing methods modify datasets or prompts, whereas in-training methods modify the optimization process and model parameters.
  • Intra-processing methods modify model weights or decoding behavior without training or fine-tuning.
  • Post-processing methods operate on outputs from a black-box model without access to the model itself.

5.1 Pre-Processing Mitigation

Pre-processing mitigations alter data or prompts before model training or use, including augmentation, filtering, reweighting, generation, and instruction tuning. These methods can reduce bias but face substantial representational, scalability, and effectiveness limitations.

  • Pre-processing mitigations affect model inputs without intrinsically changing trainable parameters.They modify data or prompts through augmentation, filtering, reweighting, generation, and prompt changes.
  • Data augmentation: Data augmentation extends training distributions for underrepresented groups, commonly using counterfactual replacement of protected-attribute terms.Counterfactual data augmentation can create matched pairs by flipping gendered or definitionally gendered words while preserving grammatical and semantic correctness.
  • Data augmentation: Selective substitution and interpolation target data efficiency by replacing examples probabilistically or mixing counterfactual and original examples.Counterfactual data substitution randomly replaces gendered text with probability 0.5, while mixup interpolates augmented examples with originals and labels.
  • Filtering and reweighting: Filtering and reweighting target selected examples by demographic representation, bias level, labels, or protected attributes and their proxies.Methods may curate underrepresented or low-bias text, select highly biased examples, mask identity words and proxies, or reweight classes and socially associated examples.
  • Discussion and limitations: Pre-processing methods may be limited by incomplete or non-equivalent word lists, grammatical and factuality errors, binary-group assumptions, and new distribution imbalances.Modified prompting can also show limited effectiveness for removing bias from outputs, although control tokens may remain interpretable.
  • Discussion and limitations: Future alternatives include cost-sensitive reweighting and documenting dataset authors’ identities, backgrounds, and perspectives during data curation.

5.2 In-Training Mitigation

In-training mitigations alter the training procedure and update model parameters through gradient-based learning. The survey covers architectural, loss-based, representation, attention, and parameter-filtering approaches, while noting risks from aggressive pruning and implicit fairness assumptions.

  • In-training methods modify optimization through loss functions, next-word probabilities, parameter freezing, or removal of neurons contributing to harmful outputs.All such methods change model parameters through gradient-based training updates.
  • Architecture modifications: Architecture modifications change model configuration, including layers, encoders, decoders, or added adapter modules for parameter-efficient fine-tuning.ADELE adds debiasing adapter layers while updating only the injected layers.
  • Loss function modification: Equalizing objectives use embedding, attention, or predicted-distribution losses to encourage independence between social groups and stereotypical outputs.These objectives may use regularization, contrastive learning, adversarial learning, or reinforcement learning.
  • Representation-based mitigation: Representation-based methods reduce distances, project away protected directions, or minimize mutual information between social groups and learned semantics.Examples operate on embeddings and hidden states while sometimes preserving relative distances or pretrained linguistic knowledge.
  • Discussion and limitations: Loss functions and reinforcement-learning rewards encode fairness assumptions, often treating social groups or outcomes as invariant or interchangeable despite distinct harms.

5.3 Intra-Processing Mitigation

Intra-processing methods modify a pre-trained model’s inference behavior without further training or fine-tuning. They alter decoding, token probabilities, model weights, or modular debiasing components, but may trade bias reduction against output diversity.

  • Definition: Intra-processing methods modify pre-trained model behavior at inference without further training or fine-tuning.They include decoding strategies, post hoc parameter modifications, and modular debiasing networks.
  • Decoding Strategy Modification: Decoding interventions constrain next-token search, rerank alternatives, or use classifiers to steer generations away from harmful language.Examples include blocking offensive tokens, replacing potentially harmful tokens, and using toxicity or morality scores.
  • Modified Token Distribution: Modified token distributions adjust next-token probabilities to increase diversity or favor less biased outputs.Approaches include logit suppression, temperature sampling, expert–anti-expert comparisons, and projection-based debiasing.
  • Weight Distribution and Modular Debiasing Networks: Post hoc weight redistribution changes attention patterns, while modular networks apply stand-alone debiasing components for specific bias dimensions.These interventions can alter attention toward less stereotypical tokens or remove gender- and race-related dimensions during inference.
  • Discussion and Limitations: A central limitation is balancing bias mitigation with diverse generation, because biased-language classifiers can disproportionately filter minority dialects.Decoding heuristics may also be manipulated to increase toxic or hateful text.

5.4 Post-Processing Mitigation

Post-processing mitigations rewrite completed model outputs rather than changing model parameters, making them suitable for black-box systems. Methods include keyword replacement, machine translation, and other neural rewriting strategies, but rewriting decisions can themselves encode bias.

  • Overview: Post-processing removes bias from generated outputs without modifying the original model parameters, supporting use with black-box models.It operates after the model has produced a complete output.
  • Keyword Replacement: Keyword replacement identifies biased tokens and predicts substitutes while aiming to preserve the original output’s content and style.LIME and SHAP can identify words associated with bias before replacement.
  • Machine Translation: Machine-translation rewriters learn mappings from biased sentences to neutralized or otherwise unbiased alternatives using parallel corpora.Rules-based generation can create gender-neutral training pairs for biased sentences with gendered pronouns.
  • Other Neural Rewriters: Other neural rewriters modify encoded power, agency, or user-selected bias dimensions while reconstructing or paraphrasing the input.InterFair lets users critique word scores and adjust rewriting during inference.
  • Discussion and Limitations: Rewriting can disproportionately flag social-group language, erase important context, reduce diversity, and remain constrained by available parallel training corpora.The choice of which outputs to rewrite is subjective and value-laden.

5.5 Recommendations

The recommendations emphasize preserving social context, aligning objectives with explicit fairness criteria, protecting output diversity, and retaining important contextual information.

  • Recommendations: Pre-processing should not flatten power imbalances, and identity-word lists should validly and completely represent the social groups they model.Masking or replacing identity words may miss power dynamics specific to particular groups.
  • Recommendations: Fairness objectives should state their assumptions explicitly and align loss or regularization choices with a specific fairness criterion.Cost-sensitive learning is suggested as one way to increase the weight of minority classes.
  • Recommendations: Decoding-based mitigation should be balanced against output diversity and validated so toxic-language heuristics do not marginalize dialects or minoritized voices.The recommendation directly addresses filtering risks in modified decoding strategies.
  • Recommendations: Mitigation methods should preserve important contexts rather than treating removal of protected attributes as universally appropriate.The supplied passage introduces this recommendation but does not provide its full elaboration.

6. Open Problems & Challenges

The paper identifies open problems spanning power, values, participation, evaluation, resources, intervention design, theory, and performance–fairness trade-offs. It calls for broader, more context-sensitive, and better-grounded approaches to bias and fairness in LLMs.

  • Addressing Power Imbalances: Bias mitigation should engage historical and institutional power hierarchies and center vulnerable groups from system design through deployment.Technical interventions alone are described as incomplete responses to societal injustices.
  • Addressing Power Imbalances: Participatory research incorporates community members’ lived experiences into dataset curation, mitigation design, and evaluation strategies.HolisticBias and WinoQueer are cited as examples of participatory dataset creation.
  • Conceptualizing Fairness for NLP: Fairness concepts should make normative assumptions explicit and account for social, cultural, historical, political, and regional contexts beyond Western English settings.The paper specifically calls for stronger grounding in theories of social change and more resources for other languages and dialects.
  • Conceptualizing Fairness for NLP: Future metrics and mitigations should identify who is harmed and how, represent more social groups, and avoid exclusive or binary subgroup definitions.The paper highlights the exclusion of trans, nonbinary, and intersex identities by binary gender modeling.
  • Refining Evaluation Principles: Reporting standards remain open because metrics can disagree, and evaluation results require context about datasets, hyperparameters, harms, and unrepresented cases.The paper notes that claiming successful debiasing from any single decreasing metric can mislead.
  • Future Research Directions: Evaluation faces reliability, validity, coverage, and scalability limitations, while mitigation research needs broader resources and comparisons across intervention stages.The survey also identifies hybrid techniques and theoretical guarantees as underdeveloped areas.
  • Future Research Directions: Future work should characterize performance–fairness trade-offs and track whether performance changes affect social groups uniformly.The paper recommends analyzing Pareto frontiers across hyperparameter values.

7. Limitations

The survey’s technical focus is bounded by broader societal and pipeline considerations. It also covers English-language papers narrowly.

  • Scope of technical solutions: Technical bias solutions cannot replace broader societal action addressing power hierarchies and structural oppression.The authors warn that technical solutionism can produce surface-level repairs while preserving underlying inequity and injustice.
  • Pipeline coverage: The survey focuses mainly on model training and evaluation rather than the full development and deployment pipeline.The authors identify problem formulation, data collection, deployment, and real-world integration as additional decision points.
  • Language coverage: The survey is limited to English-language papers.

8. Conclusion

The paper surveys bias evaluation and mitigation for LLMs, formalizes social bias and fairness concepts, and organizes the literature into three taxonomies. It also consolidates datasets and identifies open problems to guide future research.

  • Conclusion: The survey synthesizes research on bias evaluation and mitigation techniques for LLMs to describe the current research landscape.
  • Conclusion: It formalizes social bias and fairness in NLP, defines distinct forms of harm, and proposes initial fairness desiderata for LLMs.
  • Conclusion: Three taxonomies organize metrics and datasets for bias evaluation and intervention stages for mitigation techniques.The survey also consolidates publicly available datasets to increase accessibility and categorizes trends within mitigation stages.
  • Conclusion: The paper outlines actionable open problems and challenges to guide future research.
Loading 2309.00770v3…