Papers with debiasing
NeuroX Library for Neuron Analysis of Deep NLP Models (2023.acl-demo)
Copied to clipboard
| Challenge: | NeuroX is an open-source toolkit to conduct neuron analysis of natural language processing models. |
| Approach: | They propose a Python toolkit to conduct neuron analysis of natural language processing models. |
| Outcome: | a new open-source toolkit enables neuron analysis of natural language processing models . the framework provides a framework for data processing and evaluation, making it easier for researchers and practitioners to perform neuron analyses. |
So Different Yet So Alike! Constrained Unsupervised Text Style Transfer (2022.acl-long)
Copied to clipboard
| Challenge: | Automated transfer of text between domains does not maintain other attributes between the source and translated text. |
| Approach: | They propose a method for automatic transfer of text between domains that preserves semantic content but changes other attributes. |
| Outcome: | The proposed method retains lexical, syntactic and domain-specific constraints between domains for multiple benchmark datasets, including ones where more than one attribute change. |
Robust Natural Language Understanding with Residual Attention Debiasing (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing ensemble-based debiasing methods do not address unintended dataset biases . attention plays a crucial role in providing robust prediction in NLU models . |
| Approach: | They propose an end-to-end debiasing method that mitigates unintended biases from attention. |
| Outcome: | The proposed method improves the OOD performance of BERT-based models on three benchmarks. |
Identifying and Mitigating Social Bias Knowledge in Language Models (2025.findings-naacl)
Copied to clipboard
| Challenge: | Existing methods for debiasing may generate incorrect or nonsensical predictions but leave aside individual commonsense facts, resulting in modified knowledge that elicits unreasonable or undesired predictions. |
| Approach: | They propose a framework that identifies encoding locations of biases within language models and then applies the Fairness-Stamp (FAST) they also propose 'BiaScope' to evaluate the retention of commonsense knowledge and generalization across paraphrased social biase. |
| Outcome: | The proposed framework surpasses state-of-the-art baselines with superior debiasing performance while not compromising the overall model capability for knowledge retention and prediction. |
Debiasing by obfuscating with 007-classifiers promotes fairness in multi-community settings (2025.coling-main)
Copied to clipboard
| Challenge: | a number of studies have focused on the mitigation of biases in text classifiers. |
| Approach: | They propose an obfuscation-based data augmentation debiasing approach to reduce bias . they add to the training data *obfuses* versions of *all* false positive instances . |
| Outcome: | The proposed approach reduces bias for almost all of the tests without sacrificing false positive rates or F1 scores for minority or majority communities. |
Adversarial Scrubbing of Demographic Information for Text Classification (2021.emnlp-main)
Copied to clipboard
Somnath Basu Roy Chowdhury, Sayan Ghosh, Yiyuan Li, Junier Oliva, Shashank Srivastava, Snigdha Chaturvedi
| Challenge: | Existing frameworks to debias contextual representations can encode undesirable attributes, like demographic associations of the users, while being trained for an unrelated task. |
| Approach: | They propose an adversarial learning framework to debias contextual representations by encoding undesirable attributes while being trained for an unrelated task. |
| Outcome: | The proposed framework debiases representations on 8 datasets while remaining informative on the target task. |
Towards Argument Mining for Social Good: A Survey (2021.acl-long)
Copied to clipboard
| Challenge: | Argument Mining is a social science-based approach to analysis and analysis of arguments. |
| Approach: | They propose a novel definition of argument quality which integrates the social science literature and the argument quality. |
| Outcome: | The proposed definition of argument quality integrates the social science literature and the argument quality debate. |
Towards Understanding Task-agnostic Debiasing Through the Lenses of Intrinsic Bias and Forgetfulness (2024.findings-acl)
Copied to clipboard
Guangliang Liu, Milad Afshari, Xitong Zhang, Zhiyu Xue, Avrajit Ghosh, Bidhan Bashyal, Rongrong Wang, Kristen Johnson
| Challenge: | Debiasing Pretrained Language Models (PLMs) are task-agnostic and can be generalizable, but its impact on language modeling ability and the risk of relearning social biases remain as the two most significant challenges. |
| Approach: | They propose a framework which can Propagate Socially-fair Debiasing to Downstream Fine-tuning to alleviate the forgetting issue of PLMs by regularizing debiased attention heads based on the PLM’s bias levels from stages of pretraining and debiase. |
| Outcome: | The proposed framework can Propagate Socially-fair Debiasing to Downstream Fine-tuning, indicating that the ineffectiveness of debiase can be alleviated by overcoming the forgetting issue through regularizing successfully debiased attention heads based on the PLMs’ bias levels from stages of pretraining and debiases. |
DUnE: Dataset for Unified Editing (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing models are susceptible to errors necessitating a comprehensive retraining process. |
| Approach: | They propose to define an edit as any natural language expression that solicits a change in the model’s outputs. |
| Outcome: | The proposed editing benchmarks show that retrieval-augmented language modeling outperforms specialized editing techniques and neither set of approaches has fully solved the generalized editing problem covered by the proposed benchmark. |
In-Contextual Gender Bias Suppression for Large Language Models (2024.findings-eacl)
Copied to clipboard
| Challenge: | Prior work has proposed debiasing methods that require human labelled examples, data augmentation and fine-tuning of LLMs, which are computationally expensive. |
| Approach: | They propose to suppress gender biases by providing textual preambles from manually designed templates and real-world statistics without accessing model parameters. |
| Outcome: | The proposed methods suppress gender biases in English LLMs using a CrowsPairs dataset without accessing model parameters. |
Towards Stable Natural Language Understanding via Information Entropy Guided Debiasing (2023.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to debiase Natural Language Understanding models use dataset biases instead of learning the intended task. |
| Approach: | They propose a debiasing framework that detects and purifies dataset biases using information entropy. |
| Outcome: | The proposed framework improves the stability of performance on out-of-distribution datasets for a set of widely adopted NLU models. |
The Gaps between Fine Tuning and In-context Learning in Bias Evaluation and Debiasing (2025.coling-main)
Copied to clipboard
| Challenge: | FT-based debiasing methods cause a performance degradation in downstream tasks . FT works by updating some or all parameters, while ICL uses prompts without modifying the model parameters. |
| Approach: | They propose to use ICL to customize PLMs to downstream tasks without parameter updates. |
| Outcome: | The proposed method lowers the performance degradation of FT-based debiasing methods compared to FT models . the proposed method improves performance on large datasets while allowing for smaller changes to PLMs . |
How Gender Debiasing Affects Internal Model Representations, and Why It Matters (2022.naacl-main)
Copied to clipboard
| Challenge: | Existing studies of gender bias in NLP focus on extrinsic or intrinsic bias, but the relationship between extrindic and intrinsic bias is relatively unknown. |
| Approach: | They propose a framework to measure extrinsic and intrinsic bias together and propose metric to measure debiasing and intrinsic debiases. |
| Outcome: | The proposed framework provides a comprehensive perspective on bias in NLP models, which can be applied to deploy NLP systems in a more informed manner. |
Bias Vector: Mitigating Biases in Language Models with Task Arithmetic Approach (2025.coling-main)
Copied to clipboard
| Challenge: | Using language models (LMs) has increased in use, and the use of biases and stereotypes is creating social problems. |
| Approach: | They propose a method to mitigate LM biases by continual training on biased data . they use masked language modeling to construct a Bias Vector as the difference between biased LMs and pre-trained LM weights . |
| Outcome: | The proposed method improves on the GLUE and SEAT benchmarks. |
DINER: Debiasing Aspect-based Sentiment Analysis with Multi-variable Causal Inference (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to debiase ABSA focus on single-variable causal inference . aspect-based sentiment analysis models are prone to learn spurious correlations from annotation biases . |
| Approach: | They propose a framework based on multivariable causal inference for debiasing ABSA . they propose to model different types of biases based upon different causal intervention methods . |
| Outcome: | The proposed framework tackles different types of biases based on different intervention methods. |
Mitigating Biases in Language Models via Bias Unlearning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent debiasing approaches target different demographic groups, harming fairness and discrimination. |
| Approach: | They propose a model debiasing framework which targets stereotypes by unlearning stereotype forgetting and anti-stereotype retention. |
| Outcome: | The proposed framework outperforms existing methods in mitigating bias while retaining language modeling capabilities. |
Generalized but not Robust? Comparing the Effects of Data Modification Methods on Out-of-Domain Generalization and Adversarial Robustness (2022.findings-acl)
Copied to clipboard
| Challenge: | Data modification has been proposed as an effective solution for generalizing to out-of-domain (OOD) inputs. |
| Approach: | They propose to use data modification to generalize to out-of-domain inputs . they also analyze their adversarial robustness using a synthetic dataset . |
| Outcome: | The proposed data modification strategies improve OOD accuracy and AR, but data filtering hurts OOD on other tasks. |
Data-Centric Explainable Debiasing for Improving Fairness in Pre-trained Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing data-centric debiasing strategies mainly leverage explicit bias words for counterfactual data augmentation to balance the training data. |
| Approach: | They propose a method which uses an explainability method to search for implicit bias words to assist in debiasing PLMs. |
| Outcome: | Extensive results show that the proposed method achieves state-of-the-art debiasing performance and strong generalization while maintaining predictive abilities. |
Social-Group-Agnostic Bias Mitigation via the Stereotype Content Model (2023.acl-long)
Copied to clipboard
Ali Omrani, Alireza Salkhordeh Ziabari, Charles Yu, Preni Golazizian, Brendan Kennedy, Mohammad Atari, Heng Ji, Morteza Dehghani
| Challenge: | Existing methods for mitigating bias require social-group-specific word pairs for each social attribute (e.g., gender) Existing approaches require only one social attribute, rendering them impractical and costly . |
| Approach: | They propose that stereotype content models capture the underlying connection between bias and stereotypes by embedding only two psychological dimensions of warmth and competence. |
| Outcome: | The proposed method performs comparably to group-specific debiasing on multiple bias benchmarks, but has theoretical and practical advantages over existing methods. |
Recent Advances in Online Hate Speech Moderation: Multimodality and the Role of Large Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | HS is any communication demeaning a person or a group based on social or ethnic characteristics that undermines social harmony and individual safety . the recent Israel-Hamas conflict has escalated both anti-Muslim and anti-Semitic sentiments worldwide . |
| Approach: | They examine the role of large language models and large multimodal models in HS moderation . they examine how text, images, and audio interact to spread hate speech . |
| Outcome: | The findings highlight the need for solutions in low-resource settings and highlight the gaps in existing methods. |
Addressing Healthcare-related Racial and LGBTQ+ Biases in Pretrained Language Models (2024.findings-naacl)
Copied to clipboard
| Challenge: | Pretrained language models (PLMs) propagate social stigmas and stereotypes, a critical concern given their widespread use. |
| Approach: | They adapt two intrinsic bias benchmarks to quantify racial and LGBTQ+ biases in prevalent PLMs and empirically evaluate the effectiveness of various debiasing methods in mitigating these biase. |
| Outcome: | The proposed methods reduce biases without compromising performance in downstream tasks. |
Neutralizing Gender Bias in Word Embeddings with Latent Disentanglement and Counterfactual Generation (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Recent research shows word embeddings have strong gender biases in embeddable spaces . a proposed method can be used to debiase word embeds without loss of semantic information . |
| Approach: | They propose a latent disentanglement method with a siamese auto-encoder structure with an adapted gradient reversal layer to debiase word embeddings. |
| Outcome: | The proposed method can preserve semantic information during debiasing while minimizing loss of semantic information for extrinsic NLP tasks. |
Dynamically Disentangling Social Bias from Task-Oriented Representations with Adversarial Attack (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing methods to learn representations from text often reflect social biases . previous methods rely on pre-specified direction or suffer from unstable training . |
| Approach: | They propose an adversarial disentangled debiasing model to decouple social bias attributes from intermediate representations trained on the main task. |
| Outcome: | The proposed model decouples social bias attributes from intermediate representations trained on the main task. |
Contrastive Learning as a Polarizer: Mitigating Gender Bias by Fair and Biased sentences (2024.findings-naacl)
Copied to clipboard
| Challenge: | Recent studies have highlighted social biases inherent in training data can lead models to learn and propagate them. |
| Approach: | They propose a contrastive learning method that uses anchor points to push further negatives and pull closer positives within the representation space. |
| Outcome: | The proposed method achieves state-of-the-art in the ICAT score on the StereoSet, a benchmark for measuring bias in models. |
Debiasing with Sufficient Projection: A General Theoretical Framework for Vector Representations (2024.naacl-long)
Copied to clipboard
| Challenge: | Pre-trained vector representations can inadvertently encode undesirable social biases. |
| Approach: | They propose a framework for reducing bias by transforming vector representations to an unbiased subspace using sufficient projection. |
| Outcome: | The proposed framework mitigates bias across debiasing and fairness tasks and across various vector representation types, including word embeddings and output representations of transformer models. |
Gender-tuning: Empowering Fine-tuning for Debiasing Pre-trained Language Models (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for debiasing are resource-intensive and costly. Existing solutions for debiansing require fine-tuning on downstream tasks. |
| Approach: | They propose to integrate Masked Language Modeling (MLM) training objectives into fine-tuning’s training process to debiase the PLMs. |
| Outcome: | The proposed approach outperforms the state-of-the-art baselines in terms of gender bias scores while improving PLMs’ performance solely using the downstream tasks’ dataset. |
Mitigating Biases of Large Language Models in Stance Detection with Counterfactual Augmented Calibration (2025.naacl-long)
Copied to clipboard
Ang Li, Jingqian Zhao, Bin Liang, Lin Gui, Hui Wang, Xi Zeng, Xingwei Liang, Kam-Fai Wong, Ruifeng Xu
| Challenge: | Large language models generate biased stances due to spurious correlations and preference towards certain individuals and topics. |
| Approach: | They propose a counterfactual Augmented Calibration Network to calibrate potential bias in stance detection of large language models. |
| Outcome: | The proposed calibration network can mitigate biases of large language models, achieving state-of-the-art results. |
Backpack Language Models (2023.acl-long)
Copied to clipboard
| Challenge: | We train a 170Mparameter Backpack language model on OpenWebText, matching the loss of a 6Bparameter Transformer. |
| Approach: | They propose a neural architecture that learns multiple non-contextual sense vectors for each word in a vocabulary and represents a word as a context-dependent, non-negative linear combination of sense vector. |
| Outcome: | The proposed model outperforms a GPT-2's word embeddings on lexical similarity evaluations and can be used to perform controllable text generation and debiasing. |
Discovering the Unknown Knowns: Turning Implicit Knowledge in the Dataset into Explicit Training Examples for Visual Question Answering (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods address this issue by introducing an auxiliary task such as visual grounding, cycle consistency, or debiasing. |
| Approach: | They propose a data augmentation pipeline to turn “known” knowledge into training examples for VQA. |
| Outcome: | The proposed model can handle multi-modal information and is based on human-annotated examples. |
Hire Me or Not? Examining Language Model’s Behavior with Occupation Attributes (2025.coling-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) have been widely integrated into production pipelines due to their impressive performance across multiple tasks. |
| Approach: | They construct a dataset using a standard occupation classification knowledge base and tested it on three families of LLMs. |
| Outcome: | The proposed framework analyzes LLMs’ behavior with respect to gender stereotypes in the context of occupation decision making. |
It’s All in the Name: Mitigating Gender Bias with Name-Based Counterfactual Data Substitution (D19-1)
Copied to clipboard
| Challenge: | Existing attempts to mitigate gender bias rely on operationalisation of gender bias as a projection over a linear subspace. |
| Approach: | They propose to operationalise gender bias as a linear subspace and augmented a corpus to remove bias by swapping all inherently-gendered words in the copy. |
| Outcome: | The proposed approach outperforms projection-based methods at the task of drawing non-biased gender analogies by an average of 19% across both corpora. |
Fair Federated Learning with Biased Vision-Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing literature ignores the inherent group unfairness within CLIP and its ethical implications on FL applications. |
| Approach: | They propose a fairness-aware adaptation framework for CLIP in federated learning . they propose to leverage biased pre-trained VLMs to build fair FL frameworks . |
| Outcome: | The proposed framework addresses unique bias in FL, triggered by data heterogeneity . it trains a fair FL model with fairness-aware deep visual prompting (DVP) Extensive results on human face attribute recognition (FAR) applications show it outperforms state-of-the-art FL models . |
Are All Spurious Features in Natural Language Alike? An Analysis through a Causal Lens (2022.emnlp-main)
Copied to clipboard
| Challenge: | 'spurious correlations' have been used in NLP to informally denote any undesirable feature-label correlations. |
| Approach: | They formalize this distinction using a causal model and probabilities of necessity and sufficiency, which delineates causal relations between a feature and a label. |
| Outcome: | The proposed model is invariant to the feature, but not sufficient for prediction. |
Prompt Tuning Pushes Farther, Contrastive Learning Pulls Closer: A Two-Stage Approach to Mitigate Social Biases (2023.acl-long)
Copied to clipboard
| Challenge: | Existing debiasing techniques use Counterfactual Data Augmentation (CDA) to balance the training corpus, but this technique slightly modifies the original corpus limiting the representation distance between different demographic groups. |
| Approach: | They propose a two-stage debiasing model using Contrastive learning with Continuous Prompt Augmentation to mitigate social biases in PLMs’ encoding. |
| Outcome: | The proposed model outperforms baselines in terms of debiasing performance while maintaining the language modeling capability of PLMs. |
Applying Intrinsic Debiasing on Downstream Tasks: Challenges and Considerations for Machine Translation (2024.emnlp-main)
Copied to clipboard
| Challenge: | In this study, we examine three considerations for intrinsic debiasing in neural machine translation models. |
| Approach: | They propose to measure the extrinsic bias of neural machine translation models by embedding them in a neural embeddable space and using different tokens to debias them. |
| Outcome: | The proposed methods over-rely on gender stereotypes and over-represent them in their models. |
Mitigating Bias in RAG: Controlling the Embedder (2025.findings-acl)
Copied to clipboard
| Challenge: | a promising modular AI system enhances factuality and privacy in large language models . however, each component introduces its own biases into the RAG system, which could cause representational harm and unsafe user interactions. |
| Approach: | They study the conflict between biases of each component and their relationship to the overall bias of the retrieval augmented generation system. |
| Outcome: | The proposed model can be controlled by the embedder while maintaining utility . the proposed model is more fair than existing models, the authors show . |
Debiasing Multilingual LLMs in Cross-lingual Latent Space (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies have evaluated their cross-lingual transferability by directly applying these methods to LLM representations, revealing their limited effectiveness across languages. |
| Approach: | They propose to perform debiasing in a joint latent space rather than directly on LLM representations by using an autoencoder trained on parallel TED talk scripts. |
| Outcome: | The proposed method improves both the overall debiasing performance and cross-lingual transferability of the proposed techniques across four languages. |
Distilling the Essence, Discarding the Dross: Improving Fairness in Multimodal Large Language Models via Historical Reflection-Guided Prompt Optimization (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to debiase MLLMs rely on handcrafted prompts that are brittle and difficult to generalize across tasks and bias types. |
| Approach: | They propose an adaptive self-debiasing framework that optimizes task-specific debiasers to suppress stereotypical outputs. |
| Outcome: | The proposed framework suppresses stereotypical outputs while maintaining performance. |
Iterative Multilingual Spectral Attribute Erasure (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for debiasing are unable to exploit this opportunity because they operate on individual languages. |
| Approach: | They propose to iterate multilingual spectral attribute error (IMSAE) to mitigate joint bias subspaces across multiple languages through iterative SVD-based truncation. |
| Outcome: | The proposed method outperforms monolingual and cross-lingual approaches while maintaining model utility. |
B-APO: Bias-Targeted Adversarial Preference Optimization for Debiasing Multimodal Large Language Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing debiasing methods create biased responses by completely removing an entire modality, forming an extreme and static training environment. |
| Approach: | They propose a method to debiase multimodal large language models by masking one modality and then enlarge the margin between clean and adversarial responses. |
| Outcome: | The proposed method achieves superior debiasing performance while maintaining general capabilities. |