Challenge: Large language models and other massively pre-trained "foundation" models can easily adapt to a wide variety of downstream tasks in a process called finetuning.
Approach: They propose to use the bias transfer hypothesis to reduce social biases internalized by large language models during pre-training into harmful task-specific behavior after fine-tuning.
Outcome: The bias transfer hypothesis is the theory that social biases internalized by large language models during pre-training transfer into harmful task-specific behavior after fine-tuning.

Similar Papers

On Transferability of Bias Mitigation Effects in Language Model Fine-Tuning (2021.naacl-main)

Copied to clipboard

Challenge: PTLMs can exhibit biases against protected groups in a host of modeling tasks . but, fine-tuned LMs may propagate bias to downstream classifiers .
Approach: They propose to use upstream bias mitigation techniques to reduce bias on downstream tasks by fine-tuning an upstream model and applying it to a downstream model.
Outcome: The proposed model reduces bias on hate speech detection, toxicity detection and coreference resolution tasks over bias factors.
Analyzing Effects of Learning Downstream Tasks on Moral Bias in Large Language Models (2024.lrec-main)

Copied to clipboard

Challenge: Existing methods for fine-tuning large language models replicate and perpetuate social biases . pre-existing moral bias may be mitigated or amplified even when presented with opposing views .
Approach: They develop methods to assess the agreement of LMs to explicit codified norms . they find that introducing downstream tasks may lead to unexpected inconsistencies .
Outcome: The proposed model can be used to improve morality in data-scarce tasks.
When Do Pre-Training Biases Propagate to Downstream Tasks? A Case Study in Text Summarization (2023.eacl-main)

Copied to clipboard

Challenge: Existing studies have shown that large language models contain linguistic and societal biases, but it is unclear how these biase amplify to downstream tasks.
Approach: They investigate how name-nationality bias propagates from pre-training to downstream tasks . they show that these biases manifest themselves as hallucinations in summarization .
Outcome: The proposed model can reduce the rate of hallucinations, but does not change the types of biases that do appear.
Causal-Debias: Unifying Debiasing in Pretrained Language Models and Fine-tuning via Causal Invariant Learning (2023.acl-long)

Copied to clipboard

Challenge: Existing methods to remove unwanted stereotypical associations from pretrained language models (PLMs) are often focused on removing unwanted stereotypes from PLMs.
Approach: They propose a framework to remove unwanted stereotypical associations in pretrained language models . they propose bias-relevant factors are causal, while labelrelevant factors causal .
Outcome: The proposed framework reduces stereotypical associations after PLMs are fine-tuned . the proposed framework mitigates bias from a causal invariant perspective .
Don’t Just Clean It, Proxy Clean It: Mitigating Bias by Proxy in Pre-Trained Models (2022.findings-emnlp)

Copied to clipboard

Challenge: Transformer-based pre-trained models can encode societal biases in their contextual representations and in downstream predictions when fine-tuned on task-specific data.
Approach: They propose an approach that selectively eliminates stereotypical associations at fine-tuning, so that the model doesn't learn to excessively rely on those signals.
Outcome: The proposed approach reduces biases from identity words and frequently co-occurring proxies by > 60% in toxicity classification, and also extends to multiple identities.
Co2PT: Mitigating Bias in Pre-trained Language Models through Counterfactual Contrastive Prompt Tuning (2023.findings-emnlp)

Copied to clipboard

Challenge: Pre-trained language models can encode unfair social biases from large pre-training corpora and even amplify biase in downstream applications.
Approach: They propose a *debias-while-prompt tuning* method for mitigating biases via counterfactual contrastive prompt tuning on downstream tasks.
Outcome: The proposed method can mitigate biases on three extrinsic bias benchmarks and adapt to existing debiased language models.
Intrinsic Bias is Predicted by Pretraining Data and Correlates with Downstream Performance in Vision-Language Encoders (2025.naacl-long)

Copied to clipboard

Challenge: Recent work has found that vision-language models trained under the Contrastive Language Image Pre-training framework contain intrinsic social biases, but how these biase relates to downstream performance has been unclear.
Approach: They present the largest comprehensive analysis to-date of how upstream pre-training factors and downstream performance of CLIP models relate to their intrinsic biases.
Outcome: The proposed model performance analysis shows that the choice of pre-training dataset is the most significant upstream predictor of bias, whereas architectural variations have minimal impact.
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.
Assessing Combinational Generalization of Language Models in Biased Scenarios (2022.aacl-short)

Copied to clipboard

Challenge: Existing work focuses on assessing in-domain knowledge, but shedding light on what pre-trained Language Models learn is important.
Approach: They propose a method to assess a PLM's generalization capacity in biased scenarios by combining component combinations where it could be easy for the PLMs to learn shortcuts from the training corpus.
Outcome: The proposed model can overcome distribution shifts in the training corpus and with sufficient data.
StereoSet: Measuring stereotypical bias in pretrained language models (2021.acl-long)

Copied to clipboard

Challenge: Existing literature on stereotypical biases in language models is limited . current evaluations focus on measuring bias without considering language modeling ability .
Approach: They propose to measure stereotypical biases in four domains: gender, profession, race, and religion . they compare stereotypical and language modeling ability of popular models like BERT, GPT-2, RoBERTa and XLnet .
Outcome: The proposed model shows strong stereotypical biases in gender, profession, race, and religion domains.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations