@inproceedings{panda-etal-2022-dont,
title = "Don`t Just Clean It, Proxy Clean It: Mitigating Bias by Proxy in Pre-Trained Models",
author = "Panda, Swetasudha and
Kobren, Ari and
Wick, Michael and
Shen, Qinlan",
editor = "Goldberg, Yoav and
Kozareva, Zornitsa and
Zhang, Yue",
booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2022",
month = dec,
year = "2022",
address = "Abu Dhabi, United Arab Emirates",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2022.findings-emnlp.372/",
doi = "10.18653/v1/2022.findings-emnlp.372",
pages = "5073--5085",
abstract = "Transformer-based pre-trained models are known to encode societal biases not only in their contextual representations, but also in downstream predictions when fine-tuned on task-specific data.We present D-Bias, an approach that selectively eliminates stereotypical associations (e.g, co-occurrence statistics) at fine-tuning, such that the model doesn`t learn to excessively rely on those signals.D-Bias attenuates biases from both identity words and frequently co-occurring proxies, which we select using pointwise mutual information.We apply D-Bias to a) occupation classification, and b) toxicity classification and find that our approach substantially reduces downstream biases (e.g. by {\ensuremath{>}} 60{\%} in toxicity classification, for identities that are most frequently flagged as toxic on online platforms).In addition, we show that D-Bias dramatically improves upon scrubbing, i.e., removing only the identity words in question.We also demonstrate that D-Bias easily extends to multiple identities, and achieves competitive performance with two recently proposed debiasing approaches: R-LACE and INLP."
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="panda-etal-2022-dont">
<titleInfo>
<title>Don‘t Just Clean It, Proxy Clean It: Mitigating Bias by Proxy in Pre-Trained Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Swetasudha</namePart>
<namePart type="family">Panda</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ari</namePart>
<namePart type="family">Kobren</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Michael</namePart>
<namePart type="family">Wick</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qinlan</namePart>
<namePart type="family">Shen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2022-12</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2022</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yoav</namePart>
<namePart type="family">Goldberg</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zornitsa</namePart>
<namePart type="family">Kozareva</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Abu Dhabi, United Arab Emirates</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Transformer-based pre-trained models are known to encode societal biases not only in their contextual representations, but also in downstream predictions when fine-tuned on task-specific data.We present D-Bias, an approach that selectively eliminates stereotypical associations (e.g, co-occurrence statistics) at fine-tuning, such that the model doesn‘t learn to excessively rely on those signals.D-Bias attenuates biases from both identity words and frequently co-occurring proxies, which we select using pointwise mutual information.We apply D-Bias to a) occupation classification, and b) toxicity classification and find that our approach substantially reduces downstream biases (e.g. by \ensuremath> 60% in toxicity classification, for identities that are most frequently flagged as toxic on online platforms).In addition, we show that D-Bias dramatically improves upon scrubbing, i.e., removing only the identity words in question.We also demonstrate that D-Bias easily extends to multiple identities, and achieves competitive performance with two recently proposed debiasing approaches: R-LACE and INLP.</abstract>
<identifier type="citekey">panda-etal-2022-dont</identifier>
<identifier type="doi">10.18653/v1/2022.findings-emnlp.372</identifier>
<location>
<url>https://aclanthology.org/2022.findings-emnlp.372/</url>
</location>
<part>
<date>2022-12</date>
<extent unit="page">
<start>5073</start>
<end>5085</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Don‘t Just Clean It, Proxy Clean It: Mitigating Bias by Proxy in Pre-Trained Models
%A Panda, Swetasudha
%A Kobren, Ari
%A Wick, Michael
%A Shen, Qinlan
%Y Goldberg, Yoav
%Y Kozareva, Zornitsa
%Y Zhang, Yue
%S Findings of the Association for Computational Linguistics: EMNLP 2022
%D 2022
%8 December
%I Association for Computational Linguistics
%C Abu Dhabi, United Arab Emirates
%F panda-etal-2022-dont
%X Transformer-based pre-trained models are known to encode societal biases not only in their contextual representations, but also in downstream predictions when fine-tuned on task-specific data.We present D-Bias, an approach that selectively eliminates stereotypical associations (e.g, co-occurrence statistics) at fine-tuning, such that the model doesn‘t learn to excessively rely on those signals.D-Bias attenuates biases from both identity words and frequently co-occurring proxies, which we select using pointwise mutual information.We apply D-Bias to a) occupation classification, and b) toxicity classification and find that our approach substantially reduces downstream biases (e.g. by \ensuremath> 60% in toxicity classification, for identities that are most frequently flagged as toxic on online platforms).In addition, we show that D-Bias dramatically improves upon scrubbing, i.e., removing only the identity words in question.We also demonstrate that D-Bias easily extends to multiple identities, and achieves competitive performance with two recently proposed debiasing approaches: R-LACE and INLP.
%R 10.18653/v1/2022.findings-emnlp.372
%U https://aclanthology.org/2022.findings-emnlp.372/
%U https://doi.org/10.18653/v1/2022.findings-emnlp.372
%P 5073-5085
Markdown (Informal)
[Don’t Just Clean It, Proxy Clean It: Mitigating Bias by Proxy in Pre-Trained Models](https://aclanthology.org/2022.findings-emnlp.372/) (Panda et al., Findings 2022)
ACL