<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="shao-etal-2023-erasure">
    <titleInfo>
        <title>Erasure of Unaligned Attributes from Neural Representations</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Shun</namePart>
        <namePart type="family">Shao</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yftah</namePart>
        <namePart type="family">Ziser</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Shay</namePart>
        <namePart type="given">B</namePart>
        <namePart type="family">Cohen</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2023</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <genre authority="bibutilsgt">journal article</genre>
    <relatedItem type="host">
        <titleInfo>
            <title>Transactions of the Association for Computational Linguistics</title>
        </titleInfo>
        <originInfo>
            <issuance>continuing</issuance>
            <publisher>MIT Press</publisher>
            <place>
                <placeTerm type="text">Cambridge, MA</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">periodical</genre>
        <genre authority="bibutilsgt">academic journal</genre>
    </relatedItem>
    <abstract>We present the Assignment-Maximization Spectral Attribute removaL (AMSAL) algorithm, which erases information from neural representations when the information to be erased is implicit rather than directly being aligned to each input example. Our algorithm works by alternating between two steps. In one, it finds an assignment of the input representations to the information to be erased, and in the other, it creates projections of both the input representations and the information to be erased into a joint latent space. We test our algorithm on an extensive array of datasets, including a Twitter dataset with multiple guarded attributes, the BiasBios dataset, and the BiasBench benchmark. The latter benchmark includes four datasets with various types of protected attributes. Our results demonstrate that bias can often be removed in our setup. We also discuss the limitations of our approach when there is a strong entanglement between the main task and the information to be erased.1</abstract>
    <identifier type="citekey">shao-etal-2023-erasure</identifier>
    <identifier type="doi">10.1162/tacl_a_00558</identifier>
    <location>
        <url>https://aclanthology.org/2023.tacl-1.29</url>
    </location>
    <part>
        <date>2023</date>
        <detail type="volume"><number>11</number></detail>
        <extent unit="page">
            <start>488</start>
            <end>510</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Journal Article
%T Erasure of Unaligned Attributes from Neural Representations
%A Shao, Shun
%A Ziser, Yftah
%A Cohen, Shay B.
%J Transactions of the Association for Computational Linguistics
%D 2023
%V 11
%I MIT Press
%C Cambridge, MA
%F shao-etal-2023-erasure
%X We present the Assignment-Maximization Spectral Attribute removaL (AMSAL) algorithm, which erases information from neural representations when the information to be erased is implicit rather than directly being aligned to each input example. Our algorithm works by alternating between two steps. In one, it finds an assignment of the input representations to the information to be erased, and in the other, it creates projections of both the input representations and the information to be erased into a joint latent space. We test our algorithm on an extensive array of datasets, including a Twitter dataset with multiple guarded attributes, the BiasBios dataset, and the BiasBench benchmark. The latter benchmark includes four datasets with various types of protected attributes. Our results demonstrate that bias can often be removed in our setup. We also discuss the limitations of our approach when there is a strong entanglement between the main task and the information to be erased.1
%R 10.1162/tacl_a_00558
%U https://aclanthology.org/2023.tacl-1.29
%U https://doi.org/10.1162/tacl_a_00558
%P 488-510

Markdown (Informal)

[Erasure of Unaligned Attributes from Neural Representations](https://aclanthology.org/2023.tacl-1.29) (Shao et al., TACL 2023)

• Erasure of Unaligned Attributes from Neural Representations (Shao et al., TACL 2023)

ACL

• Shun Shao, Yftah Ziser, and Shay B. Cohen. 2023. Erasure of Unaligned Attributes from Neural Representations. Transactions of the Association for Computational Linguistics, 11:488–510.