<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="honda-etal-2024-eliminate">
    <titleInfo>
        <title>Not Eliminate but Aggregate: Post-Hoc Control over Mixture-of-Experts to Address Shortcut Shifts in Natural Language Understanding</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Ukyo</namePart>
        <namePart type="family">Honda</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Tatsushi</namePart>
        <namePart type="family">Oka</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Peinan</namePart>
        <namePart type="family">Zhang</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Masato</namePart>
        <namePart type="family">Mita</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2024</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>Recent models for natural language understanding are inclined to exploit simple patterns in datasets, commonly known as shortcuts. These shortcuts hinge on spurious correlations between labels and latent features existing in the training data. At inference time, shortcut-dependent models are likely to generate erroneous predictions under distribution shifts, particularly when some latent features are no longer correlated with the labels. To avoid this, previous studies have trained models to eliminate the reliance on shortcuts. In this study, we explore a different direction: pessimistically aggregating the predictions of a mixture-of-experts, assuming each expert captures relatively different latent features. The experimental results demonstrate that our post-hoc control over the experts significantly enhances the model’s robustness to the distribution shift in shortcuts. Additionally, we show that our approach has some practical advantages. We also analyze our model and provide results to support the assumption.1</abstract>
    <identifier type="citekey">honda-etal-2024-eliminate</identifier>
    <identifier type="doi">10.1162/tacl_a_00701</identifier>
    <location>
        <url>https://aclanthology.org/2024.tacl-1.70</url>
    </location>
    <part>
        <date>2024</date>
        <detail type="volume"><number>12</number></detail>
        <extent unit="page">
            <start>1268</start>
            <end>1289</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Journal Article
%T Not Eliminate but Aggregate: Post-Hoc Control over Mixture-of-Experts to Address Shortcut Shifts in Natural Language Understanding
%A Honda, Ukyo
%A Oka, Tatsushi
%A Zhang, Peinan
%A Mita, Masato
%J Transactions of the Association for Computational Linguistics
%D 2024
%V 12
%I MIT Press
%C Cambridge, MA
%F honda-etal-2024-eliminate
%X Recent models for natural language understanding are inclined to exploit simple patterns in datasets, commonly known as shortcuts. These shortcuts hinge on spurious correlations between labels and latent features existing in the training data. At inference time, shortcut-dependent models are likely to generate erroneous predictions under distribution shifts, particularly when some latent features are no longer correlated with the labels. To avoid this, previous studies have trained models to eliminate the reliance on shortcuts. In this study, we explore a different direction: pessimistically aggregating the predictions of a mixture-of-experts, assuming each expert captures relatively different latent features. The experimental results demonstrate that our post-hoc control over the experts significantly enhances the model’s robustness to the distribution shift in shortcuts. Additionally, we show that our approach has some practical advantages. We also analyze our model and provide results to support the assumption.1
%R 10.1162/tacl_a_00701
%U https://aclanthology.org/2024.tacl-1.70
%U https://doi.org/10.1162/tacl_a_00701
%P 1268-1289

Markdown (Informal)

[Not Eliminate but Aggregate: Post-Hoc Control over Mixture-of-Experts to Address Shortcut Shifts in Natural Language Understanding](https://aclanthology.org/2024.tacl-1.70) (Honda et al., TACL 2024)

• Not Eliminate but Aggregate: Post-Hoc Control over Mixture-of-Experts to Address Shortcut Shifts in Natural Language Understanding (Honda et al., TACL 2024)

ACL

• Ukyo Honda, Tatsushi Oka, Peinan Zhang, and Masato Mita. 2024. Not Eliminate but Aggregate: Post-Hoc Control over Mixture-of-Experts to Address Shortcut Shifts in Natural Language Understanding. Transactions of the Association for Computational Linguistics, 12:1268–1289.