@inproceedings{he-etal-2024-mitigating,
title = "Mitigating Shortcuts in Language Models with Soft Label Encoding",
author = "He, Zirui and
Deng, Huiqi and
Zhao, Haiyan and
Liu, Ninghao and
Du, Mengnan",
editor = "Calzolari, Nicoletta and
Kan, Min-Yen and
Hoste, Veronique and
Lenci, Alessandro and
Sakti, Sakriani and
Xue, Nianwen",
booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
month = may,
year = "2024",
address = "Torino, Italia",
publisher = "ELRA and ICCL",
url = "https://aclanthology.org/2024.lrec-main.991/",
pages = "11341--11348",
abstract = "Recent research has shown that large language models rely on spurious correlations in the data for natural language understanding (NLU) tasks. In this work, we aim to answer the following research question: Can we reduce spurious correlations by modifying the ground truth labels of the training data? Specifically, we propose a simple yet effective debiasing framework, named Soft Label Encoding (SoftLE). First, we train a teacher model to quantify each sample`s degree of relying on shortcuts. Then, we encode this shortcut degree into a dummy class and use it to smooth the original ground truth labels, generating soft labels. These soft labels are used to train a more robust student model that reduces spurious correlations between shortcut features and certain classes. Extensive experiments on two NLU benchmark tasks via two language models demonstrate that SoftLE significantly improves out-of-distribution generalization while maintaining satisfactory in-distribution accuracy. Our code is available at https://github.com/ZiruiHE99/sle"
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="he-etal-2024-mitigating">
<titleInfo>
<title>Mitigating Shortcuts in Language Models with Soft Label Encoding</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zirui</namePart>
<namePart type="family">He</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Huiqi</namePart>
<namePart type="family">Deng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haiyan</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ninghao</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mengnan</namePart>
<namePart type="family">Du</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2024-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nicoletta</namePart>
<namePart type="family">Calzolari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Min-Yen</namePart>
<namePart type="family">Kan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Veronique</namePart>
<namePart type="family">Hoste</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alessandro</namePart>
<namePart type="family">Lenci</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sakriani</namePart>
<namePart type="family">Sakti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nianwen</namePart>
<namePart type="family">Xue</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>ELRA and ICCL</publisher>
<place>
<placeTerm type="text">Torino, Italia</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Recent research has shown that large language models rely on spurious correlations in the data for natural language understanding (NLU) tasks. In this work, we aim to answer the following research question: Can we reduce spurious correlations by modifying the ground truth labels of the training data? Specifically, we propose a simple yet effective debiasing framework, named Soft Label Encoding (SoftLE). First, we train a teacher model to quantify each sample‘s degree of relying on shortcuts. Then, we encode this shortcut degree into a dummy class and use it to smooth the original ground truth labels, generating soft labels. These soft labels are used to train a more robust student model that reduces spurious correlations between shortcut features and certain classes. Extensive experiments on two NLU benchmark tasks via two language models demonstrate that SoftLE significantly improves out-of-distribution generalization while maintaining satisfactory in-distribution accuracy. Our code is available at https://github.com/ZiruiHE99/sle</abstract>
<identifier type="citekey">he-etal-2024-mitigating</identifier>
<location>
<url>https://aclanthology.org/2024.lrec-main.991/</url>
</location>
<part>
<date>2024-05</date>
<extent unit="page">
<start>11341</start>
<end>11348</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Mitigating Shortcuts in Language Models with Soft Label Encoding
%A He, Zirui
%A Deng, Huiqi
%A Zhao, Haiyan
%A Liu, Ninghao
%A Du, Mengnan
%Y Calzolari, Nicoletta
%Y Kan, Min-Yen
%Y Hoste, Veronique
%Y Lenci, Alessandro
%Y Sakti, Sakriani
%Y Xue, Nianwen
%S Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)
%D 2024
%8 May
%I ELRA and ICCL
%C Torino, Italia
%F he-etal-2024-mitigating
%X Recent research has shown that large language models rely on spurious correlations in the data for natural language understanding (NLU) tasks. In this work, we aim to answer the following research question: Can we reduce spurious correlations by modifying the ground truth labels of the training data? Specifically, we propose a simple yet effective debiasing framework, named Soft Label Encoding (SoftLE). First, we train a teacher model to quantify each sample‘s degree of relying on shortcuts. Then, we encode this shortcut degree into a dummy class and use it to smooth the original ground truth labels, generating soft labels. These soft labels are used to train a more robust student model that reduces spurious correlations between shortcut features and certain classes. Extensive experiments on two NLU benchmark tasks via two language models demonstrate that SoftLE significantly improves out-of-distribution generalization while maintaining satisfactory in-distribution accuracy. Our code is available at https://github.com/ZiruiHE99/sle
%U https://aclanthology.org/2024.lrec-main.991/
%P 11341-11348
Markdown (Informal)
[Mitigating Shortcuts in Language Models with Soft Label Encoding](https://aclanthology.org/2024.lrec-main.991/) (He et al., LREC-COLING 2024)
ACL
- Zirui He, Huiqi Deng, Haiyan Zhao, Ninghao Liu, and Mengnan Du. 2024. Mitigating Shortcuts in Language Models with Soft Label Encoding. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 11341–11348, Torino, Italia. ELRA and ICCL.