@inproceedings{mikula-etal-2024-think,
title = "Think Twice: Measuring the Efficiency of Eliminating Prediction Shortcuts of Question Answering Models",
author = "Mikula, Luk{\'a}{\v{s}} and
{\v{S}}tef{\'a}nik, Michal and
Petrovi{\v{c}}, Marek and
Sojka, Petr",
editor = "Graham, Yvette and
Purver, Matthew",
booktitle = "Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = mar,
year = "2024",
address = "St. Julian{'}s, Malta",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.eacl-long.133",
pages = "2179--2193",
abstract = "While the Large Language Models (LLMs) dominate a majority of language understanding tasks, previous work shows that some of these results are supported by modelling spurious correlations of training datasets. Authors commonly assess model robustness by evaluating their models on out-of-distribution (OOD) datasets of the same task, but these datasets might share the bias of the training dataset. We propose a simple method for measuring a scale of models{'} reliance on any identified spurious feature and assess the robustness towards a large set of known and newly found prediction biases for various pre-trained models and debiasing methods in Question Answering (QA). We find that the reported OOD gains of debiasing methods can not be explained by mitigated reliance on biased features, suggesting that biases are shared among different QA datasets. We further evidence this by measuring that performance of OOD models depends on bias features comparably to the ID model. Our findings motivate future work to refine the reports of LLMs{'} robustness to a level of known spurious features.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="mikula-etal-2024-think">
<titleInfo>
<title>Think Twice: Measuring the Efficiency of Eliminating Prediction Shortcuts of Question Answering Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Lukáš</namePart>
<namePart type="family">Mikula</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Michal</namePart>
<namePart type="family">Štefánik</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marek</namePart>
<namePart type="family">Petrovič</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Petr</namePart>
<namePart type="family">Sojka</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2024-03</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yvette</namePart>
<namePart type="family">Graham</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Matthew</namePart>
<namePart type="family">Purver</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">St. Julian’s, Malta</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>While the Large Language Models (LLMs) dominate a majority of language understanding tasks, previous work shows that some of these results are supported by modelling spurious correlations of training datasets. Authors commonly assess model robustness by evaluating their models on out-of-distribution (OOD) datasets of the same task, but these datasets might share the bias of the training dataset. We propose a simple method for measuring a scale of models’ reliance on any identified spurious feature and assess the robustness towards a large set of known and newly found prediction biases for various pre-trained models and debiasing methods in Question Answering (QA). We find that the reported OOD gains of debiasing methods can not be explained by mitigated reliance on biased features, suggesting that biases are shared among different QA datasets. We further evidence this by measuring that performance of OOD models depends on bias features comparably to the ID model. Our findings motivate future work to refine the reports of LLMs’ robustness to a level of known spurious features.</abstract>
<identifier type="citekey">mikula-etal-2024-think</identifier>
<location>
<url>https://aclanthology.org/2024.eacl-long.133</url>
</location>
<part>
<date>2024-03</date>
<extent unit="page">
<start>2179</start>
<end>2193</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Think Twice: Measuring the Efficiency of Eliminating Prediction Shortcuts of Question Answering Models
%A Mikula, Lukáš
%A Štefánik, Michal
%A Petrovič, Marek
%A Sojka, Petr
%Y Graham, Yvette
%Y Purver, Matthew
%S Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2024
%8 March
%I Association for Computational Linguistics
%C St. Julian’s, Malta
%F mikula-etal-2024-think
%X While the Large Language Models (LLMs) dominate a majority of language understanding tasks, previous work shows that some of these results are supported by modelling spurious correlations of training datasets. Authors commonly assess model robustness by evaluating their models on out-of-distribution (OOD) datasets of the same task, but these datasets might share the bias of the training dataset. We propose a simple method for measuring a scale of models’ reliance on any identified spurious feature and assess the robustness towards a large set of known and newly found prediction biases for various pre-trained models and debiasing methods in Question Answering (QA). We find that the reported OOD gains of debiasing methods can not be explained by mitigated reliance on biased features, suggesting that biases are shared among different QA datasets. We further evidence this by measuring that performance of OOD models depends on bias features comparably to the ID model. Our findings motivate future work to refine the reports of LLMs’ robustness to a level of known spurious features.
%U https://aclanthology.org/2024.eacl-long.133
%P 2179-2193
Markdown (Informal)
[Think Twice: Measuring the Efficiency of Eliminating Prediction Shortcuts of Question Answering Models](https://aclanthology.org/2024.eacl-long.133) (Mikula et al., EACL 2024)
ACL