Think Twice: Measuring the Efficiency of Eliminating Prediction
Shortcuts of Question Answering Models

D 2024

Think Twice: Measuring the Efficiency of Eliminating Prediction Shortcuts of Question Answering Models

MIKULA, Lukáš; Michal ŠTEFÁNIK; Marek PETROVIČ a Petr SOJKA

Základní údaje

Originální název

Think Twice: Measuring the Efficiency of Eliminating Prediction Shortcuts of Question Answering Models

Autoři

MIKULA, Lukáš; Michal ŠTEFÁNIK; Marek PETROVIČ a Petr SOJKA

Vydání

St. Julian's, Malta, Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), od s. 2179-2193, 15 s. 2024

Nakladatel

Association for Computational Linguistics

Další údaje

Jazyk

angličtina

Typ výsledku

Stať ve sborníku

Obor

10201 Computer sciences, information science, bioinformatics

Stát vydavatele

Spojené státy

Utajení

není předmětem státního či obchodního tajemství

Forma vydání

tištěná verze "print"

Odkazy

Openreview, Pre-print, fulltext PDF

Kód RIV

RIV/00216224:14330/24:00135399

Organizační jednotka

Fakulta informatiky

ISBN

979-8-89176-088-2

UT WoS

001356732602016

EID Scopus

2-s2.0-85189943414

Klíčová slova anglicky

LLM; Large Language Models; bias; Question Answering;

Štítky

core_A, firank_A

Příznaky

Mezinárodní význam, Recenzováno

Změněno: 30. 7. 2025 15:03, Mgr. Michal Petr

Anotace

V originále

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 QA datasets. We further evidence this by measuring that performance of OOD models depends on bias features comparably to the ID model, motivating future work to refine the reports of LLMs' robustness to a level of known spurious features.

Návaznosti

MUNI/A/1590/2023, interní kód MU

Název: Využití technik umělé inteligence pro zpracování dat, komplexní analýzy a vizualizaci rozsáhlých dat

Investor: Masarykova univerzita, Využití technik umělé inteligence pro zpracování dat, komplexní analýzy a vizualizaci rozsáhlých dat

MUNI/A/1608/2023, interní kód MU

Název: Zapojení studentů Fakulty informatiky do mezinárodní vědecké komunity 24

Investor: Masarykova univerzita, Zapojení studentů Fakulty informatiky do mezinárodní vědecké komunity 24

Citovat

MIKULA, Lukáš; Michal ŠTEFÁNIK; Marek PETROVIČ a Petr SOJKA. Think Twice: Measuring the Efficiency of Eliminating Prediction Shortcuts of Question Answering Models. In Yvette Graham, Matthew Purver. Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). St. Julian's, Malta: Association for Computational Linguistics, 2024, s. 2179-2193. ISBN 979-8-89176-088-2.

@inproceedings{2365639,
   author = {Mikula, Lukáš and Štefánik, Michal and Petrovič, Marek and Sojka, Petr},
   address = {St. Julian's, Malta},
   booktitle = {Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)},
   editor = {Yvette Graham, Matthew Purver},
   keywords = {LLM; Large Language Models; bias; Question Answering;},
   howpublished = {tištěná verze "print"},
   language = {eng},
   location = {St. Julian's, Malta},
   isbn = {979-8-89176-088-2},
   pages = {2179-2193},
   publisher = {Association for Computational Linguistics},
   title = {Think Twice: Measuring the Efficiency of Eliminating Prediction Shortcuts of Question Answering Models},
   url = {https://arxiv.org/abs/2305.06841},
   year = {2024}
}

TY  - CONF
ID  - 2365639
AU  - Mikula, Lukáš - Štefánik, Michal - Petrovič, Marek - Sojka, Petr
PY  - 2024
TI  - Think Twice: Measuring the Efficiency of Eliminating Prediction Shortcuts of Question Answering Models
PB  - Association for Computational Linguistics
CY  - St. Julian's, Malta
SN  - 9798891760882
KW  - LLM
KW  - Large Language Models
KW  - bias
KW  - Question Answering;
UR  - https://arxiv.org/abs/2305.06841
N2  - 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 QA datasets. We further evidence this by measuring that performance of OOD models depends on bias features comparably to the ID model, motivating future work to refine the reports of LLMs' robustness to a level of known spurious features.
ER  -

MIKULA, Lukáš; Michal ŠTEFÁNIK; Marek PETROVIČ a Petr SOJKA. Think Twice: Measuring the Efficiency of Eliminating Prediction Shortcuts of Question Answering Models. In Yvette Graham, Matthew Purver. \textit{Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)}. St. Julian's, Malta: Association for Computational Linguistics, 2024, s.~2179-2193. ISBN~979-8-89176-088-2.

Přehled o publikaci