@inproceedings{li-michael-2022-overconfidence,
title = "Overconfidence in the Face of Ambiguity with Adversarial Data",
author = "Li, Margaret and
Michael, Julian",
editor = "Bartolo, Max and
Kirk, Hannah and
Rodriguez, Pedro and
Margatina, Katerina and
Thrush, Tristan and
Jia, Robin and
Stenetorp, Pontus and
Williams, Adina and
Kiela, Douwe",
booktitle = "Proceedings of the First Workshop on Dynamic Adversarial Data Collection",
month = jul,
year = "2022",
address = "Seattle, WA",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2022.dadc-1.4",
doi = "10.18653/v1/2022.dadc-1.4",
pages = "30--40",
abstract = "Adversarial data collection has shown promise as a method for building models which are more robust to the spurious correlations that generally appear in naturalistic data. However, adversarially-collected data may itself be subject to biases, particularly with regard to ambiguous or arguable labeling judgments. Searching for examples where an annotator disagrees with a model might over-sample ambiguous inputs, and filtering the results for high inter-annotator agreement may under-sample them. In either case, training a model on such data may produce predictable and unwanted biases. In this work, we investigate whether models trained on adversarially-collected data are miscalibrated with respect to the ambiguity of their inputs. Using Natural Language Inference models as a testbed, we find no clear difference in accuracy between naturalistically and adversarially trained models, but our model trained only on adversarially-sourced data is considerably more overconfident of its predictions and demonstrates worse calibration, especially on ambiguous inputs. This effect is mitigated, however, when naturalistic and adversarial training data are combined.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="li-michael-2022-overconfidence">
<titleInfo>
<title>Overconfidence in the Face of Ambiguity with Adversarial Data</title>
</titleInfo>
<name type="personal">
<namePart type="given">Margaret</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Julian</namePart>
<namePart type="family">Michael</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2022-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the First Workshop on Dynamic Adversarial Data Collection</title>
</titleInfo>
<name type="personal">
<namePart type="given">Max</namePart>
<namePart type="family">Bartolo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hannah</namePart>
<namePart type="family">Kirk</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pedro</namePart>
<namePart type="family">Rodriguez</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Katerina</namePart>
<namePart type="family">Margatina</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tristan</namePart>
<namePart type="family">Thrush</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Robin</namePart>
<namePart type="family">Jia</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pontus</namePart>
<namePart type="family">Stenetorp</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Adina</namePart>
<namePart type="family">Williams</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Douwe</namePart>
<namePart type="family">Kiela</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Seattle, WA</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Adversarial data collection has shown promise as a method for building models which are more robust to the spurious correlations that generally appear in naturalistic data. However, adversarially-collected data may itself be subject to biases, particularly with regard to ambiguous or arguable labeling judgments. Searching for examples where an annotator disagrees with a model might over-sample ambiguous inputs, and filtering the results for high inter-annotator agreement may under-sample them. In either case, training a model on such data may produce predictable and unwanted biases. In this work, we investigate whether models trained on adversarially-collected data are miscalibrated with respect to the ambiguity of their inputs. Using Natural Language Inference models as a testbed, we find no clear difference in accuracy between naturalistically and adversarially trained models, but our model trained only on adversarially-sourced data is considerably more overconfident of its predictions and demonstrates worse calibration, especially on ambiguous inputs. This effect is mitigated, however, when naturalistic and adversarial training data are combined.</abstract>
<identifier type="citekey">li-michael-2022-overconfidence</identifier>
<identifier type="doi">10.18653/v1/2022.dadc-1.4</identifier>
<location>
<url>https://aclanthology.org/2022.dadc-1.4</url>
</location>
<part>
<date>2022-07</date>
<extent unit="page">
<start>30</start>
<end>40</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Overconfidence in the Face of Ambiguity with Adversarial Data
%A Li, Margaret
%A Michael, Julian
%Y Bartolo, Max
%Y Kirk, Hannah
%Y Rodriguez, Pedro
%Y Margatina, Katerina
%Y Thrush, Tristan
%Y Jia, Robin
%Y Stenetorp, Pontus
%Y Williams, Adina
%Y Kiela, Douwe
%S Proceedings of the First Workshop on Dynamic Adversarial Data Collection
%D 2022
%8 July
%I Association for Computational Linguistics
%C Seattle, WA
%F li-michael-2022-overconfidence
%X Adversarial data collection has shown promise as a method for building models which are more robust to the spurious correlations that generally appear in naturalistic data. However, adversarially-collected data may itself be subject to biases, particularly with regard to ambiguous or arguable labeling judgments. Searching for examples where an annotator disagrees with a model might over-sample ambiguous inputs, and filtering the results for high inter-annotator agreement may under-sample them. In either case, training a model on such data may produce predictable and unwanted biases. In this work, we investigate whether models trained on adversarially-collected data are miscalibrated with respect to the ambiguity of their inputs. Using Natural Language Inference models as a testbed, we find no clear difference in accuracy between naturalistically and adversarially trained models, but our model trained only on adversarially-sourced data is considerably more overconfident of its predictions and demonstrates worse calibration, especially on ambiguous inputs. This effect is mitigated, however, when naturalistic and adversarial training data are combined.
%R 10.18653/v1/2022.dadc-1.4
%U https://aclanthology.org/2022.dadc-1.4
%U https://doi.org/10.18653/v1/2022.dadc-1.4
%P 30-40
Markdown (Informal)
[Overconfidence in the Face of Ambiguity with Adversarial Data](https://aclanthology.org/2022.dadc-1.4) (Li & Michael, DADC 2022)
ACL