@inproceedings{elangovan-etal-2023-effects,
title = "Effects of Human Adversarial and Affable Samples on {BERT} Generalization",
author = "Elangovan, Aparna and
He, Estrid and
Li, Yuan and
Verspoor, Karin",
editor = "Bouamor, Houda and
Pino, Juan and
Bali, Kalika",
booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2023",
month = dec,
year = "2023",
address = "Singapore",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2023.findings-emnlp.512",
doi = "10.18653/v1/2023.findings-emnlp.512",
pages = "7637--7649",
abstract = "BERT-based models have had strong performance on leaderboards, yet have been demonstrably worse in real-world settings requiring generalization. Limited quantities of training data is considered a key impediment to achieving generalizability in machine learning. In this paper, we examine the impact of training data quality, not quantity, on a model{'}s generalizability. We consider two characteristics of training data: the portion of human-adversarial (h-adversarial), i.e. sample pairs with seemingly minor differences but different ground-truth labels, and human-affable (h-affable) training samples, i.e. sample pairs with minor differences but the same ground-truth label. We find that for a fixed size of training samples, as a rule of thumb, having 10-30{\%} h-adversarial instances improves the precision, and therefore F1, by up to 20 points in the tasks of text classification and relation extraction. Increasing h-adversarials beyond this range can result in performance plateaus or even degradation. In contrast, h-affables may not contribute to a model{'}s generalizability and may even degrade generalization performance.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="elangovan-etal-2023-effects">
<titleInfo>
<title>Effects of Human Adversarial and Affable Samples on BERT Generalization</title>
</titleInfo>
<name type="personal">
<namePart type="given">Aparna</namePart>
<namePart type="family">Elangovan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Estrid</namePart>
<namePart type="family">He</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuan</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Karin</namePart>
<namePart type="family">Verspoor</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2023-12</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2023</title>
</titleInfo>
<name type="personal">
<namePart type="given">Houda</namePart>
<namePart type="family">Bouamor</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Juan</namePart>
<namePart type="family">Pino</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kalika</namePart>
<namePart type="family">Bali</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Singapore</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>BERT-based models have had strong performance on leaderboards, yet have been demonstrably worse in real-world settings requiring generalization. Limited quantities of training data is considered a key impediment to achieving generalizability in machine learning. In this paper, we examine the impact of training data quality, not quantity, on a model’s generalizability. We consider two characteristics of training data: the portion of human-adversarial (h-adversarial), i.e. sample pairs with seemingly minor differences but different ground-truth labels, and human-affable (h-affable) training samples, i.e. sample pairs with minor differences but the same ground-truth label. We find that for a fixed size of training samples, as a rule of thumb, having 10-30% h-adversarial instances improves the precision, and therefore F1, by up to 20 points in the tasks of text classification and relation extraction. Increasing h-adversarials beyond this range can result in performance plateaus or even degradation. In contrast, h-affables may not contribute to a model’s generalizability and may even degrade generalization performance.</abstract>
<identifier type="citekey">elangovan-etal-2023-effects</identifier>
<identifier type="doi">10.18653/v1/2023.findings-emnlp.512</identifier>
<location>
<url>https://aclanthology.org/2023.findings-emnlp.512</url>
</location>
<part>
<date>2023-12</date>
<extent unit="page">
<start>7637</start>
<end>7649</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Effects of Human Adversarial and Affable Samples on BERT Generalization
%A Elangovan, Aparna
%A He, Estrid
%A Li, Yuan
%A Verspoor, Karin
%Y Bouamor, Houda
%Y Pino, Juan
%Y Bali, Kalika
%S Findings of the Association for Computational Linguistics: EMNLP 2023
%D 2023
%8 December
%I Association for Computational Linguistics
%C Singapore
%F elangovan-etal-2023-effects
%X BERT-based models have had strong performance on leaderboards, yet have been demonstrably worse in real-world settings requiring generalization. Limited quantities of training data is considered a key impediment to achieving generalizability in machine learning. In this paper, we examine the impact of training data quality, not quantity, on a model’s generalizability. We consider two characteristics of training data: the portion of human-adversarial (h-adversarial), i.e. sample pairs with seemingly minor differences but different ground-truth labels, and human-affable (h-affable) training samples, i.e. sample pairs with minor differences but the same ground-truth label. We find that for a fixed size of training samples, as a rule of thumb, having 10-30% h-adversarial instances improves the precision, and therefore F1, by up to 20 points in the tasks of text classification and relation extraction. Increasing h-adversarials beyond this range can result in performance plateaus or even degradation. In contrast, h-affables may not contribute to a model’s generalizability and may even degrade generalization performance.
%R 10.18653/v1/2023.findings-emnlp.512
%U https://aclanthology.org/2023.findings-emnlp.512
%U https://doi.org/10.18653/v1/2023.findings-emnlp.512
%P 7637-7649
Markdown (Informal)
[Effects of Human Adversarial and Affable Samples on BERT Generalization](https://aclanthology.org/2023.findings-emnlp.512) (Elangovan et al., Findings 2023)
ACL