@inproceedings{kayal-etal-2024-far,
title = "How Far Is Too Far? Studying the Effects of Domain Discrepancy on Masked Language Models",
author = "Kayal, Subhradeep and
Rakhlin, Alexander and
Dashti, Ali and
Stepaniants, Serguei",
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.718",
pages = "8192--8199",
abstract = "Pre-trained masked language models, such as BERT, perform strongly on a wide variety of NLP tasks and have become ubiquitous in recent years. The typical way to use such models is to fine-tune them on downstream data. In this work, we aim to study how the difference in domains between the pre-trained model and the task effects its final performance. We first devise a simple mechanism to quantify the domain difference (using a cloze task) and use it to partition our dataset. Using these partitions of varying domain discrepancy, we focus on answering key questions around the impact of discrepancy on final performance, robustness to out-of-domain test-time examples and effect of domain-adaptive pre-training. We base our experiments on a large-scale openly available e-commerce dataset, and our findings suggest that in spite of pre-training the performance of BERT degrades on datasets with high domain discrepancy, especially in low resource cases. This effect is somewhat mitigated by continued pre-training for domain adaptation. Furthermore, the domain-gap also makes BERT sensitive to out-of-domain examples during inference, even in high resource tasks, and it is prudent to use as diverse a dataset as possible during fine-tuning to make it robust to domain shift.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kayal-etal-2024-far">
<titleInfo>
<title>How Far Is Too Far? Studying the Effects of Domain Discrepancy on Masked Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Subhradeep</namePart>
<namePart type="family">Kayal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alexander</namePart>
<namePart type="family">Rakhlin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ali</namePart>
<namePart type="family">Dashti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Serguei</namePart>
<namePart type="family">Stepaniants</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>Pre-trained masked language models, such as BERT, perform strongly on a wide variety of NLP tasks and have become ubiquitous in recent years. The typical way to use such models is to fine-tune them on downstream data. In this work, we aim to study how the difference in domains between the pre-trained model and the task effects its final performance. We first devise a simple mechanism to quantify the domain difference (using a cloze task) and use it to partition our dataset. Using these partitions of varying domain discrepancy, we focus on answering key questions around the impact of discrepancy on final performance, robustness to out-of-domain test-time examples and effect of domain-adaptive pre-training. We base our experiments on a large-scale openly available e-commerce dataset, and our findings suggest that in spite of pre-training the performance of BERT degrades on datasets with high domain discrepancy, especially in low resource cases. This effect is somewhat mitigated by continued pre-training for domain adaptation. Furthermore, the domain-gap also makes BERT sensitive to out-of-domain examples during inference, even in high resource tasks, and it is prudent to use as diverse a dataset as possible during fine-tuning to make it robust to domain shift.</abstract>
<identifier type="citekey">kayal-etal-2024-far</identifier>
<location>
<url>https://aclanthology.org/2024.lrec-main.718</url>
</location>
<part>
<date>2024-05</date>
<extent unit="page">
<start>8192</start>
<end>8199</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T How Far Is Too Far? Studying the Effects of Domain Discrepancy on Masked Language Models
%A Kayal, Subhradeep
%A Rakhlin, Alexander
%A Dashti, Ali
%A Stepaniants, Serguei
%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 kayal-etal-2024-far
%X Pre-trained masked language models, such as BERT, perform strongly on a wide variety of NLP tasks and have become ubiquitous in recent years. The typical way to use such models is to fine-tune them on downstream data. In this work, we aim to study how the difference in domains between the pre-trained model and the task effects its final performance. We first devise a simple mechanism to quantify the domain difference (using a cloze task) and use it to partition our dataset. Using these partitions of varying domain discrepancy, we focus on answering key questions around the impact of discrepancy on final performance, robustness to out-of-domain test-time examples and effect of domain-adaptive pre-training. We base our experiments on a large-scale openly available e-commerce dataset, and our findings suggest that in spite of pre-training the performance of BERT degrades on datasets with high domain discrepancy, especially in low resource cases. This effect is somewhat mitigated by continued pre-training for domain adaptation. Furthermore, the domain-gap also makes BERT sensitive to out-of-domain examples during inference, even in high resource tasks, and it is prudent to use as diverse a dataset as possible during fine-tuning to make it robust to domain shift.
%U https://aclanthology.org/2024.lrec-main.718
%P 8192-8199
Markdown (Informal)
[How Far Is Too Far? Studying the Effects of Domain Discrepancy on Masked Language Models](https://aclanthology.org/2024.lrec-main.718) (Kayal et al., LREC-COLING 2024)
ACL