@inproceedings{najem-meyer-etal-2025-dont,
title = "Don{'}t stop pretraining! Efficiently building specialised language models in resource-constrained settings.",
author = "Najem-Meyer, Sven and
Kaplan, Fr{\'e}d{\'e}ric and
Romanello, Matteo",
editor = "Kazantseva, Anna and
Szpakowicz, Stan and
Degaetano-Ortlieb, Stefania and
Bizzoni, Yuri and
Pagel, Janis",
booktitle = "Proceedings of the 9th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL 2025)",
month = may,
year = "2025",
address = "Albuquerque, New Mexico",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.latechclfl-1.22/",
doi = "10.18653/v1/2025.latechclfl-1.22",
pages = "252--260",
ISBN = "979-8-89176-241-1",
abstract = "Developing specialised language models for low-resource domains typically involves a trade-off between two specialisation strategies: adapting a general-purpose model through continued pretraining or retraining a model from scratch. While adapting preserves the model{'}s linguistic knowledge, retraining benefits from the flexibility of an in-domain tokeniser {--} a potentially significant advantage when handling rare languages. This study investigates the impact of tokenisation, specialisation strategy, and pretraining data availability using classical scholarship {--} a multilingual, code-switching and highly domain-specific field {--} as a case study. Through extensive experiments, we assess whether domain-specific tokenisation improves model performance, whether character-based models provide a viable alternative to subword-based models, and which specialisation strategy is optimal given the constraints of limited pretraining data. Contrary to prior findings, our results show that in-domain tokenisation does not necessarily enhance performance. Most notably, adaptation consistently outperforms retraining, even with limited data, confirming its efficiency as the preferred strategy for resource-constrained domains. These insights provide valuable guidelines for developing specialised models in fields with limited textual resources."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="najem-meyer-etal-2025-dont">
<titleInfo>
<title>Don’t stop pretraining! Efficiently building specialised language models in resource-constrained settings.</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sven</namePart>
<namePart type="family">Najem-Meyer</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Frédéric</namePart>
<namePart type="family">Kaplan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Matteo</namePart>
<namePart type="family">Romanello</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 9th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL 2025)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Anna</namePart>
<namePart type="family">Kazantseva</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Stan</namePart>
<namePart type="family">Szpakowicz</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Stefania</namePart>
<namePart type="family">Degaetano-Ortlieb</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuri</namePart>
<namePart type="family">Bizzoni</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Janis</namePart>
<namePart type="family">Pagel</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Albuquerque, New Mexico</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-241-1</identifier>
</relatedItem>
<abstract>Developing specialised language models for low-resource domains typically involves a trade-off between two specialisation strategies: adapting a general-purpose model through continued pretraining or retraining a model from scratch. While adapting preserves the model’s linguistic knowledge, retraining benefits from the flexibility of an in-domain tokeniser – a potentially significant advantage when handling rare languages. This study investigates the impact of tokenisation, specialisation strategy, and pretraining data availability using classical scholarship – a multilingual, code-switching and highly domain-specific field – as a case study. Through extensive experiments, we assess whether domain-specific tokenisation improves model performance, whether character-based models provide a viable alternative to subword-based models, and which specialisation strategy is optimal given the constraints of limited pretraining data. Contrary to prior findings, our results show that in-domain tokenisation does not necessarily enhance performance. Most notably, adaptation consistently outperforms retraining, even with limited data, confirming its efficiency as the preferred strategy for resource-constrained domains. These insights provide valuable guidelines for developing specialised models in fields with limited textual resources.</abstract>
<identifier type="citekey">najem-meyer-etal-2025-dont</identifier>
<identifier type="doi">10.18653/v1/2025.latechclfl-1.22</identifier>
<location>
<url>https://aclanthology.org/2025.latechclfl-1.22/</url>
</location>
<part>
<date>2025-05</date>
<extent unit="page">
<start>252</start>
<end>260</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Don’t stop pretraining! Efficiently building specialised language models in resource-constrained settings.
%A Najem-Meyer, Sven
%A Kaplan, Frédéric
%A Romanello, Matteo
%Y Kazantseva, Anna
%Y Szpakowicz, Stan
%Y Degaetano-Ortlieb, Stefania
%Y Bizzoni, Yuri
%Y Pagel, Janis
%S Proceedings of the 9th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature (LaTeCH-CLfL 2025)
%D 2025
%8 May
%I Association for Computational Linguistics
%C Albuquerque, New Mexico
%@ 979-8-89176-241-1
%F najem-meyer-etal-2025-dont
%X Developing specialised language models for low-resource domains typically involves a trade-off between two specialisation strategies: adapting a general-purpose model through continued pretraining or retraining a model from scratch. While adapting preserves the model’s linguistic knowledge, retraining benefits from the flexibility of an in-domain tokeniser – a potentially significant advantage when handling rare languages. This study investigates the impact of tokenisation, specialisation strategy, and pretraining data availability using classical scholarship – a multilingual, code-switching and highly domain-specific field – as a case study. Through extensive experiments, we assess whether domain-specific tokenisation improves model performance, whether character-based models provide a viable alternative to subword-based models, and which specialisation strategy is optimal given the constraints of limited pretraining data. Contrary to prior findings, our results show that in-domain tokenisation does not necessarily enhance performance. Most notably, adaptation consistently outperforms retraining, even with limited data, confirming its efficiency as the preferred strategy for resource-constrained domains. These insights provide valuable guidelines for developing specialised models in fields with limited textual resources.
%R 10.18653/v1/2025.latechclfl-1.22
%U https://aclanthology.org/2025.latechclfl-1.22/
%U https://doi.org/10.18653/v1/2025.latechclfl-1.22
%P 252-260
Markdown (Informal)
[Don’t stop pretraining! Efficiently building specialised language models in resource-constrained settings.](https://aclanthology.org/2025.latechclfl-1.22/) (Najem-Meyer et al., LaTeCHCLfL 2025)
ACL