@inproceedings{alcoba-inciarte-etal-2024-utility,
title = "On the Utility of Pretraining Language Models on Synthetic Data",
author = "Alcoba Inciarte, Alcides and
Kwon, Sang Yun and
Nagoudi, El Moatez Billah and
Abdul-Mageed, Muhammad",
editor = "Habash, Nizar and
Bouamor, Houda and
Eskander, Ramy and
Tomeh, Nadi and
Abu Farha, Ibrahim and
Abdelali, Ahmed and
Touileb, Samia and
Hamed, Injy and
Onaizan, Yaser and
Alhafni, Bashar and
Antoun, Wissam and
Khalifa, Salam and
Haddad, Hatem and
Zitouni, Imed and
AlKhamissi, Badr and
Almatham, Rawan and
Mrini, Khalil",
booktitle = "Proceedings of The Second Arabic Natural Language Processing Conference",
month = aug,
year = "2024",
address = "Bangkok, Thailand",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.arabicnlp-1.23",
doi = "10.18653/v1/2024.arabicnlp-1.23",
pages = "265--282",
abstract = "Development of pre-trained language models has predominantly relied on large amounts of datasets. However, this dependence on abundant data has limited the applicability of these models in low-resource settings. In this work, we investigate the utility of exploiting synthetic datasets acquired from different sources to pre-train language models for Arabic. Namely, we leverage data derived based on four different methods: optical character recognition (OCR), automatic speech recognition (ASR), machine translation (MT), and generative language models. We use these datasets to pre-train models in three different architectures: encoder-only (BERTBase), encoder-decoder (T5), and decoder-only (GPT-2). We test the capabilities of resulting models on Arabic natural language understanding (NLU) tasks using the ORCA benchmark. Our results show that utilizing synthetic data can achieve performance comparable to, or even surpassing, those trained on gold data. For example, our model based on a GPT-2 architecture trained on a combined synthetic dataset surpasses the baseline model ARBERTv2. Overall, our models pre-trained on synthetic data demonstrate robust performance across various tasks. This highlights the potential of synthetic datasets in augmenting language model training in low-resource settings.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="alcoba-inciarte-etal-2024-utility">
<titleInfo>
<title>On the Utility of Pretraining Language Models on Synthetic Data</title>
</titleInfo>
<name type="personal">
<namePart type="given">Alcides</namePart>
<namePart type="family">Alcoba Inciarte</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sang</namePart>
<namePart type="given">Yun</namePart>
<namePart type="family">Kwon</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">El</namePart>
<namePart type="given">Moatez</namePart>
<namePart type="given">Billah</namePart>
<namePart type="family">Nagoudi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Muhammad</namePart>
<namePart type="family">Abdul-Mageed</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2024-08</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of The Second Arabic Natural Language Processing Conference</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nizar</namePart>
<namePart type="family">Habash</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<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">Ramy</namePart>
<namePart type="family">Eskander</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nadi</namePart>
<namePart type="family">Tomeh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ibrahim</namePart>
<namePart type="family">Abu Farha</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ahmed</namePart>
<namePart type="family">Abdelali</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Samia</namePart>
<namePart type="family">Touileb</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Injy</namePart>
<namePart type="family">Hamed</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yaser</namePart>
<namePart type="family">Onaizan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bashar</namePart>
<namePart type="family">Alhafni</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wissam</namePart>
<namePart type="family">Antoun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Salam</namePart>
<namePart type="family">Khalifa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hatem</namePart>
<namePart type="family">Haddad</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Imed</namePart>
<namePart type="family">Zitouni</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Badr</namePart>
<namePart type="family">AlKhamissi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rawan</namePart>
<namePart type="family">Almatham</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Khalil</namePart>
<namePart type="family">Mrini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Bangkok, Thailand</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Development of pre-trained language models has predominantly relied on large amounts of datasets. However, this dependence on abundant data has limited the applicability of these models in low-resource settings. In this work, we investigate the utility of exploiting synthetic datasets acquired from different sources to pre-train language models for Arabic. Namely, we leverage data derived based on four different methods: optical character recognition (OCR), automatic speech recognition (ASR), machine translation (MT), and generative language models. We use these datasets to pre-train models in three different architectures: encoder-only (BERTBase), encoder-decoder (T5), and decoder-only (GPT-2). We test the capabilities of resulting models on Arabic natural language understanding (NLU) tasks using the ORCA benchmark. Our results show that utilizing synthetic data can achieve performance comparable to, or even surpassing, those trained on gold data. For example, our model based on a GPT-2 architecture trained on a combined synthetic dataset surpasses the baseline model ARBERTv2. Overall, our models pre-trained on synthetic data demonstrate robust performance across various tasks. This highlights the potential of synthetic datasets in augmenting language model training in low-resource settings.</abstract>
<identifier type="citekey">alcoba-inciarte-etal-2024-utility</identifier>
<identifier type="doi">10.18653/v1/2024.arabicnlp-1.23</identifier>
<location>
<url>https://aclanthology.org/2024.arabicnlp-1.23</url>
</location>
<part>
<date>2024-08</date>
<extent unit="page">
<start>265</start>
<end>282</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T On the Utility of Pretraining Language Models on Synthetic Data
%A Alcoba Inciarte, Alcides
%A Kwon, Sang Yun
%A Nagoudi, El Moatez Billah
%A Abdul-Mageed, Muhammad
%Y Habash, Nizar
%Y Bouamor, Houda
%Y Eskander, Ramy
%Y Tomeh, Nadi
%Y Abu Farha, Ibrahim
%Y Abdelali, Ahmed
%Y Touileb, Samia
%Y Hamed, Injy
%Y Onaizan, Yaser
%Y Alhafni, Bashar
%Y Antoun, Wissam
%Y Khalifa, Salam
%Y Haddad, Hatem
%Y Zitouni, Imed
%Y AlKhamissi, Badr
%Y Almatham, Rawan
%Y Mrini, Khalil
%S Proceedings of The Second Arabic Natural Language Processing Conference
%D 2024
%8 August
%I Association for Computational Linguistics
%C Bangkok, Thailand
%F alcoba-inciarte-etal-2024-utility
%X Development of pre-trained language models has predominantly relied on large amounts of datasets. However, this dependence on abundant data has limited the applicability of these models in low-resource settings. In this work, we investigate the utility of exploiting synthetic datasets acquired from different sources to pre-train language models for Arabic. Namely, we leverage data derived based on four different methods: optical character recognition (OCR), automatic speech recognition (ASR), machine translation (MT), and generative language models. We use these datasets to pre-train models in three different architectures: encoder-only (BERTBase), encoder-decoder (T5), and decoder-only (GPT-2). We test the capabilities of resulting models on Arabic natural language understanding (NLU) tasks using the ORCA benchmark. Our results show that utilizing synthetic data can achieve performance comparable to, or even surpassing, those trained on gold data. For example, our model based on a GPT-2 architecture trained on a combined synthetic dataset surpasses the baseline model ARBERTv2. Overall, our models pre-trained on synthetic data demonstrate robust performance across various tasks. This highlights the potential of synthetic datasets in augmenting language model training in low-resource settings.
%R 10.18653/v1/2024.arabicnlp-1.23
%U https://aclanthology.org/2024.arabicnlp-1.23
%U https://doi.org/10.18653/v1/2024.arabicnlp-1.23
%P 265-282
Markdown (Informal)
[On the Utility of Pretraining Language Models on Synthetic Data](https://aclanthology.org/2024.arabicnlp-1.23) (Alcoba Inciarte et al., ArabicNLP-WS 2024)
ACL
- Alcides Alcoba Inciarte, Sang Yun Kwon, El Moatez Billah Nagoudi, and Muhammad Abdul-Mageed. 2024. On the Utility of Pretraining Language Models on Synthetic Data. In Proceedings of The Second Arabic Natural Language Processing Conference, pages 265–282, Bangkok, Thailand. Association for Computational Linguistics.