@inproceedings{ueda-etal-2024-token-length,
title = "Token-length Bias in Minimal-pair Paradigm Datasets",
author = "Ueda, Naoya and
Mita, Masato and
Oka, Teruaki and
Komachi, Mamoru",
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.1410",
pages = "16224--16236",
abstract = "Minimal-pair paradigm datasets have been used as benchmarks to evaluate the linguistic knowledge of models and provide an unsupervised method of acceptability judgment. The model performances are evaluated based on the percentage of minimal pairs in the MPP dataset where the model assigns a higher sentence log-likelihood to an acceptable sentence than to an unacceptable sentence. Each minimal pair in MPP datasets is controlled to align the number of words per sentence because the sentence length affects the sentence log-likelihood. However, aligning the number of words may be insufficient because recent language models tokenize sentences with subwords. Tokenization may cause a token length difference in minimal pairs, introducing token-length bias that skews the evaluation results. This study demonstrates that MPP datasets suffer from token-length bias and fail to evaluate the linguistic knowledge of a language model correctly. The results proved that sentences with a shorter token length would likely be assigned a higher log-likelihood regardless of their acceptability, which becomes problematic when comparing models with different tokenizers. To address this issue, we propose a debiased minimal pair generation method, allowing MPP datasets to measure language ability correctly and provide comparable results for all models.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ueda-etal-2024-token-length">
<titleInfo>
<title>Token-length Bias in Minimal-pair Paradigm Datasets</title>
</titleInfo>
<name type="personal">
<namePart type="given">Naoya</namePart>
<namePart type="family">Ueda</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Masato</namePart>
<namePart type="family">Mita</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Teruaki</namePart>
<namePart type="family">Oka</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mamoru</namePart>
<namePart type="family">Komachi</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>Minimal-pair paradigm datasets have been used as benchmarks to evaluate the linguistic knowledge of models and provide an unsupervised method of acceptability judgment. The model performances are evaluated based on the percentage of minimal pairs in the MPP dataset where the model assigns a higher sentence log-likelihood to an acceptable sentence than to an unacceptable sentence. Each minimal pair in MPP datasets is controlled to align the number of words per sentence because the sentence length affects the sentence log-likelihood. However, aligning the number of words may be insufficient because recent language models tokenize sentences with subwords. Tokenization may cause a token length difference in minimal pairs, introducing token-length bias that skews the evaluation results. This study demonstrates that MPP datasets suffer from token-length bias and fail to evaluate the linguistic knowledge of a language model correctly. The results proved that sentences with a shorter token length would likely be assigned a higher log-likelihood regardless of their acceptability, which becomes problematic when comparing models with different tokenizers. To address this issue, we propose a debiased minimal pair generation method, allowing MPP datasets to measure language ability correctly and provide comparable results for all models.</abstract>
<identifier type="citekey">ueda-etal-2024-token-length</identifier>
<location>
<url>https://aclanthology.org/2024.lrec-main.1410</url>
</location>
<part>
<date>2024-05</date>
<extent unit="page">
<start>16224</start>
<end>16236</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Token-length Bias in Minimal-pair Paradigm Datasets
%A Ueda, Naoya
%A Mita, Masato
%A Oka, Teruaki
%A Komachi, Mamoru
%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 ueda-etal-2024-token-length
%X Minimal-pair paradigm datasets have been used as benchmarks to evaluate the linguistic knowledge of models and provide an unsupervised method of acceptability judgment. The model performances are evaluated based on the percentage of minimal pairs in the MPP dataset where the model assigns a higher sentence log-likelihood to an acceptable sentence than to an unacceptable sentence. Each minimal pair in MPP datasets is controlled to align the number of words per sentence because the sentence length affects the sentence log-likelihood. However, aligning the number of words may be insufficient because recent language models tokenize sentences with subwords. Tokenization may cause a token length difference in minimal pairs, introducing token-length bias that skews the evaluation results. This study demonstrates that MPP datasets suffer from token-length bias and fail to evaluate the linguistic knowledge of a language model correctly. The results proved that sentences with a shorter token length would likely be assigned a higher log-likelihood regardless of their acceptability, which becomes problematic when comparing models with different tokenizers. To address this issue, we propose a debiased minimal pair generation method, allowing MPP datasets to measure language ability correctly and provide comparable results for all models.
%U https://aclanthology.org/2024.lrec-main.1410
%P 16224-16236
Markdown (Informal)
[Token-length Bias in Minimal-pair Paradigm Datasets](https://aclanthology.org/2024.lrec-main.1410) (Ueda et al., LREC-COLING 2024)
ACL
- Naoya Ueda, Masato Mita, Teruaki Oka, and Mamoru Komachi. 2024. Token-length Bias in Minimal-pair Paradigm Datasets. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 16224–16236, Torino, Italia. ELRA and ICCL.