MorSeD: Morphological Segmentation of Danish and its Effect on Language Modeling

Rob van der Goot; Anette Jensen; Emil Allerslev Schledermann; Mikkel Wildner Kildeberg; Nicolaj Larsen; Mike Zhang; Elisa Bassignana

MorSeD: Morphological Segmentation of Danish and its Effect on Language Modeling

Rob van der Goot
, Anette Jensen
, Emil Allerslev Schledermann
, Mikkel Wildner Kildeberg
, Nicolaj Larsen
, Mike Zhang
, Elisa Bassignana

Publikation: Bidrag til bog/antologi/rapport/konference proceeding › Konferenceartikel i proceeding › Forskning › peer review

8 Downloads (Pure)

Abstract

Current language models (LMs) mostly exploit subwords as input units based on statistical co-occurrences of characters. Adjacently, previous work has shown that modeling morphemes can aid performance for Natural Language Processing (NLP) models. However, morphemes are challenging to obtain as there is no annotated data in most languages. In this work, we release a wide-coverage Danish morphological segmentation evaluation set. We evaluate a wide variety of unsupervised token segmenters and evaluate the downstream effect of using morphemes as input units for transformer-based LMs. Our results show that popular subword algorithms perform poorly on this task, scoring at most an $F_1$ of 44.8 compared to 58.6 for an unsupervised morphological segmented (Morfessor). Furthermore, we confirm that the more closely our tokenizer resembles morphemes, the higher the performance of LMs.

Originalsprog	Engelsk
Titel	Proceedings of the 25th Nordic Conference on Computational Lingustics and 11th Baltic Conference on Human Language Technologies : NoDaLiDa/Baltic-HLT 2025
Redaktører	Richard Johansson, Sara Stymne
Forlag	Association for Computational Linguistics (ACL)
Publikationsdato	mar. 2025
Sider	223-229
ISBN (Elektronisk)	978-9908-53-109-0
Status	Udgivet - mar. 2025
Begivenhed	Proceedings of the Joint 25th Nordic Conference on Computational Linguistics and 11th Baltic Conference on Human Language Technologies - Tartu, Estland Varighed: 3 mar. 2025 → 4 mar. 2025

Konference

Konference	Proceedings of the Joint 25th Nordic Conference on Computational Linguistics and 11th Baltic Conference on Human Language Technologies
Land/Område	Estland
By	Tartu
Periode	03/03/2025 → 04/03/2025

Adgang til dokumentet

pdfForlagets udgivne version, 179 KBLicens: CC BY 4.0

https://aclanthology.org/2025.nodalida-1.23/Licens: CC BY 4.0

AUB Link

Søg efter materialet i Aalborg Universitetsbiblioteks søgemaskine

1 Igangværende

Digital Twins for Abundant Feedback: Novel Feedback Paradigms via Explainable Multilingual Natural Language Processing
Bjerva, J. (PI (principal investigator)), Lindsay, E. (PI (principal investigator)) & Zhang, M. (Projektdeltager)
Villum Fonden
01/01/2024 → 31/12/2026
Projekter: Projekt › Forskning

Citationsformater

van der Goot, R., Jensen, A., Allerslev Schledermann, E., Wildner Kildeberg, M., Larsen, N., Zhang, M., & Bassignana, E. (2025). MorSeD: Morphological Segmentation of Danish and its Effect on Language Modeling. I R. Johansson, & S. Stymne (red.), Proceedings of the 25th Nordic Conference on Computational Lingustics and 11th Baltic Conference on Human Language Technologies : NoDaLiDa/Baltic-HLT 2025 (s. 223-229). Association for Computational Linguistics (ACL). https://aclanthology.org/2025.nodalida-1.23/

van der Goot, Rob ; Jensen, Anette ; Allerslev Schledermann, Emil et al. / MorSeD : Morphological Segmentation of Danish and its Effect on Language Modeling. Proceedings of the 25th Nordic Conference on Computational Lingustics and 11th Baltic Conference on Human Language Technologies : NoDaLiDa/Baltic-HLT 2025. red. / Richard Johansson ; Sara Stymne. Association for Computational Linguistics (ACL), 2025. s. 223-229

@inproceedings{cdd51d4a01534cd6b5ca2c7f2c97e306,

title = "MorSeD: Morphological Segmentation of Danish and its Effect on Language Modeling",

abstract = "Current language models (LMs) mostly exploit subwords as input units based on statistical co-occurrences of characters. Adjacently, previous work has shown that modeling morphemes can aid performance for Natural Language Processing (NLP) models. However, morphemes are challenging to obtain as there is no annotated data in most languages. In this work, we release a wide-coverage Danish morphological segmentation evaluation set. We evaluate a wide variety of unsupervised token segmenters and evaluate the downstream effect of using morphemes as input units for transformer-based LMs. Our results show that popular subword algorithms perform poorly on this task, scoring at most an \$F\_1\$ of 44.8 compared to 58.6 for an unsupervised morphological segmented (Morfessor). Furthermore, we confirm that the more closely our tokenizer resembles morphemes, the higher the performance of LMs.",

author = "\{van der Goot\}, Rob and Anette Jensen and \{Allerslev Schledermann\}, Emil and \{Wildner Kildeberg\}, Mikkel and Nicolaj Larsen and Mike Zhang and Elisa Bassignana",

year = "2025",

month = mar,

language = "English",

pages = "223--229",

editor = "Richard Johansson and Sara Stymne",

booktitle = "Proceedings of the 25th Nordic Conference on Computational Lingustics and 11th Baltic Conference on Human Language Technologies : NoDaLiDa/Baltic-HLT 2025",

publisher = "Association for Computational Linguistics (ACL)",

address = "United States",

note = "Proceedings of the Joint 25th Nordic Conference on Computational Linguistics and 11th Baltic Conference on Human Language Technologies ; Conference date: 03-03-2025 Through 04-03-2025",

}

van der Goot, R, Jensen, A, Allerslev Schledermann, E, Wildner Kildeberg, M, Larsen, N, Zhang, M & Bassignana, E 2025, MorSeD: Morphological Segmentation of Danish and its Effect on Language Modeling. i R Johansson & S Stymne (red), Proceedings of the 25th Nordic Conference on Computational Lingustics and 11th Baltic Conference on Human Language Technologies : NoDaLiDa/Baltic-HLT 2025. Association for Computational Linguistics (ACL), s. 223-229, Proceedings of the Joint 25th Nordic Conference on Computational Linguistics and 11th Baltic Conference on Human Language Technologies, Tartu, Estland, 03/03/2025. <https://aclanthology.org/2025.nodalida-1.23/>

MorSeD: Morphological Segmentation of Danish and its Effect on Language Modeling. / van der Goot, Rob; Jensen, Anette; Allerslev Schledermann, Emil et al.
Proceedings of the 25th Nordic Conference on Computational Lingustics and 11th Baltic Conference on Human Language Technologies : NoDaLiDa/Baltic-HLT 2025. red. / Richard Johansson; Sara Stymne. Association for Computational Linguistics (ACL), 2025. s. 223-229.

Publikation: Bidrag til bog/antologi/rapport/konference proceeding › Konferenceartikel i proceeding › Forskning › peer review

TY - GEN

T1 - MorSeD

T2 - Proceedings of the Joint 25th Nordic Conference on Computational Linguistics and 11th Baltic Conference on Human Language Technologies

AU - van der Goot, Rob

AU - Jensen, Anette

AU - Allerslev Schledermann, Emil

AU - Wildner Kildeberg, Mikkel

AU - Larsen, Nicolaj

AU - Zhang, Mike

AU - Bassignana, Elisa

PY - 2025/3

Y1 - 2025/3

N2 - Current language models (LMs) mostly exploit subwords as input units based on statistical co-occurrences of characters. Adjacently, previous work has shown that modeling morphemes can aid performance for Natural Language Processing (NLP) models. However, morphemes are challenging to obtain as there is no annotated data in most languages. In this work, we release a wide-coverage Danish morphological segmentation evaluation set. We evaluate a wide variety of unsupervised token segmenters and evaluate the downstream effect of using morphemes as input units for transformer-based LMs. Our results show that popular subword algorithms perform poorly on this task, scoring at most an $F_1$ of 44.8 compared to 58.6 for an unsupervised morphological segmented (Morfessor). Furthermore, we confirm that the more closely our tokenizer resembles morphemes, the higher the performance of LMs.

AB - Current language models (LMs) mostly exploit subwords as input units based on statistical co-occurrences of characters. Adjacently, previous work has shown that modeling morphemes can aid performance for Natural Language Processing (NLP) models. However, morphemes are challenging to obtain as there is no annotated data in most languages. In this work, we release a wide-coverage Danish morphological segmentation evaluation set. We evaluate a wide variety of unsupervised token segmenters and evaluate the downstream effect of using morphemes as input units for transformer-based LMs. Our results show that popular subword algorithms perform poorly on this task, scoring at most an $F_1$ of 44.8 compared to 58.6 for an unsupervised morphological segmented (Morfessor). Furthermore, we confirm that the more closely our tokenizer resembles morphemes, the higher the performance of LMs.

M3 - Article in proceeding

SP - 223

EP - 229

BT - Proceedings of the 25th Nordic Conference on Computational Lingustics and 11th Baltic Conference on Human Language Technologies : NoDaLiDa/Baltic-HLT 2025

A2 - Johansson, Richard

A2 - Stymne, Sara

PB - Association for Computational Linguistics (ACL)

Y2 - 3 March 2025 through 4 March 2025

ER -

van der Goot R, Jensen A, Allerslev Schledermann E, Wildner Kildeberg M, Larsen N, Zhang M et al. MorSeD: Morphological Segmentation of Danish and its Effect on Language Modeling. I Johansson R, Stymne S, red., Proceedings of the 25th Nordic Conference on Computational Lingustics and 11th Baltic Conference on Human Language Technologies : NoDaLiDa/Baltic-HLT 2025. Association for Computational Linguistics (ACL). 2025. s. 223-229

MorSeD: Morphological Segmentation of Danish and its Effect on Language Modeling

Abstract

Konference

Adgang til dokumentet

AUB Link

Fingeraftryk

Projekter

Digital Twins for Abundant Feedback: Novel Feedback Paradigms via Explainable Multilingual Natural Language Processing

Citationsformater