@inproceedings{andersen-etal-2026-manuscript,
title = "From Manuscript to Model: Developing {HTR} for Medieval {G}reek",
author = "Andersen, Nicklas Sindlev and
MacDougall, Byron and
Yousef, Tariq and
Pizzone, Aglae",
editor = "Sprugnoli, Rachele and
Passarotti, Marco",
booktitle = "Proceedings of the Fourth Workshop on Language Technologies for Historical and Ancient Languages ({LT}4{HALA} 2026) @ {LREC} 2026",
month = may,
year = "2026",
address = "Palma, Mallorca (Spain)",
publisher = "ELRA Language Resources Association (ELRA)",
url = "https://aclanthology.org/2026.lt4hala-1.13/",
doi = "10.63317/47tkobv5mgbu",
pages = "139--151",
abstract = "We develop and evaluate manuscript-specific text line detection (TLD) and handwritten text recognition (HTR) models for two 14th-century Medieval Greek manuscripts, Vat. gr. 2228 and Phil. gr. 130, comprising 1,356 handwritten pages. From these, we curate and document 36 pages with complete, manually curated text line annotations, together with 10 additional pages with layout annotations only for TLD, forming two manuscript-specific ground truth (GT) datasets. To ensure representative evaluation despite limited annotations, validation splits are optimized for character coverage and distributional similarity using Jensen-Shannon divergence. Using the Transkribus platform, we train manuscript-specific TLD models from scratch and manuscript-specific HTR models, comparing HTR training from scratch with fine-tuning of a publicly available Medieval Greek base model. TLD achieves validation pixel-wise misclassification rates of 5.42{\%} for Vat. gr. 2228 and 8.76{\%} for the more layout-variable Phil. gr. 130. For HTR, fine-tuning consistently outperforms training from scratch. On validation pages with manually curated text line annotations, Vat. gr. 2228 reaches 5.13{\%} character error rate (CER) and 23.66{\%} word error rate (WER), while Phil. gr. 130 reaches 27.13{\%} CER and 65.72{\%} WER after continued training. A supplementary held-out evaluation on Vat. gr. 2228 shows that the fine-tuned model reaches 5.97{\%} CER and 23.52{\%} WER on test pages with manually corrected line polygons, degrading to 12.12{\%} CER and 44.21{\%} WER under automatic TLD-based segmentation. The study also provides a reproducible workflow and evaluation protocol for Medieval Greek HTR under low-resource conditions."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="andersen-etal-2026-manuscript">
<titleInfo>
<title>From Manuscript to Model: Developing HTR for Medieval Greek</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nicklas</namePart>
<namePart type="given">Sindlev</namePart>
<namePart type="family">Andersen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Byron</namePart>
<namePart type="family">MacDougall</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tariq</namePart>
<namePart type="family">Yousef</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aglae</namePart>
<namePart type="family">Pizzone</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Fourth Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA 2026) @ LREC 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Rachele</namePart>
<namePart type="family">Sprugnoli</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marco</namePart>
<namePart type="family">Passarotti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>ELRA Language Resources Association (ELRA)</publisher>
<place>
<placeTerm type="text">Palma, Mallorca (Spain)</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>We develop and evaluate manuscript-specific text line detection (TLD) and handwritten text recognition (HTR) models for two 14th-century Medieval Greek manuscripts, Vat. gr. 2228 and Phil. gr. 130, comprising 1,356 handwritten pages. From these, we curate and document 36 pages with complete, manually curated text line annotations, together with 10 additional pages with layout annotations only for TLD, forming two manuscript-specific ground truth (GT) datasets. To ensure representative evaluation despite limited annotations, validation splits are optimized for character coverage and distributional similarity using Jensen-Shannon divergence. Using the Transkribus platform, we train manuscript-specific TLD models from scratch and manuscript-specific HTR models, comparing HTR training from scratch with fine-tuning of a publicly available Medieval Greek base model. TLD achieves validation pixel-wise misclassification rates of 5.42% for Vat. gr. 2228 and 8.76% for the more layout-variable Phil. gr. 130. For HTR, fine-tuning consistently outperforms training from scratch. On validation pages with manually curated text line annotations, Vat. gr. 2228 reaches 5.13% character error rate (CER) and 23.66% word error rate (WER), while Phil. gr. 130 reaches 27.13% CER and 65.72% WER after continued training. A supplementary held-out evaluation on Vat. gr. 2228 shows that the fine-tuned model reaches 5.97% CER and 23.52% WER on test pages with manually corrected line polygons, degrading to 12.12% CER and 44.21% WER under automatic TLD-based segmentation. The study also provides a reproducible workflow and evaluation protocol for Medieval Greek HTR under low-resource conditions.</abstract>
<identifier type="citekey">andersen-etal-2026-manuscript</identifier>
<identifier type="doi">10.63317/47tkobv5mgbu</identifier>
<location>
<url>https://aclanthology.org/2026.lt4hala-1.13/</url>
</location>
<part>
<date>2026-05</date>
<extent unit="page">
<start>139</start>
<end>151</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T From Manuscript to Model: Developing HTR for Medieval Greek
%A Andersen, Nicklas Sindlev
%A MacDougall, Byron
%A Yousef, Tariq
%A Pizzone, Aglae
%Y Sprugnoli, Rachele
%Y Passarotti, Marco
%S Proceedings of the Fourth Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA 2026) @ LREC 2026
%D 2026
%8 May
%I ELRA Language Resources Association (ELRA)
%C Palma, Mallorca (Spain)
%F andersen-etal-2026-manuscript
%X We develop and evaluate manuscript-specific text line detection (TLD) and handwritten text recognition (HTR) models for two 14th-century Medieval Greek manuscripts, Vat. gr. 2228 and Phil. gr. 130, comprising 1,356 handwritten pages. From these, we curate and document 36 pages with complete, manually curated text line annotations, together with 10 additional pages with layout annotations only for TLD, forming two manuscript-specific ground truth (GT) datasets. To ensure representative evaluation despite limited annotations, validation splits are optimized for character coverage and distributional similarity using Jensen-Shannon divergence. Using the Transkribus platform, we train manuscript-specific TLD models from scratch and manuscript-specific HTR models, comparing HTR training from scratch with fine-tuning of a publicly available Medieval Greek base model. TLD achieves validation pixel-wise misclassification rates of 5.42% for Vat. gr. 2228 and 8.76% for the more layout-variable Phil. gr. 130. For HTR, fine-tuning consistently outperforms training from scratch. On validation pages with manually curated text line annotations, Vat. gr. 2228 reaches 5.13% character error rate (CER) and 23.66% word error rate (WER), while Phil. gr. 130 reaches 27.13% CER and 65.72% WER after continued training. A supplementary held-out evaluation on Vat. gr. 2228 shows that the fine-tuned model reaches 5.97% CER and 23.52% WER on test pages with manually corrected line polygons, degrading to 12.12% CER and 44.21% WER under automatic TLD-based segmentation. The study also provides a reproducible workflow and evaluation protocol for Medieval Greek HTR under low-resource conditions.
%R 10.63317/47tkobv5mgbu
%U https://aclanthology.org/2026.lt4hala-1.13/
%U https://doi.org/10.63317/47tkobv5mgbu
%P 139-151
Markdown (Informal)
[From Manuscript to Model: Developing HTR for Medieval Greek](https://aclanthology.org/2026.lt4hala-1.13/) (Andersen et al., LT4HALA 2026)
ACL
- Nicklas Sindlev Andersen, Byron MacDougall, Tariq Yousef, and Aglae Pizzone. 2026. From Manuscript to Model: Developing HTR for Medieval Greek. In Proceedings of the Fourth Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA 2026) @ LREC 2026, pages 139–151, Palma, Mallorca (Spain). ELRA Language Resources Association (ELRA).