Grammatical error correction of Slovenian school essays using large language models

Klemen, Matej (author), Božič, Martin (author), Arhar Holdt, Špela (author), Robnik Šikonja, Marko (author)

Grammatical error correction (GEC) is the task of automatically detecting and correcting grammatical errors in text. Large language models have enabled the development of accurate automated methods for detecting and correcting certain types of errors. In the educational domain, the aim of GEC is to aid teachers in correcting student errors. Excessive paraphrasing is a property of Generative Pre-trained Transformer-based models and is undesirable in the language education context. To avoid this, we develop multiple Slovenian models for correcting errors in spelling, word case (capitalization), word form, and word order. We describe the training data construction, training process, and model evaluation approach using the Šolar-Eval 1.0 corpus of school essays authored by primary and secondary school students. Our quantitative evaluation shows that the developed models have reasonably high accuracy levels, and our qualitative evaluation highlights the strengths and weaknesses of the models and the evaluation process. The analysis reveals multiple challenges and promising future directions for improving both model development and the evaluation process.

• Language: English
• Keywords: large language models, grammatical error correction, educational domain, synthetic data construction
• Work type: Article
• Typology: 1.02 - Review Article
• Organization: FRI - Faculty of Computer and Information Science ; FF - Faculty of Arts
• Publication status: Published
• Publication version: Version of Record
• Year: 2025
• Number of pages: Str. 162–176
• Numbering: Letn. 76, št. 3
• PID: 20.500.12556/RUL-182472
• UDC: 371.68
• ISSN on article: 0038-0474
• DOI: 10.63384/sptB53z793a
• COBISS.SI-ID: 259208195
• Publication date in RUL: 19.05.2026

Record is a part of a journal

• Title: Sodobna pedagogika
• Shortened title: Sodob. pedagog.
• Publisher: Zveza društev pedagoških delavcev Slovenije
• ISSN: 0038-0474
• COBISS.SI-ID: 761348

Licences

• License: CC BY-SA 4.0, Creative Commons Attribution-ShareAlike 4.0 International
• Link: http://creativecommons.org/licenses/by-sa/4.0/
• Description: This Creative Commons license is very similar to the regular Attribution license, but requires the release of all derivative works under this same license.

Secondary language

• Language: Slovenian

Abstract

Strojno popravljanje slovničnih napak je naloga, ki zajema samodejno zaznavanje in popravljanje slovničnih napak v besedilu. Na področju izobraževanja je cilj metod pomagati učiteljem pri popravljanju napak učencev. Veliki jezikovni modeli omogočajo razvoj natančnih avtomatskih metod za zaznavanje in popravljanje določenih vrst napak. Da bi se izognili pretiranemu parafraziranju, ki je značilno za modele tipa GPT, in je v kontekstu poučevanja jezika nezaželeno, predstavimo več razvitih slovenskih modelov tipa BERT in T5 za popravljanje različnih vrst napak. Te vključujejo črkovalne napake, napake v rabi velikih začetnic, besednih oblik in besednega reda. V članku opišemo postopek ustvarjanja učnih podatkov, postopek učenja ter postopek evalvacije modelov na korpusu Šolar-Eval 1.0, ki vsebuje šolske spise osnovnošolcev in srednješolcev. Avtomatska evalvacija kaže razmeroma visoko natančnost razvitih modelov, medtem ko ročna kvalitativna evalvacija razkrije prednosti in slabosti razvitih modelov ter evalvacijskega postopka. Analiza razkriva številne izzive in obetavne smeri za nadaljnje izboljšave tako pri razvoju modelov kot pri postopku evalvacije.

• Keywords: veliki jezikovni modeli, popravljanje slovničnih napak, izobraževalna domena, sintetiziranje podatkov