TY - JOUR

T1 - Cross-Lingual Transfer of Named Entity Markup with Large Language Models

AU - Barakhnin, Vladimir

AU - Mussabayev, Rustam

AU - Mengliev, Davlatyor

AU - Krassovitskiy, Alexander

AU - Toleu, Alymzhan

AU - Lyutaev, Daniil

AU - Akhmetov, Iskander

AU - Ibragimov, Bahodir

N1 - Barakhnin, V.; Mussabayev, R.; Mengliev, D.; Krassovitskiy, A.; Toleu, A.; Lyutaev, D.; Akhmetov, I.; Ibragimov, B. Cross-Lingual Transfer of Named Entity Markup with Large Language Models. Informatics 2026, 13, 70. https://doi.org/10.3390/informatics13050070 This research was funded by the Science Committee of the Ministry of Science and Higher Education of the Republic of Kazakhstan (grant No. AP23486904). The research was also conducted within the state assignment of the Ministry of Science and Higher Education of the Russian Federation for the Federal Research Center for Information and Computational Technologies.

PY - 2026/5/7

Y1 - 2026/5/7

N2 - This paper investigates the problem of cross-lingual named entity recognition (NER), which involves automatically identifying entities such as persons, organizations, locations, and other structured elements in text. High-quality NER typically requires manually annotated corpora; however, for many low-resource languages, such data are scarce and costly to produce. The study addresses the following question: can annotated sentences in one language be used to transfer NER markup to their machine-translated counterparts in other languages? To explore this, we propose an approach based on a large language model (LLM) that performs two tasks simultaneously: translating a source sentence and generating BIOES-formatted entity tags for the translated output. To improve robustness and reduce semantic drift, a back-translation step is incorporated to verify meaning preservation by comparing the reconstructed source sentence with the original. The proposed method is compared with two baseline approaches: (1) annotation projection via machine translation and (2) automatic tagging using pre-existing NER tools. Performance is evaluated using standard metrics, including precision, recall, and F1-score. Experimental results demonstrate that the LLM-based approach provides a practical and efficient mechanism for transferring NER annotations across languages. While the method achieves strong and balanced performance, its quality remains influenced by translation accuracy and adherence to annotation constraints. Methodologically, the approach can be considered relatively language-independent, as it relies on general LLM capabilities, a universal tagging scheme, and multilingual semantic representations rather than language-specific model training.

AB - This paper investigates the problem of cross-lingual named entity recognition (NER), which involves automatically identifying entities such as persons, organizations, locations, and other structured elements in text. High-quality NER typically requires manually annotated corpora; however, for many low-resource languages, such data are scarce and costly to produce. The study addresses the following question: can annotated sentences in one language be used to transfer NER markup to their machine-translated counterparts in other languages? To explore this, we propose an approach based on a large language model (LLM) that performs two tasks simultaneously: translating a source sentence and generating BIOES-formatted entity tags for the translated output. To improve robustness and reduce semantic drift, a back-translation step is incorporated to verify meaning preservation by comparing the reconstructed source sentence with the original. The proposed method is compared with two baseline approaches: (1) annotation projection via machine translation and (2) automatic tagging using pre-existing NER tools. Performance is evaluated using standard metrics, including precision, recall, and F1-score. Experimental results demonstrate that the LLM-based approach provides a practical and efficient mechanism for transferring NER annotations across languages. While the method achieves strong and balanced performance, its quality remains influenced by translation accuracy and adherence to annotation constraints. Methodologically, the approach can be considered relatively language-independent, as it relies on general LLM capabilities, a universal tagging scheme, and multilingual semantic representations rather than language-specific model training.

KW - BIOES tagging scheme

KW - annotation transfer

KW - back-translation

KW - cross-lingual transfer

KW - large language models

KW - low-resource languages

KW - multilingual markup

KW - named entity recognition

KW - prompt engineering

KW - межъязыковой перенос

KW - перенос аннотаций

KW - распознавание именованных сущностей

KW - многоязычная разметка

KW - большие языковые модели

KW - разработка промптов (промпт-инжиниринг)

KW - схема разметки BIOES

KW - языки с ограниченными ресурсами

KW - обратный перевод

UR - https://www.mendeley.com/catalogue/d83ab1d7-c358-33b4-ab95-18b942d95292/

UR - https://www.scopus.com/pages/publications/105039992434

U2 - 10.3390/informatics13050070

DO - 10.3390/informatics13050070

M3 - Article

VL - 13

JO - Informatics

JF - Informatics

SN - 2227-9709

IS - 5

M1 - 70

ER -