Comparison of Korean NER models and Cross-Lingual Data Adaptation for NER in Low-Resource Language

Park, Yoojin (author), Bohak, Ciril (mentor), Šmajdek, Uroš (comentor), Kim, Min Young (comentor)

This study was designed to systematically evaluate and compare the performance of various Korean NER models in the same domain. By doing so, we aim to address the research gap surrounding domain-specific performance variations, which remain under-discussed in existing Korean NER literature. The focus of this research is not merely to report performance metrics but to draw meaningful inferences about the generalizability, robustness, and practical applicability of Korean NER systems in real-world, domain-diverse scenarios. To achieve this, we have selected several state-of-the-art Korean NER models, primarily based on pre-trained transformer architectures. These models were first fine-tuned and validated using a general-purpose Korean NER corpus to establish a consistent baseline. Subsequently, their performance was tested on a domain-specific dataset based on Korean Wikipedia articles. Model evaluation was carried out using standard metrics such as precision, recall, and F1-score. Loss curves were employed to better interpret model training performance. As an additional experiment, we employed a translation process to indirectly address the limitation of Korean NER models in a specific domain. Through this process, we were able to confirm the cross-lingual transferability.

• Language: English
• Keywords: Korean NER models, Cross-Lingual NER system, Cross-lingual Transfer
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2025
• PID: 20.500.12556/RUL-174626
• COBISS.SI-ID: 255327747
• Publication date in RUL: 07.10.2025

Secondary language

• Language: Slovenian
• Title: Primerjava korejskih modelov NER in medjezikovne prilagoditve podatkov za NER v jeziku z malo viri

Abstract

Ta študija je bila zasnovana za sistematično ocenjevanje in primerjavo uspešnosti različnih korejskih modelov NER na istem področju. S tem želimo odpraviti raziskovalno vrzel v zvezi z variacijami uspešnosti, specifičnimi za podro čje, ki ostajajo premalo obravnavane v obstoječi korejski literaturi NER. Poudarek te raziskave ni le poročanje o meritvah uspešnosti, temveč tudi smiselne sklepe o posploševanju, robustnosti in praktični uporabnosti korejskih sistemov NER v resničnih scenarijih. Da bi to dosegli, smo izbrali več najsodobnejših korejskih modelov NER, ki temeljijo predvsem na vnaprej usposobljenih arhitekturah transformatorjev. Ti modeli so bili najprej natan čno nastavljeni in potrjeni z uporabo splošnega korejskega korpusa NER, da bi vzpostavili dosledno izhodišče. Nato je bila njihova uspešnost preizku šena na naboru podatkov, specifičnem za domeno, ki temelji na člankih korejske Wikipedije. Vrednotenje modela je bilo izvedeno z uporabo standardnih meritev, kot so natančnost, odpoklic in F1-rezultat. Krivulje izgub so bile uporabljene za boljšo razlago uspešnosti usposabljanja modela. Kot dodaten poskus smo uporabili prevajalski postopek, da bi posredno obravnavali omejitev korejskih modelov NER na določeni domeni. S tem postopkom smo lahko potrdili medjezikovno prenosljivost.

• Keywords: modeli NER za Korejščino, medjezikovni modeli NER, medjezikovni prenos