Junsub Kim  Dongnyeok Choi  Sung Gu Kim  Deuk Hwa Kim

10.5626/JOK.2026.53.2.117

embedding model contrastive learning Hard Negative BGE-M3

Korean language models specifically designed for the defense sector are still limited, even with the rapid advancements in text embeddings. In this study, we fine-tune the multilingual BGE-M3 model to better understand military terminology and investigate how negative sampling in contrastive learning impacts downstream performance. We evaluate three strategies: Easy (random negatives), Hard (lexicographic adjacency), and Harder (similarity-mined negatives). Our analysis, based on clustering metrics such as Accuracy, NMI, and ARI using a defense news dataset, reveals that the similarity-based Harder strategy consistently outperforms the others. Further evaluations on the KorSTS dataset demonstrate that the Harder approach maintains strong Spearman and Pearson correlations, indicating successful domain adaptation without compromising overall semantic competence. Interestingly, the three Harder variants—negatives mined with BGE-M3, ko-sroberta, and multilingual-e5—produce nearly identical similarity distributions and comparable improvements, while the Easy strategy plateaus and the Hard strategy shows only moderate performance. These findings suggest that mining sufficiently similar negatives, as opposed to using random or adjacent ones, is crucial for effective, domain-specific fine-tuning of multilingual embedding models.