NTP-Based DPO Training for Suppressing Korean-Chinese Code-Switching 


Vol. 15,  No. 1, pp. 54-60, Jan.  2026
10.3745/TKIPS.2026.15.1.54


PDF
  Abstract

Large Language Models (LLMs) have demonstrated outstanding performance across various natural language processing (NLP) tasks. However, when Chinese-pretrained LLMs are applied in Korean environments, unintended code-switching between Korean and Chinese frequently occurs. This not only undermines user trust but also poses critical issues in domains requiring linguistic accuracy, such as translation, education, and official document writing.To address this problem, we propose a Direct Preference Optimization (DPO) training method based on Next Token Prediction (NTP). Specifically, we leverage NTP to detect confusion points—the moments when Chinese tokens first appear in Korean outputs. Using these points, we construct the NTP-CS dataset by pairing responses with code-switching (rejected) against responses without code-switching (chosen), and train the model accordingly. Experimental results show that our proposed NTP-CS approach consistently outperforms both LLM-CS, which induces code-switching via LLM prompts, and LLM-TX, which translates entire sentences. Notably, across all datasets, chosen responses exhibited positive log-likelihood values, while rejected responses showed negative values, forming an ideal probability pattern. This demonstrates that datasets constructed around local confusion points are more effective than simple translation datasets for mitigating code-switching. We expect that this research will improve the consistency of Korean-centric outputs in multilingual settings and help reduce the quality gap between Korean and Chinese responses.

  Statistics
  Cite this article

[IEEE Style]

S. Ko and Y. Shin, "NTP-Based DPO Training for Suppressing Korean-Chinese Code-Switching," The Transactions of the Korea Information Processing Society, vol. 15, no. 1, pp. 54-60, 2026. DOI: 10.3745/TKIPS.2026.15.1.54.

[ACM Style]

Sungmin Ko and Youhyun Shin. 2026. NTP-Based DPO Training for Suppressing Korean-Chinese Code-Switching. The Transactions of the Korea Information Processing Society, 15, 1, (2026), 54-60. DOI: 10.3745/TKIPS.2026.15.1.54.

Forms

Search
(IN TITLE, AUTHOR, ABSTRACT,KEYWORDS)

Advanced Search ▼

POPULAR KEYWORDS
(TOP 10 KEYWORDS)

Recent Publications
(LAST 3 YEARS)

Old Journals

Indexing

All publications of TKIPS is indexed in DOI, EBSCO, Google Scholar, Crossref, and CrossCheck.

TKIPS is also selected as the Journal for Accreditation by NRF (National Research Foundation of Korea).

Related Journals