KoPolitical-LLM: 한국 정치 연구를 위한 거대언어모델 탐구

초록

This paper presents KoPolitical-LLM, a framework consisting of two concrete research cases designed to explore the potential of Korean Computational Political Science in applying large language models to the study of Korean politics. While the use of LLMs in political research has been actively pursued in Anglophone and European academia, its adoption in Korean political research remains at a relatively nascent stage. Departing from the recognition that political discourse is deeply embedded in the unique historical and institutional contexts of individual nations, this paper aims to demonstrate that methodologies developed within the tradition of computational political science can be meaningfully extended to Korean political data. The first research case, KoP-Stance, is a Korean political stance detection framework built on official spokesperson press releases and briefings from the Democratic Party of Korea and the People Power Party, spanning approximately nine years. Both encoder-based fine-tuning and decoder-based in-context learning were applied, with fine-tuned encoder models achieving macro F1 scores above 0.83, outperforming decoder-based in-context learning by approximately eight to ten percentage points. The fine-tuned models have been made publicly available via HuggingFace. The second research case, KoPAA (Korean Political Actor Agent), assigns the ideological identities of the two major parties to large language models and predicts each party's stance on actual bills introduced in the National Assembly of Korea. Evaluated across Korean-specialized models and larger frontier models including GPT and Claude series, the results reveal a notable performance gap, with frontier models achieving near-perfect accuracy while Korean-specialized models exhibited lower and more uneven performance. This paper further discusses the rationale for Korean computational political science, grounded in two distinctive properties of political data, namely originality and societal impact, and proposes future directions including interdisciplinary collaboration, benchmark development, and the conceptualization of Korean politics as a complex system.

키워드