Earticle

This study investigates the performance of a state-of-the-art Large Language Model(LLM) in classifying Korean neutral sentiment without additional fine-tuning and proposes an effective prompt design to improve neutral sentiment classification. To enhance the accuracy of neutral sentiment detection, we introduce a prompt based on Aspect-Based Sentiment Analysis(ABSA) and conduct a comparative evaluation using the proposed approach. The proposed prompt consists of a five-step procedure, including aspect identification and sentiment ratio–based classification, enabling more fine-grained sentiment reasoning. Experimental results demonstrate that the proposed prompt significantly improves classification accuracy when applied to the GPT-4-turbo model, thereby validating the effectiveness of prompt-based control for neutral sentiment analysis in Korean.

This study examines the usage patterns and discourse functions of the ‘ani X’ construction as a complex discourse marker in Korean spoken discourse. It identifies and categorizes instances in which the discourse marker ‘ani’ appears in combination with other expressions. The analysis shows that 37.8% of discourse marker tokens beginning with ‘ani’ are realized in complex forms, indicating that such combinations are not marginal. Focusing on the most frequent types—‘ani keunde’, ‘ani geu’, ‘ani mwo’, ‘ani geureonikka’, and ‘ani geunyang’—the study illustrates how these constructions are used in various interactional contexts, including topic shifts, justification, mitigation, and the management of interpersonal stance. The study provides a descriptive basis for understanding complex discourse markers in Korean and highlights the analytical value of moving beyond single-word forms.

This study verifies the necessity of adding 'educational institution' as a metadata variable to the National Institute of Korean Language's Korean Learner Corpus by comparing a single-institution corpus (Yonsei University Korean Language Institute) with the multi-institution integrated corpus (NIKLC). Sub-corpora were constructed from beginner-level (1-2) writing samples, controlling for variables such as proficiency, nationality, topic, genre, and token count. Analyses included lexical diversity, average sentence length, chi-square tests, and log-likelihood ratio. Results showed statistically significant differences (p<0.001) in lexical distribution and morpheme usage, manifesting as institutional variations in style (declarative vs. polite endings) and vocabulary choice ('한국어' vs. '한국말'). These findings demonstrate the independent impact of institutional factors on learner language, proposing the addition of institution variables to enhance the balance and research potential of the NIKLC corpus.

The study reviews the evolution of Korean corpus construction from the Sejong Project to the AI era, highlighting the shift from linguistic analysis to application-driven datasets. It emphasizes the need for culturally grounded Korean corpora that capture honorifics, idioms, and social context to enhance AI understanding. A seven-stage framework is proposed to build a comprehensive Korean language–culture corpus for improving LLMs’ cultural competence.