An Analysis of Korean Homonym Comprehension in Generative AI: Focusing on Semantic Disambiguation through ChatGPT-4o.

This study aims to experimentally examine the semantic disambiguation ability of the generative AI model ChatGPT-4o with regard to Korean homonyms, thereby exploring its linguistic accuracy and limitations. The target homonyms selected for the study include Bae (Pear vs. Multiple), Nun (Eye vs. Snow), Sijang (Mayor vs. Market), and Jungsik (Lunch vs. Chinese food). From the 2023 National Institute of Korean Language newspaper corpus, a total of 400 context-based example sentences were collected, with an equal number of examples representing each sense of the target words. The experiment was designed with two independent variables: input type (file-based vs. text-based) and data volume (100 vs. 50 sentences). Results showed that text-based inputs yielded relatively higher semantic disambiguation accuracy than file-based inputs, while the amount of data did not significantly affect performance. The study provides the following implications: First, homonyms can serve as effective linguistic tools for evaluating the contextual comprehension ability of generative AI. Second, the input format and data volume can have a practical impact on AI's semantic processing. Third, semantic disambiguation was primarily guided by a rule-based approach, relying on specific keywords rather than deep contextual integration. These findings highlight both the potential and limitations of generative AI in handling languages with high contextual dependency, such as Korean.