대형언어모델(LLM)의 역사지리 지식과 AI 역사지도 제작 가능성의 모색

Abstract

The purpose of this study is to analyze the current state of historical geographic knowledge of Large Language Model (LLM) and to explore whether it is possible to create AI historical maps using LLM historical geographic knowledge. First, I analyzed the answers to historical geographic queries by point, line, and area geometry type for OpenAPI's ChatGPT, Google's Gemini, and Naver's ClovaX, which are major LLM-based GPT services, and estimated the causes of errors. Second, to explore the possibility of AI historical mapping, I created a map of Eupchi(邑治, administrative center) in Gyeonggi-do during the Joseon Dynasty. Then, I compared the current location information of the Eupchi provided by LLM with the results of the Eupchi restored by historical GIS for each county. The results showed that LLM is capable of answering a limited number of queries related to historical geographic information in Korea. However, various errors were found, the main reasons for which are, first, that the characteristics of geographic data are difficult to describe in text alone, and second, that the current LLM technology has limitations in handling historical geographic information consisting of multiple time layers, and third, there is a lack of publicly available historical geographic information. To improve these issues, I propose the construction of historical GIS (HGIS) layers on an ongoing basis and the knowledge expansion of LLM models through semantic data of historical geography.