2025 Volume 34 Issue 2 Pages 51-67
We conducted a smart agriculture demonstration project in Japan to clarify the effects of introducing smart agriculture technology. Agricultural management indicators were developed on the basis of the project’s results. In the process, we considered that visualized agricultural information on the causal relationship among agricultural technology, farm work, and the effects of technology adoption (“technology–work–effect table”) would facilitate the development of the indicators. However, such information is based mostly on documents, the manual deciphering of which is considerably labor intensive, time consuming, and limited. Large language models (LLMs) may be effective in solving this problem. Therefore, we investigated the possibility of using an LLM to generate technology–work–effect tables from agricultural documents. Specifically, experiments used GPT-3.5 Turbo, developed by OpenAI, to generate tables from document data in the “smart agriculture demonstration project results portal”, provided by NARO, and we evaluated and analyzed the accuracy of the tables. Analysis of 4400 technology–work–effect tables generated from agricultural documents by the LLM revealed that they were not always correct, and that the content of the documents significantly influenced the output variability. Furthermore, the appropriate experimental conditions are to use plain-text documents in a more deterministic LLM setting, such as setting the temperature parameter to 0, under which conditions the frequency of the average output is high. These findings indicate that the accuracy of LLM output tables must be improved if these tables are to be put to practical use.
