[1Yin-B] Poster 2

[1Yin-B-03] DAGRI Subtask 2: Task setup for farm management simulation and construction of a baseline system

〇Fuka Nakajima¹, Ryoma Itakura¹, Hiroki Sakaji¹, Akio Kobayashi², Kenta Baba², Masahiro Ohotomo², Jyun-ichi Ishihara², Tetsuo Katsuragi², Takehiro Takayanagi³Yasutomo Kimura⁴, Itsuki Noda¹ (1. Hokkaido University, 2. Agricultural Research Center, National Agriculture and Food Research Organization, 3. Simulacra Co., Ltd., 4. Otaru University of Commerce)

Keywords:

Agriculture,Large Language Model

In Japan, the importance of agricultural extension officers who provide farming guidance has been increasing, in order to foster the next generation of agricultural workers and to pass on agricultural technologies. To address this challenge, we systematize the task definition and evaluation methodology for implementing DAGRI Subtask 2, aiming to realize question answering that more faithfully reflects real-world decision-making processes in farming, and we organize them into concrete system specifications. The task extends questions beyond multi-table data referencing to include condition changes and constraints that arise in practice, thereby covering situations that require judgments close to those made in actual operations. Based on the defined evaluation framework, we build a baseline system that generates answers in a QA format. This baseline derives answers through the extraction and integration of relevant tables and a reasoning process, enabling verification of both the task feasibility and the reproducibility of the evaluation. In this paper, we report the progress in constructing a set of question–answer pairs that are close to real farming practice, the evaluation design, and the evaluation results of the baseline system.