[1Win4] Poster session 1

[1Win4-93] Construction of Causal-Chain Presentation Systems using Large Language Models

〇Akio Ikegami¹, Takehiro Takayanagi¹, Yuri Murayama¹, Kiyoshi Izumi¹ (1.The University of Tokyo)

Keywords:Causal Network, Large Language Models, Retrieval-Augmented Generation

Evidence-based decision-making is increasingly emphasized in corporate activities and policymaking. To achieve this, it is essential to understand causal relationships between social phenomena widely recognized as cause-and-effect. Conventionally, causal relationships have been collected through text mining, in addition to analyzing governmental statistics. However, ideally, the extracted causal relationships should not be overly dependent on the nature or quantity of the text corpus. Therefore, this study proposes a novel causal-chain presentation system leveraging Large Language Models (LLMs), which encapsulate general knowledge, to iteratively generate causal knowledge related to a given specific economic phenomenon and present it as a causal-chain. The experiments using financial statements revealed that while the combination of LLMs and retrieval-augmented generation (RAG) did not improve the accuracy of causal relationship generation, the use of LLMs enabled the output to be more “serendipitous and diverse” compared to conventional output. Furthermore, the results were discussed through case analyses.