[4I1-GS-11] AI and Society:

[4I1-GS-11-04] An Exploration of the Generation and Utilization of Simple Causal Graphs Using Object-Based Nodes

〇Toshimitsu Tanaka Tanaka¹, Yuichiro Okugawa¹, Tomomi Nagao¹, Akihiko Katsuno¹, Hiroaki Kawata¹ (1. NTT Space Environment and Energy Laboratories)

Keywords:Causal Inference Graphs, Large Language Model, Scenario Analysis

In recent years, companies are increasingly required to develop long-term strategies to address uncertain futures, including extreme weather and conflicts. Scenario analysis is effective for strategy formulation, but the extensive information collection and analysis can be a significant burden.

This study proposes a method to construct readable causal inference graphs (causal graphs) from multiple documents, enabling comprehensive analysis of large information volumes. The causal graphs use object-type nodes, where each node represents a subject with attribute information about states or actions. The attribute information is extracted using structured summaries generated by Large Language Models (LLMs). The causal graphs are built with three simple elements: causes, interventions, and outcomes.

To network causal graphs, nodes with high similarity in causes and outcomes are identified and merged based on a similarity threshold. By leveraging node attributes, the method reduces computational costs for node matching and networking while enabling classification based on attributes.

This presentation outlines the proposed method, details the networking approach, and discusses its advantages and challenges.