[3Xin4] Poster session 1

[3Xin4-29] Impact Inference on ESG Indicators from Business Documents using Large Language Models and Verification of the Relationship between Statistical Causal Inference

〇Yusuke Ohtsubo¹ (1.Hitachi)

Keywords:NLP, fundation model, ESG

In recent years, companies have been required to disclose the impact of their businesses on ESG indicators. The automation potential of creating such disclosure documents was investigated using large language models (LLMs), which have seen significant development in recent years. Previous studies that examined the relationship between LLMs and causal inference focused only on problems with head-to-tail or tail-to-tail causal structures. Therefore, we comprehensively examined the relationship between LLMs and basic items of statistical causal inference, including intervention and backdoor paths, in addition to the remaining causal structure of head-to-head. As a result, we found that while LLMs can infer the impact of businesses without any sense of discomfort, the resulting inference is not based on concepts such as intervention or backdoor paths.