[2I6-GS-10] AI application: Finance

[2I6-GS-10-04] Text Mining of Business Risks Description During the Pandemic Using Large Language Models

〇Takeshi UMEHARA^2,3, Hideaki Takeda^1,2 (1. National Institute of Informatics, 2. Graduate Institute for Advanced Studies, 3. Nikkei Research Inc.)

Keywords:Large Language Model, Text Mining, Financial Data, Securities Report

In this paper, we try to analyse business risks descriptions in annual reports of listed companies in Japan. Especially, we analyse time series changes of business risk descriptions between the pandemic of COVID-19. Since business risk descriptions are qualitative information, it is hard to recognize that the listed companys consider what incidents as the business risks and how the incidents affect their business by information extraction techniques. We extract business risk descriptions from XBRL files, choose description of infectious disease risk by pretrained BERT and K-means, and execute dimension reduction by pretrained Sentence BERT and PacMAP. As a result, we found the changes of recognitions about the risk, which related works did not mention about.