Computer science

The modification of the SBERT language model for identifying ESG risks based on textual data from companies and supervisory activities

A. V. Buzmakov^a, D. А. Kirpishchikov^a, Yu. N. Naidenova^a, S. N. Paklina^a, P. A. Parshakov^a, R. I. Solomatin^a, N. S. Sotiriadi<sup>b

a</sup> HSE University, 37, bul. Gagarinа, Perm’, 614000, Russian Federation
^b PJSC Sberbank, 19, ul. Vavilovа, Моscow, 117312, Russian Federation

Abstract: An approach has been developed to identify risks associated with companies’ environmental impact, social responsibility, and governance quality (Environmental, Social, and Governance — ESG risks) based on textual information about the company. To achieve this, a modification of the SBERT language model is proposed with a clearly defined distance function for the embedding space. The model is trained on data from supervisory activities and texts of corporate websites. An example of interpretation of the model’s result is provided.

Keywords: ESG, natural language processing model, model training, topic modeling, website.

UDC: 004.852, 004.891, 51-77

MSC: 68T50, 91B99

Received: March 7, 2024
Accepted: December 12, 2024

DOI: 10.21638/spbu10.2025.106