Optimising Sustainability Accounting: Using Language Models to Match and Merge Survey Indicators

Abstract

[Context] To assess the sustainability performance of companies, diverse environmental, social and governance accounting (ESGA) methods exist, each with their own set of topics and indicators. In earlier research, we have shown that several ESGA methods contain overlapping indicators. [Aim] We aim to develop a semi-automated approach for identifying the overlap between ESGA methods, and then merging the methods into a single combined method that has no redundant indicators. [Method] We have approached this goal as a model management challenge. We have surveyed companies to formulate the problem statement, conducted a literature study on model management operations, created ESGA method models according to our openESEA domain-specific language, and developed algorithms that leverage the power of language models to match and merge the methods. The matching threshold is determined by performing an experiment with 16 experts. Lastly, we validate our algorithms by merging 4 real-life ESGA methods. [Result] The algorithm has proven capable of successfully identifying overlap between ESGA methods. While we would prefer to further reduce the number of false positives, the results already provide valuable insights into the optimisation of sustainability accounting. Moreover, our findings demonstrate how language models can be used for model management.