Environmental, Social and Governance (ESG) Scores Automation in Global Reporting Initiative (GRI) with Natural Language Processing

The increasing importance of Environmental, Social, and Governance (ESG) factors in making investment decisions has triggered a noteworthy transformation in the global investing scenario. This study investigates various distilled BERT models in Python, including all-MiniLM - L 12-v2, all-MiniLM-L6-v2, all-mpnet-base-v2, and paraphrase-MiniLM-L6v2, to improve the accuracy and efficiency of ESG scoring. The performance of these models will undertake a detailed evaluation employing the F1 score criteria, which will offer an understanding of the accuracy and consistency of the scoring outputs produced by each model. The reason for this detailed review is to pledge a strong valuation of the effectiveness of the automated scoring procedure. The study shows that utilising textual similarity for ESG rating considerably decreases the time needed in comparison to the conventional human scoring procedure. Throughout the time of this study, all-MiniLM-L6-v2 was used as the base comparison model, however, it was observed that the all-mpnet-base-v2 model provided a relatively better F1 score due to its value being determined by its 768-dimensional embeddings. Embeddings with better length possess the volume to cover more difficult and subtle semantic representations, therefore potentially yielding superior performance. In the result finding summary, all-mpnet-base-v2 has the highest F1 score compared with the experiment's sentence transformers.