This thesis presents a research project to develop and implement an automated scoring system to assess the ESG factors across Swiss municipalities. Initially designed as a framework applicable to all municipalities in Switzerland, we apply it to certain medium-sized cities: Nyon, Rolle, and Vevey. By fine-tuning large language models based on the transformers architecture, we classified municipal council transcripts into ESG categories and subsequently applied sentiment analysis to derive ESG ratings. Our methodology included creating a new balanced dataset translated into French from a collection of English article headlines and fine-tuning CamemBERT models. Multiple models were produced using HPC, and a weighted voting scheme was employed to combine and enhance classification accuracy.

The results indicated strong performance in identifying environmental, governance, and non-ESG-related content, with some challenges in distinguishing social aspects. Key findings revealed a significant representation of environmental topics and a notable increase in governance discussions. The social category may be underrepresented, possibly due to overlaps with the non-ESG category. Emerging trends showed stable ratings for Nyon and Vevey, while Rolle exhibited slightly more variability due to the lower number of available council transcripts.

The automated system demonstrated efficiency in analysing large volumes of municipal transcripts, providing valuable insights for policymakers. Limitations included occasional misclassification of social content and a focus on French-language data. By translating the dataset into German or Italian, the framework could be broadened to encompass the entire country, and beyond. This work offers a novel approach for automatic ESG assessment in the public sector in Switzerland, facilitating more informed decision-making and setting a benchmark for future research in automated ESG classification.