ESG-BERT

🏦 ESG-BERT v1.0

A high-quality language model for Environmental, Social, and Governance (ESG) sentiment analysis.
Each pre-trained model is fine-tuned on around 3,500 labeled ESG-related sentences, designed to establish benchmark standards for ESG-related text modeling.

---

📑 Contents

1. Introduction
2. Access To The Model
3. Acknowledgements
4. Contact Information
5. References and Citation

---

1. Introduction

Many statistical and machine learning methods require high-quality, labeled training data.
However, in the ESG (Environmental, Social, and Governance) domain, well-annotated textual resources are scarce. A good baseline model to analyse the compliance to ESG is also hard to obtain.

ESG-BERT contributes to bridging the gap by providing a benchmark in the form of BERT model fine-tuned on ESG-Bank specialised in analysing sentence sentiment with respect to ESG compliance.

---

2. Access To The Model

1. Choose the base BERT models: BERT-Base, BERT-1, BERT-2, BERT-Mini.
2. Download the dataset from https://huggingface.co/datasets/TUM-TFAI/esg-bank.
3. Choose which agreement dataset the BERT model wants to be trained on and put the dataset on the corresponding folder. For example, for BERT-Base, the choices would be ESG-Bank-100, ESG-Bank-75, ESG-Bank-66
4. In the corresponding folder, you find a Jupyter Notebook (for example: ESG-BERT-Base-100) script to train the selected base-model on the selected dataset.

---

📦 Included Files

Into ESG-BERT-Base and ESG-BERT-1, we also add the fine-tuned weights to the model.

With this, one can immediately perform inference on the fine-tuned models on the script by changing the inference path.

---

3. Acknowledgements

The development of ESG-BERT v1.0 was supported and funded by the AI4BuildingESG project from TUM Georg Nemetschek Institute Artificial Intelligence for the Built World
We thank all annotators and contributors from the Technical University of Munich for their effort and domain expertise.

---

4. Contact Information

For questions, feedback, or collaboration inquiries regarding ESG-BERT, please contact:

📧 [[email protected]]

---

5. References and Citation

If you use this model in your research, please cite:

@article{esgbert2025, title = {Decoding Sustainability: ESG-Bank and ESG-BERT for Understanding Formal ESG Disclosures}, author = {Timothy, Michael and Sabanayagam, Mahalakshmi and Eisel, Niklas and Zhu, Bing and Ghoshdastidar, Debarghya}, year = {2025}, version = {1.0}, url = {https://huggingface.co/datasets/TUM-TFAI/ESG-BERT} }

For commercial or non-academic use, please contact us for licensing terms.

---

Version: 1.0
Last modified: 16/10/2024

---

license: apache-2.0