astrohumanoid/esg_bank

🏦 ESGBank v1.0

A high-quality, human-annotated dataset for Environmental, Social, and Governance (ESG) sentiment analysis.
This resource provides over 4,800 sentences annotated for sentiment at the phrase level, designed to establish benchmark standards for ESG-related text modeling.

---

📑 Contents

1. Introduction
2. Data
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.

ESGBank contributes to addressing this gap by providing a collection of approximately 5,000 sentences with human-annotated sentiment labels for benchmarking alternative modeling techniques.

The annotation task classifies each sentence as positive, negative, or neutral, considering only information explicitly stated in the text.
Annotators evaluated sentences through an ESG perspective, identifying whether the content indicates a positive, negative, or neutral contribution to ESG principles.
Sentences irrelevant to ESG were labeled as neutral.

---

2. Data

ESGBank consists of 4,889 sentence–sentiment pairs, annotated by 8 individuals with backgrounds in ESG-related topics.
All annotators are researchers and students at the Technical University of Munich, Chair of Real Estate Management.

Each sentence was randomly assigned to 5 annotators for labeling, using the following encoding:

Code	Meaning
0	Neutral
1	Positive
2	Negative
9	Not a Sentence

If a sentence was labeled 9 (Not a Sentence) by a majority, it was excluded from the dataset.

---

🧭 Gold Standard Labeling

Final sentiment labels were determined using a majority-vote scheme, following the same methodology as the Financial Phrase Bank.
All remaining sentences have been validated by a majority of annotators as meaningful, interpretable sentences.

We provide five reference subsets of ESGBank based on the strength of majority agreement:

Agreement Threshold	Dataset File	# Sentences
=100%	esg_bank_100_agreement	1458
>= 75%	esg_bank_75_agreement	3074
>= 67%	esg_bank_66_agreement	3696
> 50%	esg_bank_50+_agreement	4238
≥ 50%	esg_bank_50_agreement	4766

---

📦 Included Files

Each agreement folder contains:

1. Full Dataset — all sentences meeting the majority criteria
2. Training Data (80%) — random stratified split (by label and length)
3. Training Data + Metadata — includes word count and length classification
4. Test Data (20%) — random stratified split (by label and length)
5. Test Data + Metadata — includes word count and length classification

---

🧩 Additional Resources

• Full Dataset — all annotated sentence–sentiment pairs
  → esg_bank.csv

  Labels are determined by majority vote; ties are resolved using priority:

0 (Neutral) > 1 (Positive) > 2 (Negative) > 9 (Not a Sentence)

• Dual-Label Subset — 529 sentences with a 50–50 tie between two classes
  → esg_bank_dual_label.csv

All datasets are provided in CSV format.

---

3. Acknowledgements

The development of ESGBank v1.0 was supported by [Add organization or funding source here].
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 ESGBank, please contact:

📧 [Your contact email or contact page link here]

---

5. References and Citation

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

@article{esgbank2025, 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://github.com//ESGBank} }

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

---

Version: 1.0
Last modified: 16/10/2024