---
license: cc-by-nc-sa-4.0
task_categories:
- text-classification
- zero-shot-classification
language:
- en
tags:
- finance
size_categories:
- 1K<n<10K
configs:
- config_name: esg-prospectus-clarity-category
  data_files: "esg-prospectus-clarity-category.csv"
- config_name: esg-prospectus-clarity-granular-category
  data_files: "esg-prospectus-clarity-granular-category.csv"
---

# Dataset Card for ESG-Prospectus-Clarity-Category

### Dataset Summary

This dataset is manually annotated quality training dataset of 1155 ESG language instances (4 classes), obtained via a data extraction pipeline from summary prospectuses of sustainable (ESG) funds.
The ESG sentences extracted from ‘Principal Investment Strategy’ sections of the documents. Following are the four classes.
1. Specific ESG Language
2. Ambiguous ESG Language
3. Generic ESG language
4. Risk ESG language

All the instances are related to ESG investment language present in prospectus of funds. Further all instances were annotated for language clarity classes.

### Supported Tasks and Leaderboards

Text Classification (Language style classification)
Few Shot Classification

### Languages

English

## Dataset Structure

### Data Instances

Total instances: 1155

classwise instances:

'Specific ESG': 320
'Ambiguous ESG': 283
'Generic ESG': 264
'Risk ESG': 288

### Data Fields

```
{ "Text": "The Sub-fund's weighted carbon footprint score is equal or better than that of the Custom Bloomberg Climate Transition Benchmark.",
  "Label": "specific"

  "Text": "The Sub-fund invests a minimum of 5% in green, social, sustainable, and/or sustainability-linked bonds.",
  "Label": "specific"

  "Text": "The Fund will seek to invest in companies with sustainable business models which have a strong consideration for ESG risks and opportunities.",
  "Label": "ambiguous"
}
```

### Data Splits

There's no train/validation/test split.

However the dataset is available two level of categorizations:

`esg-prospectus-clarity-category.csv`: Number of classes: 4 ('specific', 'ambiguous', 'generic', 'risk')
`esg-prospectus-clarity-granular-category.csv`: Number of classes: 7 ('specific', 'ambiguous', 'generic', 'general-risk', 'performance-risk', 'data-risk', 'disclaimer-risk')

## Dataset Creation

### Source Data

#### Initial Data Collection and Normalization

The process begins with downloading the public ‘Summary Prospectuses’ from literature sections of the official websites of various Asset Management Companies (AMCs). 
We collected approximately 250 sustainable products prospectuses.

#### Who are the source language producers?

The source data was written and published by various fund issuers (Asset Management Companies).

### Annotations

#### Annotation process

The dataset was divided into three subsets and each annotator was allocated 2 subset of sentences and was given few weeks to label the sentences. 
Consequently, each of the 1155 instances was annotated by 2 annotators. We release standard dataset of sentences after 100% agreement.

#### Who are the annotators?

The open-sourced dataset was annotated by 3 people with adequate knowledge of ESG investing and were fluent in English with previous exposure of analyzing financial documents.

## Considerations for Using the Data

The dataset can be used to investigate the transparency in sustainability intention of language mentioned in ESG disclosures of sustainable funds.

### Discussion of Biases

The data instances might cover languages from certain fund issuers (not all). It was extracted from randomly chosen prospectuses from the collected corpus.  
The dataset might be revised with broader coverage of prospectus language in future.

### Licensing Information

This work is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 Unported License. To view a copy of this license, visit http://creativecommons.org/licenses/by-nc-sa/3.0/.

If you are interested in commercial use of the data, please contact the following author for an appropriate license:
- [Abhijeet Kumar](mailto:abhijeet.kumar@fmr.com)

### Citation Information

[More Information Needed]

### Contributions

Thanks to [Nazia Nafis](https://www.linkedin.com/in/nazianafis/) and [Mayank Singh](https://www.linkedin.com/in/mayank-singh-43761b155/)  for contributing to the dataset creation process.

Any contribution or further research by the community are welcome.