Dataset Card for [EDGAR-CORPUS]

Dataset Summary

This dataset card is based on the paper EDGAR-CORPUS: Billions of Tokens Make The World Go Round authored by Lefteris Loukas et.al, as published in the ECONLP 2021 workshop.

This dataset contains the annual reports of public companies from 1993-2020 from SEC EDGAR filings.

There is supported functionality to load a specific year.

Care: since this is a corpus dataset, different train/val/test splits do not have any special meaning. It's the default HF card format to have train/val/test splits.

If you wish to load specific year(s) of specific companies, you probably want to use the open-source software which generated this dataset, EDGAR-CRAWLER: https://github.com/nlpaueb/edgar-crawler.

Citation

If this work helps or inspires you in any way, please consider citing the relevant paper published at the 3rd Economics and Natural Language Processing (ECONLP) workshop at EMNLP 2021 (Punta Cana, Dominican Republic):

@inproceedings{loukas-etal-2021-edgar,
    title = "{EDGAR}-{CORPUS}: Billions of Tokens Make The World Go Round",
    author = "Loukas, Lefteris  and
      Fergadiotis, Manos  and
      Androutsopoulos, Ion  and
      Malakasiotis, Prodromos",
    booktitle = "Proceedings of the Third Workshop on Economics and Natural Language Processing",
    month = nov,
    year = "2021",
    address = "Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.econlp-1.2",
    pages = "13--18",
}

Supported Tasks

This is a raw dataset/corpus for financial NLP. As such, there are no annotations or labels.

Languages

The EDGAR Filings are in English.

Dataset Structure

Data Instances

Refer to the dataset preview.

Data Fields

filename: Name of file on EDGAR from which the report was extracted.
cik: EDGAR identifier for a firm.
year: Year of report.
section_1: Corressponding section of the Annual Report.
section_1A: Corressponding section of the Annual Report.
section_1B: Corressponding section of the Annual Report.
section_2: Corressponding section of the Annual Report.
section_3: Corressponding section of the Annual Report.
section_4: Corressponding section of the Annual Report.
section_5: Corressponding section of the Annual Report.
section_6: Corressponding section of the Annual Report.
section_7: Corressponding section of the Annual Report.
section_7A: Corressponding section of the Annual Report.
section_8: Corressponding section of the Annual Report.
section_9: Corressponding section of the Annual Report.
section_9A: Corressponding section of the Annual Report.
section_9B: Corressponding section of the Annual Report.
section_10: Corressponding section of the Annual Report.
section_11: Corressponding section of the Annual Report.
section_12: Corressponding section of the Annual Report.
section_13: Corressponding section of the Annual Report.
section_14: Corressponding section of the Annual Report.
section_15: Corressponding section of the Annual Report.

import datasets

# Load the entire dataset
raw_dataset = datasets.load_dataset("eloukas/edgar-corpus", "full")

# Load a specific year and split
year_1993_training_dataset = datasets.load_dataset("eloukas/edgar-corpus", "year_1993", split="train")

Data Splits

Config	Training	Validation	Test
full	176,289	22,050	22,036
year_1993	1,060	133	133
year_1994	2,083	261	260
year_1995	4,110	514	514
year_1996	7,589	949	949
year_1997	8,084	1,011	1,011
year_1998	8,040	1,006	1,005
year_1999	7,864	984	983
year_2000	7,589	949	949
year_2001	7,181	898	898
year_2002	6,636	830	829
year_2003	6,672	834	834
year_2004	7,111	889	889
year_2005	7,113	890	889
year_2006	7,064	883	883
year_2007	6,683	836	835
year_2008	7,408	927	926
year_2009	7,336	917	917
year_2010	7,013	877	877
year_2011	6,724	841	840
year_2012	6,479	810	810
year_2013	6,372	797	796
year_2014	6,261	783	783
year_2015	6,028	754	753
year_2016	5,812	727	727
year_2017	5,635	705	704
year_2018	5,508	689	688
year_2019	5,354	670	669
year_2020	5,480	686	685

Dataset Creation

Source Data

Initial Data Collection and Normalization

Initial data was collected and processed by the authors of the research paper EDGAR-CORPUS: Billions of Tokens Make The World Go Round.

Who are the source language producers?

Public firms filing with the SEC.

Annotations

Annotation process

NA

Who are the annotators?

NA

Personal and Sensitive Information

The dataset contains public filings data from SEC.

Considerations for Using the Data

Social Impact of Dataset

Low to none.

Discussion of Biases

The dataset is about financial information of public companies and as such the tone and style of text is in line with financial literature.

Other Known Limitations

The dataset needs further cleaning for improved performance.

Additional Information

Licensing Information

EDGAR data is publicly available.

Shoutout

Huge shoutout to @JanosAudran for the HF Card setup!

References

• [Research Paper] Lefteris Loukas, Manos Fergadiotis, Ion Androutsopoulos, and, Prodromos Malakasiotis. EDGAR-CORPUS: Billions of Tokens Make The World Go Round. Third Workshop on Economics and Natural Language Processing (ECONLP). https://arxiv.org/abs/2109.14394 - Punta Cana, Dominican Republic, November 2021.

• [Software] Lefteris Loukas, Manos Fergadiotis, Ion Androutsopoulos, and, Prodromos Malakasiotis. EDGAR-CRAWLER. https://github.com/nlpaueb/edgar-crawler (2021)

• [EDGAR CORPUS, but in zip files] EDGAR CORPUS: A corpus for financial NLP research, built from SEC's EDGAR. https://zenodo.org/record/5528490 (2021)

• [Word Embeddings] EDGAR-W2V: Word2vec Embeddings trained on EDGAR-CORPUS. https://zenodo.org/record/5524358 (2021)

• [Applied Research paper where EDGAR-CORPUS is used] Lefteris Loukas, Manos Fergadiotis, Ilias Chalkidis, Eirini Spyropoulou, Prodromos Malakasiotis, Ion Androutsopoulos, and, George Paliouras. FiNER: Financial Numeric Entity Recognition for XBRL Tagging. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). https://doi.org/10.18653/v1/2022.acl-long.303 (2022)