Dataset Card for FinCorpus-DE10k

Dataset Details

Dataset Description

FinCorpus-DE10k is a corpus containing 12,235 PDF files of financial documents, mostly security prospectuses, along with plaintext files for approximately 10,500 of these documents. The documents are primarily in German (71%), with the rest being bilingual (German and English). This dataset aims to facilitate tasks like text analysis, language modeling, and document understanding in the financial domain.

This dataset is a subset of the above dataset, with the collections we felt comfortable releasing under permissive CC licenses. It omits the IFRS (containing 7 documents) and Informational_materials (127/129 txt/pdf files) collections. To get access to the full corpus, get in touch with us.

Curated by: Nata Kozaeva, Serhii Hamotskyi, Christian Hänig
Language(s) (NLP): German (DE), Bilingual (German and English)
License: CC BY-NC 4.0 except the monthly and annual reports, which are CC BY-NC-ND 4.0.

It's composed of multiple collections, with the text content available as dataset configs as:

Annual_reports
BBK_monthly
Base_prospectuses
Final_terms
Law

(By default, all collections are downloaded).

The entire corpus, pdf and txt files, can be downloaded here: https://huggingface.co/datasets/anhaltai/fincorpus-de-10k/resolve/main/data/corpus_safe.zip?download=true

Dataset Sources

The FinCorpus-DE10k dataset is composed of financial documents from various collections, each with its unique characteristics and source of origin. The documents were primarily sourced from the websites of financial institutions, regulatory bodies, and publicly available databases, with significant contributions from the Deutsche Bundesbank. The dataset includes:

Final Terms Prospectuses: These documents detail the terms and conditions of the issuance of financial securities, predominantly collected by the Deutsche Bundesbank. They form the largest part of the dataset, with documents ranging from 1 to 719 pages, but mainly under 100 pages.
Base Prospectuses: Containing information about the issuer, description of the security, and the summary of the prospectus. These documents are longer and fewer compared to the Final Terms but hold comprehensive information required for investors.
Bundesbank Monthly Reports: Consisting of 838 monthly reports from the German Bundesbank, spanning from 1949 to 2022. These documents offer a historical perspective on the German financial language. We didn't extract text from these documents. Licensed CC BY-NC-ND 4.0
Annual Reports: This collection includes annual (and some quarterly) reports from the Bundesbank and other institutions, covering economic and financial issues, monetary policy, and financial stability risks. Licensed CC BY-NC-ND 4.0
Law: Featuring German laws in the financial and related domains, including some English translations. This collection reflects the regulations applicable to the financial sector in Germany and EU Directives implemented into German law.

The collection as a whole is licensed CC BY-NC 4.0 except where stated otherwise.

Repository: https://github.com/AnhaltAI/fincorpus-de-10k-scripts/
Paper: TODO

Uses

Direct Use

By providing a rich collection of financial documents in PDF format, the dataset facilitates the development of algorithms that can navigate the complex layouts typically found in financial documents. FinCorpus-DE10k is also suited for developing and testing NLP models specialized in the financial domain, including but not limited to information extraction, named entity recognition, and specialized language models.

Dataset Structure

When used through load_dataset(), the dataset has two features: filename and text, one instance per .txt document.

The complete dataset, pdf and txt, can be found in corpus.zip. In the archive, metadata.csv contains the path for the PDF and its extracted .txt (if available), as well as collection name, presence of extracted text, paths to PDF and .txt files, document language(s), and financial identifiers like ISIN and country for relevant documents.

The pdf and txt subfolders contain the same mirrored directory structure, sorted by collection.

Dataset Creation

Source Data

Data Collection and Processing

Extensive preprocessing was applied to ensure the quality and uniformity of the dataset. It's described in our paper: TODO

Who are the source data producers?

The documents were produced by various financial institutions, regulatory bodies, companies, and the Deutsche Bundesbank.

Personal and Sensitive Information

The dataset contains financial documents that are publicly available.

Licensing

We diligently adhered to the licensing guidelines to the best of our understanding. However, the responsibility for the use of the documents and compliance with applicable laws rests with you.

Get in touch with us if any of the documents need to be removed from the collection.

Relevant links are:

Bundesbank monthly+annual allows using its documents if they are unchanged, hence the CC BY-NC-ND license: Nutzungsbedingungen - Für den allgemeinen Gebrauch der Website | Deutsche Bundesbank
German laws are public domain: Act on Copyright and Related Rights (Urheberrechtsgesetz – UrhG)
Final terms documents (and their Basisprospekte) can be considered public domain (at least their textual content), since the relevant EU regulation mandates they are published and freely accessible: L_2017168EN.01001201.xml

Citation

Temporary citation until paper is published:

@inproceedings{hamotskyi-etal-2024-fincorpus,
  title = {{{FinCorpus-DE10k}}: {{A}} Corpus for the German Financial Domain},
  booktitle = {The 2024 {{Joint International Conference}} on {{Computational Linguistics}}, {{Language Resources}} and {{Evaluation}} ({{LREC-COLING}} 2024)},
  author = {Hamotskyi, Serhii and Kozaeva, Nata and H{\"a}nig, Christian},
  year = {2024},
  month = may,
  publisher = {European Language Resources Association},
  address = {Torino, Italy},
  abstract = {We introduce a predominantly German corpus comprising 12.5k PDF documents sourced from the financial domain. The corresponding extracted textual data encompasses more than 165 million tokens derived predominantly from German, and to a lesser extent, bilingual documents.  We provide detailed information about the document types included in the corpus, such as final terms, base prospectuses, annual reports, information materials, law documents, international financial reporting standards, and monthly reports from the Bundesbank, accompanied by comprehensive statistical analysis. To our knowledge, it is the first non-email German financial corpus available, and we hope it will fill this gap and foster further research in the financial domain both in the German language and in multilingual contexts.}
}