- The original PDF files
- All the corresponding text files based on the structure of the PDF files
- All the text files, cleaned and merged in one single file
- If you encounter any difficulties in cloning the data from GitHub, you can also find all the merged and cleaned text files on Google Drive
- 💥 Two Language models as described in the paper trained with the following hyper-parameters:
hiddensize=2048
#layers=1
sequencelenght=250
minibatchsize=100
epochs=3
CoFiF is the first corpus comprising company reports in the French language. It contains over 188 million tokens in 2655 reports, covering four types of documents:
- Reference documents (documents de référence) published annually, usually in the months following the end of the calendar year, and contain information regarding the financial situation and perspectives of a company
- Annual report (résultats annuels) which summarises a company’s business and activities throughout the previous year
- Semestrial (résultats semestriels): similar to annual reports in content but published every 6 months
- Trimestrial reports (résultats trimestriels): similar to annual reports but published every 3 months
These documents are collected from the 60 largest French companies listed in France’s main stock indices CAC40 and CAC Next 20. The corpus spans over 20 years, ranging from 1995 to 2018.
In addition to the PDF files which were collected from enterprises (all rights reserved), we provide the reports in raw text without further pre-processing. The following image illustrates CoFiF corpus directories where file and folder names may contain metadata such as date of publication (year), type of document (DR: reference document, A: annual, S: semestrial, Q or T: trimestrial):
In addition to the PDF files and the raw text, we also provide a cleaned dataset CoFiF_cleaned_all.txt
which was used for training our language model reported in the paper. Find the language models and the merged cleaned text files at https://github.com/CoFiF/Corpus/tree/master/CoFiF.
For further information on how to use the language model, please check out https://github.com/flairNLP/flair/tree/master/resources/docs.
If you're using CoFiF in your researches, please don't forget to cite this paper:
@inproceedings{daudert-ahmadi-2019-cofif,
title = "{C}o{F}i{F}: A Corpus of Financial Reports in {F}rench Language",
author = "Daudert, Tobias and Ahmadi, Sina",
booktitle = "Proceedings of the First Workshop on Financial Technology and Natural Language Processing",
month = "12 " # aug,
year = "2019",
address = "Macao, China",
url = "https://www.aclweb.org/anthology/W19-5504",
pages = "21--26",
}
This corpus is openly available for non-commercial use under the Attribution-NonCommercial-ShareAlike 4.0 International.
If you have any questions or suggestions, you can contact Sina Ahmadi (sina.ahmadi@insight-centre.org) and Tobias Daudert (tobias.daudert@insight-centre.org).