README.md

metadata

license: cc0-1.0
task_categories:
  - translation
language:
  - cs
  - de
  - es
  - pl

This dataset is the product of our work focused on the investigation of the development of the representation of gender forms, specifically male and female forms of occupations, in textual data.

We used the newscrawl dataset (Newscrawl doc-split data; WMT overview paper) for Czech, German, Spanish and Polish and employed clustering and filtering techniques (based on temporal and topic information) to extract a gender-balanced portion of the data. The dataset is therefore released under the same licence as newscrawl - the Creative Commons CC0 license ("no rights reserved").

We provide 2 types of files:

• {lang}_processed_data.json - postprocessed filtered data (removed links, very short sentences and sentences consisting only of numbers); json files containing a list of sentences
• {lang}_clusters_data.txt - original filtered data without postprocessing, txt files containing documents separated by 2 empty lines

The table illustrates the difference in the percentage of female occupations between the original data (highest achieved percentage in any file of the original newscrawl data) and the postprocessed filtered data. We also report the size of the filtered data.

Language	Original data (%)	Filtered data (%)	Filtered data size (sentences)
Czech	19.17	32.12	~2.9M
German	17.61	24.79	~28M
Spanish	33.11	40.77	~1.3M
Polish	15.20	27.53	~1.5M