license: apache-2.0
language:
- pl
pretty_name: Full-sentence Polish job title examples with gender-neutral equivalents
size_categories:
- 1K<n<10K
WARNING:
This version of the dataset is not proofread and it hasn't been fully normalized. It contains some annotation erroros that should be corrected before using it for training.
Dataset structure
Dataset has a simple structure. It contains CSV files with 3 columns each: type,source_sentence,target_sentence separated by commas.
There are 2 distinct example types (1 and 0) which are described below.
All of the examples have been shuffled at two distinct stages:
- after all the examples have been sorted by type, but before the splitting (to ensure random distribution of differnet sources over each split),
- after the data was split (to ensure random order of 0s and 1s witin each split).
Type 1
Those are the sentences that contain nouns designating job titles. In the [source_sentence] column, the job title is gendered. In the [target_sencence] column, the gendered job title has been replaced with so-called "personative" (see below).
An effort has been made to pick examples containing only one job title per sentence, preferebly in nominative case of the singular number, and to avoid including edge cases. However, those criteria are not met for every type 1 sentence in the set.
Semantic edge cases include personal nouns that are similar to job titles but are not job titles per se, eg. [terrorysta], [kolekcjoner], [anarchista], [zwolennik]. In general, they are nouns which describe people in relation to their activities, interests or political/buisness/organizational affiliations. Some semantic edge cases have been kept, eg. [kierowca]. Morphological edge cases include job titles that are very hard to neutralise, eg. [listonosz].
Type 0
Those are the sentences that do not contain any job titles. Both sentences in the same row ([source_sentence] and [target_sentence] columns) are exactly the same.
Sentences with job titles or semantic/morphological edge cases have been mostly removed or edited, so that they do not contain words such as:
- [gitarzyst(k)a], [pianist(k)a], [wokalist(k)a] and other names of musicians, hobbyists
- [partner(ka)], [członek/członkini], [przedstawiciel(k)a] and other names denoting people from the POV of political or business relationships
- [mieszkaniec/mieszkanka] and names denoting people by ethnicity, nationality or locality
Some nouns designating persons were left. Typye 0 sentences may still contain gendered and non-gendered personal nouns, such as:
- [kobieta], [mężczyzna]
- [matka], [ojcicec], [rodzic], [syn], [córka], [dziecko]
- [mąż], [żona], [małżonek], [małżonka]
- [osoba], [człowiek], [ludzie]
- [pan], [pani], [państwo]
- [bóg], [bogini], [bóstwo]
Splits
The dataset is split into 3 parts with the following amount of example sentences:
| SPLIT | type 1 | type 0 | TOTAL |
|---|---|---|---|
| train.csv | 610 | 4221 | 4831 |
| validation.csv | 50 | 500 | 550 |
| test.csv | 250 | 250 | 500 |
| TOTAL | 910 | 4971 | 5881 |
Sources
Type 1 examples come from the NKJP corpus enriched with a small amount targeted searches in Google (estimated amount of examples is 20-60). Some sentences have been altered so that better meet the criteria.
Type 0 examples mostly come from Wikipedia targeted scraping made with [WikipediaScraper.py] from [https://github.com/UW-IMLLA/week-7-ArielUW/tree/main/pipeline], as well as from 2 short novels Nasza szkapa and Kamizelka published in Wolnelektury. All of them have been manually reviewed to edit or delete sentences containing job titles or semantic/morphological edge cases. Some residual sentcences from NKJP have been added – they were irrelevant search reasults obtained while looking for type 1 examples with the use of regex.
SOURCE | type 1 | type 0 | TOTAL
Wikipedia | 0 | 3830 | 3830
Short novels | 0 | 1100 | 1100
NKJP | 850-890 | 41 | 891-931
Google | 20-60 | 0 | 20–60
Splitting source texts into example sentences
Most of the rows contain single sentences. However, some of the automatic or manual delimitations may contain errors. In other cases sentence delimitation may be controversial (especially for literary texts or spoken language transcripts).
Sentences from Wikipedia and from WolneLektury have been split into sentences using Sentence Splitter Python library: [https://github.com/mediacloud/sentence-splitter]. Sentence splitting of WolneLektury was additionally corrected manually. Sentences from NKJP have been split into sentences in majority using Spacy library (without corrections), while a small amount was split using an ad-hoc heuristic method combined with plenty of manual corrections. Sentences found with the help of Google Search Engine have been manually copied and did not need anny automated processing.