hasancanbiyik/turkish-potentially-euphemistic-terms

Turkish Potentially Euphemistic Terms (PETs)

Dataset Description

This dataset contains 2,436 annotated instances across 70 Potentially Euphemistic Terms (PETs) in Turkish. Each instance presents a PET in natural context, labeled as either euphemistic (1) or non-euphemistic (0). The dataset supports the task of euphemism detection: given a sentence and a marked term, classify whether the term is used euphemistically or literally.

This is the expanded version of the Turkish PETs dataset originally introduced in Biyik et al. (2024), which contained approximately 40 PETs. The expansion adds 30 new PETs with examples collected from SketchEngine's Turkish Web 2020 Corpus (trTenTen20), bringing the total to 70 PETs across 9 semantic categories.

The dataset was used in the EACL 2026 paper:

When Semantic Overlap Is Not Enough: Cross-Lingual Euphemism Transfer Between Turkish and English Hasan Can Biyik, Libby Barak, Jing Peng, Anna Feldman Proceedings of the Second Workshop on Natural Language Processing for Turkic Languages (SIGTURK 2026), EACL 2026 arXiv:2602.16957

Dataset Structure

Fields

Field	Description
ID	Unique instance identifier
PET	The base form of the Potentially Euphemistic Term
VARIATION	The inflected/surface form of the PET as it appears in the text
CATEGORY	Semantic category of the PET (e.g., death, bodily functions)
ORIG_TEXT	Original raw text from the source corpus
CLEAN_TEXT	Lowercased and lightly normalized version of the text
CHAR_COUNT	Character count of the clean text
WORD_COUNT	Word count of the clean text
TEXT	Text with the PET span marked using [PET_BOUNDARY]...[PET_BOUNDARY] tags
LABEL	Binary label: 1 = euphemistic, 0 = non-euphemistic (literal)
EUPH_STATUS	Whether the PET is always_euph (used only euphemistically) or sometimes_euph (can be used both ways)

Label Distribution

	Count	%
Euphemistic (1)	1,457	59.8%
Non-euphemistic (0)	979	40.2%

Semantic Categories

Category	Instances	PETs	Euphemistic %
Death	750	21	61.5%
Bodily Functions	567	14	52.6%
Physical/Mental Attributes	285	11	58.9%
Sexual Activity	272	8	70.6%
Employment/Finances	205	7	62.9%
Miscellaneous	198	5	57.1%
Substances	83	1	51.8%
Illness	65	2	73.8%
Politics	11	1	45.5%

Euphemistic Status

• 11 PETs are always euphemistic (281 instances) — these terms are used exclusively in euphemistic contexts (e.g., vefat etmek, hayata gözlerini yummak, hakkın rahmetine kavuşmak).
• 59 PETs are sometimes euphemistic (2,155 instances) — these terms can appear in both euphemistic and literal contexts (e.g., aybaşı, kaybetmek, uyutmak).

Usage

from datasets import load_dataset

dataset = load_dataset("hasancanbiyik/turkish-potentially-euphemistic-terms")

Each instance in the TEXT field contains [PET_BOUNDARY] markers around the target term. To extract the PET span and its context:

import re

example = dataset["train"][0]
text = example["TEXT"]
pet_span = re.search(r'\[PET_BOUNDARY\](.*?)\[PET_BOUNDARY\]', text).group(1)
clean_text = text.replace("[PET_BOUNDARY]", "")

Example Instances

Euphemistic use (LABEL=1):

Geride bırakmaya hazırlandığımız 2013 yılı genelinde sanat, spor, siyaset, medya ve bilim dünyasından pek çok ünlü isim aramızdan ayrıldı.

("...many famous figures left us." — euphemism for death)

Non-euphemistic use (LABEL=0):

Gelenler olacağı gibi aramızdan ayrılanlar da olacak.

("Just as there will be newcomers, there will be those who leave us too." — literal use referring to player transfers)

Dataset Construction

The original 40 PETs from Biyik et al. (2024) were retained without modification; for these PETs, only instance counts were adjusted to achieve a more balanced euphemistic/non-euphemistic distribution. The remaining 30 PETs were added by collecting contextual examples from SketchEngine's trTenTen20 corpus across multiple semantic categories. Model performance remained stable after expansion (F1 change ≤ 0.04), validating the quality of the new data (see Table 2 in the paper).

Related Resources

• English PETs Dataset: Gavidia et al. (2022)
• Multilingual PETs (ZH, ES, YO): Lee et al. (2024)
• Code and Experiments: github.com/hasancanbiyik/PETs-investigating-LLMs-generalization-capabilities
• Original Turkish PETs (SIGTURK 2024): Biyik et al. (2024)

Citation

If you use this dataset, please cite both the EACL 2026 paper and the original SIGTURK 2024 paper:

@inproceedings{biyik-etal-2026-semantic,
    title = "When Semantic Overlap Is Not Enough: Cross-Lingual Euphemism Transfer Between {T}urkish and {E}nglish",
    author = "Biyik, Hasan Can and Barak, Libby and Peng, Jing and Feldman, Anna",
    booktitle = "Proceedings of the Second Workshop on Natural Language Processing for Turkic Languages (SIGTURK 2026)",
    year = "2026",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/2602.16957",
}

@inproceedings{biyik-etal-2024-turkish,
    title = "{T}urkish Delights: A Dataset on {T}urkish Euphemisms",
    author = "Biyik, Hasan and Lee, Patrick and Feldman, Anna",
    booktitle = "Proceedings of the First Workshop on Natural Language Processing for Turkic Languages (SIGTURK 2024)",
    year = "2024",
    publisher = "Association for Computational Linguistics",
    pages = "71--80",
    url = "https://aclanthology.org/2024.sigturk-1.9/",
}

License

This dataset is released under CC BY 4.0. The textual examples are sourced from SketchEngine's trTenTen20 corpus and publicly available web text.

Contact

For questions or issues, contact Hasan Can Biyik at biyikh1@montclair.edu.