Elamur Research⌕
☰
home / papers / 2026 / Challenges in Japanese Euphemism Classification: An Analysis of Pretrained Japanese and Multilingual Models

Challenges in Japanese Euphemism Classification: An Analysis of Pretrained Japanese and Multilingual Models

Paper · 2026 · Noriko Takahashi, Whitney Poh, Libby Barak et al.

Summary

Euphemisms are difficult for NLP because whether a phrase is meant literally or euphemistically depends on context, social norms and pragmatic inference. Datasets built around Potentially Euphemistic Terms (PETs) already exist for English, Turkish and other languages. Japanese had no such resource, even though its euphemisms interact closely with honorifics, register and script choice. The authors build JP-PET from sentences in BCCWJ, retrieved via Shonagon. Three native speakers labeled each PET occurrence as literal or euphemistic, and the data carries domain, register and orthography metadata. PET spans are detected with Sudachi plus custom rules and wrapped in [PET_BOUNDARY] markers. Two Japanese models (Rinna RoBERTa, Tohoku BERT) and multilingual XLM-R are then fine-tuned for binary classification. They are evaluated on PET-level splits that either allow or forbid the same PETs in training and test data. Results are good for familiar PETs but drop considerably for unseen ones, which suggests the models lean on memorized lexical items rather than contextual reasoning. JP-PET is the first benchmark for this pragmatic task in Japanese.

Key points

  • JP-PET dataset: 1,672 annotated sentences, ~100 PETs (the abstract says 101, the methods section says 100), and ten semantic domains such as Death & Dying, Bodily Functions and Politics & War.
    • Register is coded as formal, neutral or informal.
    • Each PET has on average 16.7 sentences (range 3–68), with a mean ambiguity entropy of 0.45 bits.
  • Japanese-specific phenomena: the dataset covers orthographic softening, such as writing 障害 as 障がい or using katakana ブツ for contraband. It also includes mimetic and phonologically altered forms, such as ぽっちゃり and ボンビー.
  • Evaluation splits: the methods section and appendix describe two splits, one with lexical overlap and one with no PET overlap, while the abstract mentions three.
  • Results for "sometimes-euphemistic" PETs (appendix):
    • With overlap, XLM-R scores best: F1 of 0.84 for the euphemistic label and 0.80 for the literal label.
    • Without overlap, euphemistic-label F1 falls to about 0.63–0.68, and literal-label F1 falls to 0.40–0.56 (lowest for Rinna).
    • For always-euphemistic PETs, euphemistic-label F1 stays at about 0.88–0.93.
  • Error analysis and limitations: most errors involve expressions that have become conventional in political discourse, euphemisms rooted in metaphor, and softening achieved through spelling or script choice. The authors note that the small dataset and its uneven domain balance limit firm conclusions about individual domains or registers.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Montclair State UniversityMontclair · United States5 works · 94 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSocial Sciences · 1,772 worksCommunication · 97 works

🧭 Topics 1Research subjects and languages the work is about.Multilingual & cross-lingual NLP · 828 works

Cite

Cite as Noriko Takahashi, Whitney Poh, Libby Barak, Jing Peng and Anna Feldman. 2026. Challenges in Japanese Euphemism Classification: An Analysis of Pretrained Japanese and Multilingual Models. Proceedings of Learning Non-Literal Expressions with Small Data @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/234ssv2bkqd2

@inproceedings{takahashi2026challenges,
  title = {{Challenges in Japanese Euphemism Classification: An Analysis of Pretrained Japanese and Multilingual Models}},
  author = {Noriko Takahashi, Whitney Poh, Libby Barak, Jing Peng and Anna Feldman},
  booktitle = {Proceedings of Learning Non-Literal Expressions with Small Data @ LREC 2026},
  year = {2026},
  doi = {10.63317/234ssv2bkqd2},
  url = {https://doi.org/10.63317/234ssv2bkqd2}
}

References

9 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Patrick Lee, A. Chirino Trujillo, D. Cuevas Plancarte, Olumide Ojo, Xinyi Liu, Iyanuoluwa Shode, Yuan Zhao, Anna Feldman, and Jing Meds for pets: Multilingual Peng. 2024. euphemism disambiguation for potentially euphemistic terms. In Findings of the Association for Computational Linguistics: EACL 20… find
  2. Keith Allan and Kate Burridge. 1991. Euphemism and dysphemism: Language used as shield and weapon. Oxford University Press. find
  3. Keith Allan and Kate Burridge. 2006. Forbidden words: Taboo and the censoring of language. Cambridge University Press, Cambridge. find
  4. Douglas Biber and Susan Conrad. 2019. Register, genre, and style, 2 edition. Cambridge University Press, Cambridge. find
  5. Patrick Lee, Anna Feldman, and Jing Peng. 2022. A report on the euphemisms detection shared task. In Proceedings of the 3rd Workshop on Figurative Language Processing, pages 184–190. Association for Computational Linguistics. find
  6. Hasan C. Biyik, Patrick Lee, and Anna Feldman. 2024. Turkish delights: A dataset on turkish euphemisms. In Proceedings of the First Workshop on Natural Language Processing for Turkic Languages (SIGTURK 2024), pages 71–80, Bangkok, Thailand and Online. Association for Computational Linguistics. find
  7. Patrick Lee, Iyanuoluwa Shode, Alain Trujillo, Yuan Zhao, Olumide Ojo, Diana Plancarte, Anna Feldman, and Jing Peng. 2023. FEED PETs: Further experimentation and expansion on the disambiguation of potentially euphemistic A. terms. In Proceedings of the 12th Joint Conference on Lexical and Computatio… find
  8. Emmy Liu, Chenxuan Cui, Kenneth Zheng, and Graham Neubig. 2022. Testing the ability of language models to interpret figurative language. In Proceedings of NAACL-HLT 2022, pages 4437– 4452. Association for Computational Linguistics.
  9. Per-PET statistics Statistic Sentences per PET Euphemistic instances Non-euphemistic instances PET ambiguity entropy (bits) Tokens per sentence † Lexical density † Senko K. Maynard. 1997. Japanese communication: Language and thought in context. University of Hawai’i Press. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/takahashi-2026-challenges-japanese-euphemism-classification-analysis/