Elamur Research⌕
☰
home / papers / 2026 / HOTATE: A Japanese Dialogue Corpus Annotated with Responses of Private Thoughts and Public Statements

HOTATE: A Japanese Dialogue Corpus Annotated with Responses of Private Thoughts and Public Statements

Summary

In Japanese conversation, speakers often separate honne, what they actually think, from tatemae, the socially smoothing version they say aloud. Mixing up the two can lead to misunderstandings. Earlier work covered indirectness, sarcasm and politeness, but tatemae differs because it changes what is said, not only how. No dataset targeted this contrast. The authors built HOTATE from 3,982 dialogues taken from two existing corpora: JDD (daily-life topics) and BSD (business). Crowdworkers stopped each dialogue at a chosen turn and wrote two alternative replies, one voicing the speaker's private thought and one giving a public statement. The authors then reviewed the results in two stages. Using the resulting 7,964 dialogues, they tested LLMs on two tasks:

  • labelling the final turn as one type or the other;
  • rewriting one type into the other.

Off-the-shelf Japanese LLMs essentially failed at labelling, but training on HOTATE improved accuracy substantially. In rewriting, producing honne from tatemae was harder than the reverse, according to both automatic metrics and human raters. The corpus offers a benchmark for socially aware dialogue understanding.

Key points

  • Size and sources: HOTATE contains 7,964 human-annotated dialogues, two variants for each of 3,982 source dialogues. The sources are JDD (Dailylife, School, Travel, Health, Entertainment) and BSD (Business). The corpus is released at github.com/EhimeNLP/HOTATE.
  • Annotation: 33 top-tier certified workers from the Lancers platform did the writing at ¥1,400 per hour. They followed guidelines requiring:
    • natural replies;
    • no new background information;
    • clearly distinct meanings within each pair.
  • Statistics: Dialogues average 7.26 utterances and about 15.5 words per utterance. The vocabulary covers 14,808 words, with word segmentation done by Sudachi. Business dialogues are longest, at roughly 18.6 utterances on average.
  • Classification results: Current Japanese LLMs could barely distinguish private thoughts from public statements. Fine-tuning on HOTATE brought a significant gain, and few-shot prompting also helped for most topics. gpt-oss-20b was an exception, improving only slightly over its 0-shot score.
  • Conversion results: Turning public statements into private thoughts was the harder direction under both automatic and human evaluation.

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

Fields: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/3q6pji5z9zn7

@inproceedings{toda2026hotate,
  title = {{HOTATE: A Japanese Dialogue Corpus Annotated with Responses of Private Thoughts and Public Statements}},
  author = {Toda, Yuko and Maekawa, Daisuke and Manabe, Kota and Yoneyama, Eito and Nonomura, Kanade and Fujiwara, Yuki and Kajiwara, Tomoyuki},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/3q6pji5z9zn7},
  url = {https://doi.org/10.63317/3q6pji5z9zn7}
}

References

19 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ibrahim Abu Farha, Steven Wilson, Silviu Oprea, and Walid Magdy. 2022. Sarcasm Detection Is Way Too Easy! An Empirical Comparison of Human and Machine Sarcasm Detection. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 5284–5295. find our card
  2. Penelope Brown and Stephen C. Levinson. 1987. Politeness: Some Universals in Language Usage. Studies in Interactional Sociolinguistics. find our card
  3. Kazuki Fujii, Taishi Nakamura, Mengsay Loem, Hiroki Iida, Masanari Ohi, Kakeru Hattori, Hirai Shota, Sakae Mizuki, Rio Yokota, and Naoaki Okazaki. 2024. Continual Pre-training for Cross-lingual Llm Adaptation: Enhancing Japanese Language Capabilities. In Proceedings of the First Conference on Langua… find our card
  4. LLM-jp. 2024. Llm-jp: A Cross-organizational Project for the Research and Development of Fully Open Japanese Llms. arXiv:2407.03963. — [arXiv:2407.03963](https://arxiv.org/abs/2407.03963) find
  5. Andrzej Maćkiewicz and Waldemar Ratajczak. 1993. Principal Components Analysis (Pca). Computers & Geosciences, 19(3):303–342. find
  6. Raden Regine Melansyah and Nuria Haristiani. 2020. Analysis of Japanese Refusal Speech Acts to an Invitation as a Tatemae. In Proceedings of the 3rd International Conference on Language, Literature, Culture, and Education (ICOLLITE 2019), pages 112–115. find
  7. OpenAI. 2025. Gpt-oss-120b & Gpt-oss-20b Model Card. arXiv:2508.10925. find our card
  8. Silviu Oprea and Walid Magdy. 2019. Exploring Author Context for Detecting Intended Vs Perceived Sarcasm. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2854–2859. find our card
  9. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. BLEU: A Method for Automatic Evaluation of Machine Translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318. find our card
  10. Matt Post. 2018. A Call for Clarity in Reporting Bleu Scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186–191. find our card
  11. Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence Embeddings Using Siamese Bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. find
  12. Nils Reimers and Iryna Gurevych. 2020. Making Monolingual Sentence Embeddings Multilingual Using Knowledge Distillation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing. find our card
  13. Matīss Rikters, Ryokan Ri, Tong Li, and Toshiaki Nakazawa. 2019. Designing the Business Conversation Corpus. In Proceedings of the 6th Workshop on Asian Translation, pages 54–61. find
  14. Anirudh Srinivasan and Eunsol Choi. 2022. TYDIP: A Dataset for Politeness Classification in Nine Typologically Diverse Languages. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 5723–5738. find our card
  15. Haruya Suzuki, Yuto Miyauchi, Kazuki Akiyama, Tomoyuki Kajiwara, Takashi Ninomiya, Noriko Takemura, Yuta Nakashima, and Hajime Nagahara. 2022. A JApanese Dataset for Subjective and Objective Sentiment Polarity Classification in Micro Blog Domain. In Proceedings of the Thirteenth Language Resources a… find
  16. Kazuma Takaoka, Sorami Hisamoto, Noriko Kawahara, Miho Sakamoto, Yoshitaka Uchida, and Yuji Matsumoto. 2018. Sudachi: A Japanese Tokenizer for Business. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018). find
  17. Junya Takayama, Tomoyuki Kajiwara, and Yuki Arase. 2021. Direct: Direct and Indirect Responses in Conversational Text Corpus. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1980–1989. find
  18. Benjamin Warner, Antoine Chaffin, Benjamin Clavié, Orion Weller, Oskar Hallström, Said Taghadouini, Alexis Gallagher, Raja Biswas, Faisal Ladhak, Tom Aarsen, Griffin Thomas Adams, Jeremy Howard, and Iacopo Poli. 2025. Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory E… find our card
  19. Deirdre Wilson. 2006. The Pragmatics of Verbal Irony: Echo or Pretence? Lingua, 116(10):1722–1743. Language in Mind: A Tribute to Neil Smith on the Occasion of his Retirement. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/toda-2026-hotate-japanese-dialogue-corpus-annotated/