Elamur Research⌕
☰
home / papers / 2026 / LLM-based Literal Example Generation for Japanese Multiword Expressions

LLM-based Literal Example Generation for Japanese Multiword Expressions

Paper · 2026 · Mio Ohashi, Hajime Kiyama, Zhidong Ling et al.

Summary

Many Japanese multiword expressions (MWEs), such as te o nobasu ("extend one's hand" vs. "reach out/expand into"), can be read literally or idiomatically. Corpora are dominated by the idiomatic readings, and Japanese free word order lets literal uses appear split apart. Literal examples are therefore rare and hard to extract, which makes it difficult to test whether models actually understand MWE meaning. The authors examine whether GPT-5 can supply such literal examples.

They select MWEs that appear in both the JMWEL lexicon and JParaCrawl. Using a heuristic, they split Nihon Kokugo Daijiten senses into literal and non-literal ones. GPT-5 is then prompted under five conditions that vary which hints it receives. Gemini-2.5-pro judges the outputs for literalness, and humans annotate a sample.

Giving GPT-5 contrastive non-literal information made literal generation more stable and of higher quality than literal-only or hint-free prompts. Gemini agreed with humans more on idiomatic readings than on literal ones. The work yields a contrastive sentence-level resource and a controlled framework for probing MWE understanding.

Key points

  • Generation setup: GPT-5 (gpt-5-2025-08-07) produced five literal sentences per target MWE. Targets came from the overlap of JMWEL and JParaCrawl. NKD (2nd edition) sense descriptions were available for 449 candidate expressions.
  • Sense splitting: Definition segments introduced by cues like no yoo ni or koto kara were treated as non-literal. Senses that keep the components' compositional meaning were treated as literal.
  • Prompting conditions: There were five conditions: Dn + Ex (non-literal definition plus non-literal corpus example), Ex, Dn, Dl (literal definition only) and N/A (surface form only). Dl proved markedly less reliable and often yielded pragmatically odd or ambiguous sentences.
  • Evaluation: Gemini-2.5-pro served as LLM-as-a-judge, complemented by sampled human annotation (Q1–Q3) and manual inspection. The judge modeled idiomatic interpretations better than literal ones.
  • Limitations: The study uses only one proprietary generator, and its scores reflect relative trends rather than absolute correctness. Models may also rely on superficial cues, for instance which case marker comes before te o hiku (no versus kara).

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

🎓Hitotsubashi UniversityTokyo · Japan6 works · 28 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 worksPsychology · 1,552 worksDevelopmental and Educational Psychology · 499 works

🧭 Topics 1Research subjects and languages the work is about.Idioms & multiword expressions · 217 works

Cite

Cite as Mio Ohashi, Hajime Kiyama, Zhidong Ling and Mamoru Komachi. 2026. LLM-based Literal Example Generation for Japanese Multiword Expressions. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-srw.25

@inproceedings{ohashi2026llm,
  title = {{LLM-based Literal Example Generation for Japanese Multiword Expressions}},
  author = {Mio Ohashi, Hajime Kiyama, Zhidong Ling and Mamoru Komachi},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-srw.25},
  url = {https://doi.org/10.18653/v1/2026.acl-srw.25}
}

References

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

  1. Timothy Baldwin and Su Nam Kim. 2010. Multiword Expressions. In Nitin Indurkhya and Fred J. Damerau, editors, Handbook of Natural Language Processing, pages 267–292. Chapman and Hall/CRC. find
  2. Pierluigi Cassotti and Nina Tahmasebi. 2025. Sense-specific Historical Word Usage Generation. Transactions of the Association for Computational Linguistics, 13:690–708. find
  3. Paul Cook, Afsaneh Fazly, and Suzanne Stevenson. 2008. The VNC-Tokens Dataset. find
  4. Hessel Haagsma, Johan Bos, and Malvina Nissim. 2020. MAGPIE: A large corpus of potentially idiomatic expressions. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 279–287, Marseille, France. European Language Resources Association.
  5. Chikara Hashimoto and Daisuke Kawahara. 2008. Construction of an Idiom Corpus and its Application to Idiom Identification based on WSD Incorporating Idiom-Specific Features. In Proceedings of the 2008 Conference on Empirical Methods in Natural Kosho Shudo, Akira Kurahone, and Toshifumi Tanabe. 2011.… find
  6. Kosho Shudo, Toshifumi Tanabe, and Masahito Takahashi. 2018. Overview and Current Status of the Japanese Multiword Expression Lexicon JMWEL: Focusing on verbal multiword expressions. In Proceedings of the 2018 Workshop on Language Resource Utilization, pages 601–610. National Institute for Japanese… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ohashi-2026-llm-based-literal-example-generation/