Elamur Research⌕
☰
home / papers / 2026 / Automatic Detection of Metaphorical Expressions in Classical Japanese Using WLSP-Enhanced BERT

Automatic Detection of Metaphorical Expressions in Classical Japanese Using WLSP-Enhanced BERT

Summary

Research on automatic metaphor detection has focused mostly on modern languages. Classical Japanese has received little attention, even though its vocabulary, grammar and figurative conventions differ from contemporary Japanese. The authors take a BERT-based detector that Zhu et al. (2025) first built for modern Japanese and apply it to medieval texts. The model follows MIP-style reasoning. It identifies a word's prototypical sense from WLSP basic-sense confidence annotations, which reflect modern Japanese. It then retrieves a typical usage example from BCCWJ-WLSP-auto and compares the encoding of the target in context with the encoding of that example. WLSP semantic codes are added to the input after [SEP]. Layered token type IDs mark the target, local context, semantic features and background. A weighted loss compensates for how rare metaphorical tokens are. Experiments use CHJ-Metaphor, a corpus still under construction that adds metaphor labels to CHJ-WLSP. Across three medieval works, 5-fold cross-validation gives an F1 of 82.18. Domain experts found that some of the model's apparent false positives were real metaphors the human annotators had missed. This suggests the system could help check annotation quality in large historical corpora.

Key points

  • Data: CHJ-Metaphor (85,102 tokens), built on CHJ-WLSP, which itself derives from the Corpus of Historical Japanese. It covers three works:

    • Konjaku Monogatari-shu: 79,892 tokens
    • Hojoki: 2,668 tokens
    • Toraakira-bon Kyogen-shu: 2,542 tokens

    The corpus will only be released publicly once annotation is finished.

  • Annotation scheme: follows BCCWJ-Metaphor, using MIP/MIPVU plus Nakamura's metaphor typology. Six categories are labelled: simile, metonymy, synecdoche, discourse-level metaphor, word-level metaphor, and "others", which includes makurakotoba.

  • Metaphor rates differ widely by work: 4.41% for Konjaku, 25.82% for Hojoki, 12.79% for the Kyogen collection, and 5.33% overall.

  • Results: overall F1 is 82.18 under 5-fold CV. Per-work accuracy is 98.41 (Konjaku), 92.28 (Hojoki) and 94.77 (Kyogen), and performance holds up across the different metaphor rates. The same approach previously reached an F1 of 75.07 on modern Japanese.

  • Limitations:

    • Only three of the six planned works are annotated so far.
    • Prototypical senses come from modern-Japanese WLSP data.
    • Whether the method transfers to non-medieval periods has not yet been tested.

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: Natural language processing · History & humanities

Topics Idioms & multiword expressions

 Institutions

🎓Tokyo University of Agriculture and TechnologyTokyo · Japan2 works · 24 citations🔬National Institute for Japanese Language and LinguisticsTokyo · Japan4 works · 2 citations

Also on the byline: Chuo University · Hokkaido University · National Institute of Technology, Nagano College.

Cite

Cite as Hang Zhu, Mitoki Ohara, Rei Kikuchi, Kanako Komiya, Masayuki Asahara and Sachi Kato. 2026. Automatic Detection of Metaphorical Expressions in Classical Japanese Using WLSP-Enhanced BERT. Proceedings of the Fourth Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA 2026) @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/2ptf9cfgpkgz

@inproceedings{zhu2026automatic,
  title = {{Automatic Detection of Metaphorical Expressions in Classical Japanese Using WLSP-Enhanced BERT}},
  author = {Hang Zhu, Mitoki Ohara, Rei Kikuchi, Kanako Komiya, Masayuki Asahara and Sachi Kato},
  booktitle = {Proceedings of the Fourth Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA 2026) @ LREC 2026},
  year = {2026},
  doi = {10.63317/2ptf9cfgpkgz},
  url = {https://doi.org/10.63317/2ptf9cfgpkgz}
}

References

7 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Baker, Collin F. and Fillmore, Charles J. and Lowe, John B. 2024. FrameNet. International Computer Science Institute. find
  2. Maekawa, Kikuo and Yamazaki, Makoto and Ogiso, Takehiko and Maruyama, Masaya and Ogura, Hideki and Kashino, Wakako and Koiso, Hiroshi and Yamaguchi, Makoto and Tanaka, Masahiro and Den, Yasuharu. 2014. Balanced Corpus of Contemporary Written Japanese. National Institute for Japanese Language and Lin… find
  3. Akira Nakamura. 1977. Theory and Classification of Figurative Expressions. Number 57 in Report of the National Institute for Japanese Language and Linguistics. Shuei Shuppan, Tokyo. In Japanese. find
  4. Gerard J. Steen, Aletta G. Dorst, J. Berenike Herrmann, Anna A. Kaal, Tina Krennmayr, and Tom Pasma. 2010. A Method for Linguistic Metaphor Identification: From MIP to MIPVU. John Benjamins Publishing. find
  5. National Institute for Japanese Language and Linguistics. 1964. Word List by Semantic Principles. Shuei Shuppan. find
  6. National Institute for Japanese Language and Linguistics. 2023. Corpus of Historical Japanese. National Institute for Japanese Language and Linguistics. find
  7. Xiaoyu Tong, Ekaterina Shutova, and Martha Lewis. 2021. Recent advances in neural metaphor processing: A linguistic, cognitive and social perspective. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhu-2026-automatic-detection-metaphorical-expressions-classical/