自動変換に向けた浅いデコーダを持つ 文字単位エンコーダ・デコーダモデルによる ニューラルかな漢字変換の高速化
DOI:
https://doi.org/10.51094/jxiv.7043キーワード:
ニューラルかな漢字変換、 自動変換、 ライブ変換、 日本語入力、 IME抄録
入力に追従して変換結果を更新する自動変換(macOS のライブ変換など)では,打鍵のたびにかな漢字変換を実行するため,変換 1 回あたりの時間(レイテンシ)が表示の遅延を決める.代表的なニューラルかな漢字変換モデルである zenz はデコーダのみの Transformer であり,自己回帰生成の 1 ステップごとに全 12 層を計算する.本稿では,読みと左文脈をエンコーダで 1 度だけ符号化し,2 層のデコーダで出力を生成する文字単位の T5(約 1.14 億パラメータ)を,zenz の作者が公開している学習データで学習した.学習データと出典が重ならない 9 種のコーパスから作成した 3,253 項目の評価セットにおいて,このモデルの正解率(acc@1)は 68.3%(f16)であり,zenz-v3.1-small(67.8%),zenz-v3.2-small,および azooKey の Zenzai との差はいずれも統計的に有意でなかった.一方,Apple M2 と M4 上で同一の推論エンジン(llama.cpp)を用い,読みの長さを 1〜80 文字の範囲で変えて変換時間を測定したところ,10〜80 文字の各長さにおける zenz-v3.1-small に対する速度比の中央値は,f16 で 1.6〜1.8 倍,Q5_K_M で 1.4 倍であった.読みが 40 文字のとき,Apple M4(f16)での変換時間は 44.9ms(zenz-v3.1-small は 81.2ms)であった.利益相反に関する開示
特になしダウンロード *前日までの集計結果を表示します
引用文献
azooKey. AzooKeyKanaKanjiConverter. https://github.com/azooKey/AzooKeyKanaKanjiConverter.
三輪敬太, 高橋直希. ニューラルかな漢字変換システム Zenzai. 言語処理学会第31回年次大会発表論文集, pp. 271–276, 2025.
Keita Miwa. zenz-v3.1-small / zenz-v3.2-small. https://huggingface.co/Miwa-Keita.
JIS X 4064:2002 仮名漢字変換システムの基本機能. 日本規格協会, 2002.
小林龍生. 日本の情報処理技術の足跡:漢字・日本語処理技術の発展:仮名漢字変換技術. 情報処理, Vol. 43, No. 10, pp. 1099–1103, 2002.
Armin Sarhangzadeh and Taro Watanabe. Alignment-based decoding policy for low-latency and anticipation-free neural Japanese input method editors. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Findings of the Association for Computational Linguistics: ACL 2024, pp. 8043–8054, Bangkok, Thailand, August 2024. Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.findings-acl.479
Topi Kaaresoja, Stephen Brewster, and Vuokko Lantz. Towards the temporally perfect virtual button: Touch-feedback simultaneity and perceived quality in mobile touchscreen press interactions. ACM Transactions on Applied Perception, Vol. 11, No. 2, pp. 1–25, June 2014. https://doi.org/10.1145/2611387
Jungo Kasai, Nikolaos Pappas, Hao Peng, James Cross, and Noah A. Smith. Deep encoder, shallow decoder: Reevaluating non-autoregressive machine translation. In International Conference on Learning Representations, 2021.
azooKey. AJIMEE-Bench. https://github.com/azooKey/AJIMEE-Bench.
森信介, 土屋雅稔, 山地治, 長尾真. 確率的モデルによる仮名漢字変換. 情報処理学会論文誌, Vol. 40, No. 7, pp. 2953–2964, 1999.
Google. Mozc. https://github.com/google/mozc.
Yoh Okuno. Neural IME: Neural Input Method Engine. https://github.com/yohokuno/neural_ime, 2016.
Jiali Yao, Raphael Shu, Xinjian Li, Katsutoshi Ohtsuki, and Hideki Nakayama. Enabling real-time neural IME with incremental vocabulary selection. In Anastassia Loukina, Michelle Morales, and Rohit Kumar, editors, Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Industry Papers), pp. 1–8, Minneapolis, Minnesota, June 2019. Association for Computational Linguistics. https://doi.org/10.18653/v1/N19-2001
A. Radford, Jeff Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever. Language models are unsupervised multitask learners. Technical report, OpenAI, 2019.
Kikuo Maekawa, Makoto Yamazaki, Toshinobu Ogiso, Takehiko Maruyama, Hideki Ogura, Wakako Kashino, Hanae Koiso, Masaya Yamaguchi, Makiro Tanaka, and Yasuharu Den. Balanced corpus of contemporary written Japanese. Language Resources and Evaluation, Vol. 48, No. 2, pp. 345–371, December 2013. https://doi.org/10.1007/s10579-013-9261-0
田中慎太郎, 鈴木海友, 松澤智史. BlindX におけるかな漢字変換 LLM の実装と評価. ソフトウェア・シンポジウム 2025 in 岡山, 2025. https://www.sea.jp/ss2025/download/2-ss2025.pdf.
Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, Vol. 21, No. 140, pp. 1–67, 2020.
Noam Shazeer. GLU variants improve transformer, 2020. arXiv:2002.05202.
ggml-org. llama.cpp. https://github.com/ggml-org/llama.cpp.
Keita Miwa. zenz-v2.5-dataset. https://huggingface.co/datasets/Miwa-Keita/zenz-v2.5-dataset.
Yu Tanaka, Yugo Murawaki, Daisuke Kawahara, and Sadao Kurohashi. Building a Japanese typo dataset from Wikipedia’s revision history. In Shruti Rijhwani, Jiangming Liu, Yizhong Wang, and Rotem Dror, editors, Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, pp. 230–236, Online, July 2020. Association for Computational Linguistics. https://doi.org/10.18653/v1/2020.acl-srw.31
LLM-jp. LLM-jp Corpus v4. https://gitlab.llm-jp.nii.ac.jp/datasets/llm-jp-corpus-v4.
Sanae Yamashita, Koji Inoue, Ao Guo, Shota Mochizuki, Tatsuya Kawahara, and Ryuichiro Higashinaka. RealPersonaChat: A realistic persona chat corpus with interlocutors' own personalities. In Chu-Ren Huang, Yasunari Harada, Jong-Bok Kim, Si Chen, Yu-Yin Hsu, Emmanuele Chersoni, Pranav A, Winnie Huiheng Zeng, Bo Peng, Yuxi Li, and Junlin Li, editors, Proceedings of the 37th Pacific Asia Conference on Language, Information and Computation, pp. 852–861, Hong Kong, China, December 2023. Association for Computational Linguistics.
Taro Tsuda, Sanae Yamashita, Koji Inoue, Tatsuya Kawahara, and Ryuichiro Higashinaka. Constructing a multi-party conversational corpus focusing on interlocutor relationships. In Proceedings of the 29th Workshop on the Semantics and Pragmatics of Dialogue, pp. 193–202, 2025.
Takashi Kodama, Ribeka Tanaka, and Sadao Kurohashi. Construction of hierarchical structured knowledge-based recommendation dialogue dataset and dialogue system. In Song Feng, Hui Wan, Caixia Yuan, and Han Yu, editors, Proceedings of the Second DialDoc Workshop on Document-grounded Dialogue and Conversational Question Answering, pp. 83–92, Dublin, Ireland, May 2022. Association for Computational Linguistics. https://doi.org/10.18653/v1/2022.dialdoc-1.9
Fuka Narita, Shiki Sato, Ryoko Tokuhisa, and Kentaro Inui. Construction and evaluation of news commentary chat corpus. Journal of Natural Language Processing, Vol. 31, No. 3, pp. 1015–1048, 2024. https://doi.org/10.5715/jnlp.31.1015
Nobuhiro Ueda, Hideko Habe, Akishige Yuguchi, Seiya Kawano, Yasutomo Kawanishi, Sadao Kurohashi, and Koichiro Yoshino. J-CRe3: A Japanese conversation dataset for real-world reference resolution. In Nicoletta Calzolari, Min-Yen Kan, Veronique Hoste, Alessandro Lenci, Sakriani Sakti, and Nianwen Xue, editors, Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pp. 9489–9502, Torino, Italia, May 2024. ELRA and ICCL.
Kazuma Takaoka, Sorami Hisamoto, Noriko Kawahara, Miho Sakamoto, Yoshitaka Uchida, and Yuji Matsumoto. Sudachi: a Japanese tokenizer for business. In Nicoletta Calzolari, Khalid Choukri, Christopher Cieri, Thierry Declerck, Sara Goggi, Koiti Hasida, Hitoshi Isahara, Bente Maegaard, Joseph Mariani, Hélène Mazo, Asuncion Moreno, Jan Odijk, Stelios Piperidis, and Takenobu Tokunaga, editors, Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan, May 2018. European Language Resources Association (ELRA).
Awni Hannun, Jagrit Digani, Angelos Katharopoulos, and Ronan Collobert. MLX: Efficient and flexible machine learning on Apple silicon. https://github.com/ml-explore, 2023.
Quinn McNemar. Note on the sampling error of the difference between correlated proportions or percentages. Psychometrika, Vol. 12, No. 2, pp. 153–157, June 1947. https://doi.org/10.1007/BF02295996
ダウンロード
公開済
投稿日時: 2026-11-02 01:34:56 UTC
公開日時: 2026-11-16 06:58:09 UTC
ライセンス
Copyright(c)2026
馬場, 哲晃
この作品は、Creative Commons Attribution 4.0 International Licenseの下でライセンスされています。
