Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Kanbun-LM: Reading and Translating Classical Chinese in Japanese Methods by Language Models

Hao Wang, Hirofumi Shimizu, Daisuke Kawahara

TL;DR

Kanbun-LM tackles the scarce Kanbun resources by constructing the first Classical-Chinese-to-Kanbun parallel dataset and defining two core tasks: character reordering to Japanese reading order and machine translation to Kanbun. A two-stage pipeline (reorder then translate) is evaluated using five reordering models and two translation models (mT5, mGPT), with both automatic metrics and human annotations guiding assessment. The study reports state-of-the-art results on both tasks, demonstrates that accurate preordering improves translation, and identifies BERTScore as the metric best aligned with human judgment while acknowledging limitations and proposing future dataset expansion and expert evaluation. This work advances Kanbun study by enabling quantitative evaluation and providing a practical translation pipeline, with potential impact on classical Japanese literature analysis and education.

Abstract

Recent studies in natural language processing (NLP) have focused on modern languages and achieved state-of-the-art results in many tasks. Meanwhile, little attention has been paid to ancient texts and related tasks. Classical Chinese first came to Japan approximately 2,000 years ago. It was gradually adapted to a Japanese form called Kanbun-Kundoku (Kanbun) in Japanese reading and translating methods, which has significantly impacted Japanese literature. However, compared to the rich resources for ancient texts in mainland China, Kanbun resources remain scarce in Japan. To solve this problem, we construct the first Classical-Chinese-to-Kanbun dataset in the world. Furthermore, we introduce two tasks, character reordering and machine translation, both of which play a significant role in Kanbun comprehension. We also test the current language models on these tasks and discuss the best evaluation method by comparing the results with human scores. We release our code and dataset on GitHub.

Kanbun-LM: Reading and Translating Classical Chinese in Japanese Methods by Language Models

TL;DR

Kanbun-LM tackles the scarce Kanbun resources by constructing the first Classical-Chinese-to-Kanbun parallel dataset and defining two core tasks: character reordering to Japanese reading order and machine translation to Kanbun. A two-stage pipeline (reorder then translate) is evaluated using five reordering models and two translation models (mT5, mGPT), with both automatic metrics and human annotations guiding assessment. The study reports state-of-the-art results on both tasks, demonstrates that accurate preordering improves translation, and identifies BERTScore as the metric best aligned with human judgment while acknowledging limitations and proposing future dataset expansion and expert evaluation. This work advances Kanbun study by enabling quantitative evaluation and providing a practical translation pipeline, with potential impact on classical Japanese literature analysis and education.

Abstract

Recent studies in natural language processing (NLP) have focused on modern languages and achieved state-of-the-art results in many tasks. Meanwhile, little attention has been paid to ancient texts and related tasks. Classical Chinese first came to Japan approximately 2,000 years ago. It was gradually adapted to a Japanese form called Kanbun-Kundoku (Kanbun) in Japanese reading and translating methods, which has significantly impacted Japanese literature. However, compared to the rich resources for ancient texts in mainland China, Kanbun resources remain scarce in Japan. To solve this problem, we construct the first Classical-Chinese-to-Kanbun dataset in the world. Furthermore, we introduce two tasks, character reordering and machine translation, both of which play a significant role in Kanbun comprehension. We also test the current language models on these tasks and discuss the best evaluation method by comparing the results with human scores. We release our code and dataset on GitHub.
Paper Structure (48 sections, 1 equation, 2 figures, 10 tables)

This paper contains 48 sections, 1 equation, 2 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Kaeriten
  3. Yomigana
  4. Okurigana
  5. Related Work
  6. Work for Classical Chinese
  7. Work for Kanbun
  8. Our Dataset and Tasks
  9. Experimental Setup
  10. Implementation for Tasks
  11. Pre-trained Models
  12. Models for Character Reordering
  13. BERT-japanese-char
  14. RoBERTa-japanese-char-wwm
  15. BERT-chinese
  16. ...and 33 more sections

Figures (2)

  • Figure 1: An example of Kanbun. "春眠不覚暁" (en: This morning of spring in bed I’m lying) is original Classical Chinese. To transform it into Kanbun, we first add Kaeriten, Yomigana, and Okurigana to the sentence. Two "レ" on the left side are Kaeriten, indicating the characters need to be read in reverse. On the right side, there is a Yomigana "ず", meaning "不" should be written as "ず". "エ" and "ヲ" are Okurigana, making the sentence from isolating to agglutinative. Now if we read the sentence following the above rules, the sentence becomes "春眠暁を覚えず" (While adding marks, we use Katakana like "エ" and "ヲ", but in a complete sentence, we use Hiragana like "え" and "を". They have no difference except for their looks).
  • Figure 2: An overview of the pipeline. (A) is the character reordering module and (B) is the machine translation module. (A) receives original Classical Chinese sentences and reorders them into Japanese reading order. (B) receives reordered sentences from (A) and translates them into Kanbun.