Can AI Read Classical Chinese?

Source Library holds about 1,600 Chinese-language works — pharmacology, mathematics, agriculture, military technology, natural history, the Confucian and Daoist classics. Most of them have never been translated into English. The obvious question is whether today’s AI can actually read and translate them. So we measured — not with a vibe check, but against external ground truth.

Two findings, up front. On clean printed text, our OCR reproduces the canonical characters at about 98.5% accuracy, measured against ctext.org, the leading digital library of classical Chinese. And on the same texts, our translations match or beat the standard scholarly English editions. The honest catch — the part a partner is needed for — comes later.

The collection

The ~1,600 Chinese works include 448 first English translations and about 105,000 translated pages. The center of gravity isn’t literature — it’s the history of Chinese science, technology, and medicine, with a Ming–Qing core and a Song–Yuan manuscript tail. A few of the marquee texts:

• Li Shizhen’s Bencao Gangmu (Compendium of Materia Medica, 1596) — the great pharmacopoeia of the Ming.
• Mao Yuanyi’s Wubei Zhi (Treatise on Armament Technology, 1621) — a vast military-technology compendium.
• Song Yingxing’s Tiangong Kaiwu (1637) — the encyclopedia of crafts and industry pictured above.
• The Sancai Tuhui (Illustrated Encyclopedia of the Three Realms, 1609).

You can browse the whole set in the Chinese Classics collection. Every work is a full facsimile with source provenance — which is what makes it possible to grade ourselves honestly.

Reading the page: OCR vs. ctext

To measure OCR you need a known-correct transcription to compare against. For classical Chinese, that exists: ctext.org has scholar-verified digital editions of the canon. So for a sample of canonical works, we compared our OCR of a real woodblock page against ctext’s text of the same passage, and counted character errors.

The first attempt produced nonsense — it scored the Book of Odes at 6% accuracy, which is obviously wrong. The reason is a feature of the corpus: most of our editions are commentary editions, where the canonical main text is printed interleaved with small-character annotation. ctext stores only the main text, so a naive character-by-character comparison treats every (correctly-read) commentary character as an error and reports a catastrophe where there isn’t one.

The fix is to match the canonical text as an in-order subsequence of our OCR, skipping the extra commentary characters for free — so the metric scores OCR error on the canonical text only. With that, the picture clears:

Character accuracy of our OCR vs. ctext.org canonical text, by subsequence alignment. Seven cleanly-aligned canonical works.

Chapter 1 of the Daodejing is a clean illustration. Our OCR of the woodblock page reads:

道可道，非常道。名可名，非常名。無名天地之始；有名萬物之母。故常無欲，以觀其妙；常有欲，以觀其徼。此兩者，同出而異名，同謂之玄。玄之又玄，衆妙之門。

That is character-for-character identical to ctext’s text of the chapter. The Book of Odes opening (Guan Ju) was the same but for a single character — our edition prints 鐘 (“bell”) where ctext’s base text has the variant 鍾, which is a legitimate textual variant, not a misread. The errors that remain cluster on a handful of harder characters; nothing approaching a failure mode.

Translating the page: vs. James Legge

Reading the characters is half the job. How good is the translation? The fair test is to compare against a published human translation. For the Chinese classics, the standard reference — and conveniently public domain — is James Legge’s 19th-century corpus. So we put them side by side.

Take the closing lines of Guan Ju, the first poem in the Book of Odes. Legge (1871):

“The modest, retiring, virtuous, young lady: With lutes, small and large, let us give her friendly welcome.”

Our translation of the same page:

“The modest, retiring, virtuous young lady; With lutes and zithers, we befriend her.”

The phrase is 琴瑟 — the qin and the se, two different zithers. Legge’s “lutes, small and large” blurs them into one instrument; “lutes and zithers” is the more accurate reading. It is a small thing, but it is the kind of small thing that tells you the model is construing the text, not paraphrasing a memory of it.

The opening of the Daodejing is sharper still. Legge:

“The Dao that can be trodden is not the enduring and unchanging Dao.”

Our translation of the page:

“The Way that can be spoken of is not the constant Way; the name that can be named is not the constant name. The nameless is the beginning of Heaven and Earth; the named is the mother of the ten thousand things.”

This isn’t Legge at all. 道可道 as “can be spoken of” (not Legge’s idiosyncratic “trodden”), 常 as “constant,” 萬物 as “the ten thousand things” — it tracks the modern scholarly idiom rather than the dated 1891 reading. And it did something Legge’s bare-text edition never attempted: it also translated the surrounding Heshang Gong commentary printed on the same page.

A note on method: the instinct is to score translation with BLEU or other n-gram overlap metrics. For literary translation that is actively misleading — Legge and Arthur Waley render the same poem with almost no shared phrasing, and both are correct. Two good translations can look maximally different to an n-gram metric. The honest test is a blind comparison: show a reader both renderings, unlabeled, and ask which is better. On clean canonical text, ours holds its own against the standard scholar.

Where it still fails

None of this means the corpus is finished and citable. Two failure modes are real, and worth naming plainly.

Complex layouts. On dense interlinear commentary and on multi-column tables, the model tends to translate the main text well and then quietly omit the rest. On a Bencao Gangmu page laid out as a grid of drug formulas, most of the prescriptions — ingredients, doses, indications — can simply be dropped, sometimes replaced with an editorial placeholder rather than a translation. The same is true of the OCR: the clean numbers above are for canonical printed pages, which is the easy case. Manuscripts, tables, and rare regional editions — the actual frontier — aren’t in ctext at all, and need different tooling (self-consistency, cross-model agreement, and human review) to measure.

First translations. The 448 works being rendered into English for the first time are, by definition, the ones with no published translation to check against. They are where the value is highest and the certainty is lowest — and where a human scholar’s eye is indispensable.

What this means

The pattern is the same one we found benchmarking AI OCR on Bhutanese manuscripts: the hard part is the page, not the language. On clean printed text the model reads classical Chinese at ~99% and translates it as well as the standard scholarly edition. What it struggles with is messy layout — commentary woven through the main text, formulas laid out as tables — and the works for which no reference exists to anchor it.

That is an encouraging place to be. It means the core capability is in hand, and the remaining work is the kind scholars are uniquely good at: catching the dropped formula, the commentary the model skipped, the term it construed a little too cleverly. These books — the roots of Chinese pharmacology, mathematics, and engineering — have sat unread outside China for centuries. The reading is now tractable. The careful finishing is what comes next, and it is work we would like to do with partners who know these texts.

The OCR benchmark is reproducible from the open evaluation framework in the Source Library repository (scripts/eval): ground truth is built from ctext, and scored with the subsequence aligner described above. Translation comparisons quote the exact page text, linked above.

Related notes

This is part of a running series on whether AI can read historical scripts:

• Benchmarking AI OCR on 232,000 Pages of Bhutanese Manuscripts — the same question for handwritten Tibetan.
• Can AI Read Cuneiform?
• Can AI Read Hieroglyphs? (No.)
• The Rashi Problem: When AI OCR Hallucinates in Hebrew
• How Consistent Is AI OCR? — the consistency metric, in depth.
• The Confident Hallucinator — how AI translation fails, and how to catch it.