Transcription quality转录质量报告

This page documents the failure modes of this edition, with measured data. Publishing our errors is not a disclaimer but a precondition for using this material responsibly — a reader needs to know which fields can be trusted and which must be checked against the original scan.

本页公开本项目的失败模式与实测数据。公开错误不是免责， 而是让这批材料可被负责任地使用的前提——使用者需要知道哪些字段可信、 哪些必须回原件核对。

Two quite different systems have read these pages, and they fail in opposite ways. Keeping them apart is the single most important thing to understand about this site.

这批档案先后被两套完全不同的系统读过，而它们的出错方式恰好相反。 分清这两者，是理解本站最重要的一件事。

It is worth being precise rather than dramatic. Measured across all 3,863 files (33 million word-tokens), 69.2% of the words the bnAsie OCR produces are recognisable English; the median file scores the same, and only 3% of files fall below 50%. It is poor, not useless — it usually captures most letters of a word and misses one or two.

The consequence is specific: enough to half-read a page, nowhere near enough to search one. enguirics for enquiries, revesleu for revealed — a human guesses, a search index cannot. Errors also compound across a phrase: if roughly one word in three is mangled, a two-word phrase survives intact about half the time and a three-word name about a third. Chinese-language enclosures were never recognised at all.

这里应当精确而不是渲染。全库 3,863 件、3,310 万个词形的实测结果： 那份原有 OCR 产出的词有 69.2% 是可识别的英文，中位数同样是 69.2%， 低于 50% 的卷宗只占 3%。它是不好，而不是不能用—— 通常能认出一个词的大部分字母，错一两个。

但后果很具体：够勉强读懂一页，远远不够检索。 enguirics 之于 enquiries、 revesleu 之于 revealed——人能猜出来，检索引擎不能。 误差在词组上还会叠加：若每三个词错一个，两词词组完整存活约一半， 三字人名只剩三分之一。至于夹带的中文件，则从未被识别过。

Conventional OCR fails visibly, as garbled characters. A language model confronted with a damaged page instead produces fluent, plausible, wrong text that reads perfectly. This has been systematically evaluated in the digital humanities and named over-historicization — the model invents in the direction of what it takes to "fit the period" (Levchenko, Evaluating LLMs for Historical Document OCR, LM4DH 2025, arXiv:2510.06743).

That study also found that post-hoc "OCR correction" by an LLM degrades rather than improves results. This project therefore transcribes once and applies no second polishing pass.

传统 OCR 出错会变成一眼可见的乱码；大模型面对残缺页面时， 会生成通顺、像样、但错误的内容，读起来毫无破绽。这一现象在数字人文领域已被 系统评测并命名为 over-historicization（过度历史化）——模型朝着它认为 「符合那个时代」的方向编造（Levchenko, LM4DH 2025, arXiv:2510.06743）。

该研究同时发现：事后用大模型做「OCR 纠错」会让结果变差。 因此本项目只做一次转录，不做二次润色。

Same page, same model, same prompt — only the resolution of the input image differs:

同一页、同一模型、同一提示词，只改变输入影像的分辨率：

At the lower resolution, a German Jewish refugee who ran an electroplating works was transcribed as running an opium and illegal lottery factory — wrong in a way that fits every cliché about old Shanghai, and therefore invisible. This edition transcribes at native scanning resolution throughout, at the cost of roughly 87% more input tokens.

低分辨率下，一名经营电镀厂的德籍犹太难民被转录成经营 鸦片与地下彩票厂——错得完全符合对旧上海的想象，因而看不出来。 本站一律使用原生扫描分辨率转录，代价是输入 token 增加约 87%。

In the 22-page file Report Re Wang Cheng-Hsiang, the same person's name appears in three different forms on three pages:

卷宗《Report Re Wang Cheng-Hsiang》（22 页）中， 同一个人的姓名在三页上出现了三个版本：

The correct form is 汪正祥, attested three times independently: the Chinese letter on p2, the file title Wang Cheng-Hsiang (正祥 = Cheng-Hsiang), and the handwritten annotations themselves. The model got it right once out of three.

正确写法为汪正祥，由三处独立佐证：p2 的中文正文、 卷宗标题 Wang Cheng-Hsiang（正祥＝Cheng-Hsiang）、以及手写批注本身。 模型三次只对了一次。

The case above points to a method that scales: a file that contradicts itself can be detected with no external reference at all. If one person's name appears in three forms within a single file, at least two of them are wrong.

Before comparing, the checker normalises glyph forms using an equivalence table for simplified / traditional / variant / cross-strait character standards (about 7,100 characters, merged from cjkvi-tables and a mainland–Hong Kong–Taiwan glyph comparison), so that differences such as 顧／顾 are not reported as errors.

上述案例指向一个可规模化的质检方法：同一卷宗内部的自相矛盾， 不需要任何外部参照就能自动查出。一份档案里同一个人名出现三种写法， 必有两种是错的。

比对前先用简繁／异体／两岸字形等价表归一字形（约 7,100 字， 由 cjkvi-tables 与两岸三地字形对照表合并）， 以免把「顧／顾」这类字形差异误报为错误。

Across the pages transcribed so far, the picture is lopsided in a useful way: connected English prose — the bulk of these reports — comes through well, and errors concentrate almost entirely in Chinese personal and place names, plus isolated numerals and handwriting. Reading a report to find out what happened is generally safe; taking a name, a number or a date from it is not.

就已转录的页面来看，问题的分布很不均匀，而这对使用者是好消息： 成句的英文叙述——也就是这批报告的主体——转录得相当好， 而错误几乎集中在汉字人名与地名，另有孤立的数字和手写部分。 读一份报告弄清「发生了什么」通常是安全的；从中取一个人名、一个数字、 一个日期则不然。

Temperature is set to 0, but the output is not reproducible. The same cursive signature transcribed twice yielded two different readings (陳希平 / 陳希曾): where the handwriting is genuinely illegible, the model guesses differently each time.

温度参数设为 0，但不保证可复现。实测同一页草书签名两次转录 得到不同结果（「陳希平」／「陳希曾」）——字迹本身即无法辨认时，模型每次的猜测不同。

This site offers no maps, no social-network graphs and no word-frequency analysis built on these transcriptions. The reason is established above: personal and place names are exactly the least reliable fields, and quantitative analysis built on them would hide that uncertainty behind apparently precise charts.

Once sampling verification is complete and the error rate is known, we will reassess which analyses can be supported.

本站不提供基于这批转录的地图、人物网络、词频统计等衍生分析。 理由是上文已经证明：人名与地名恰恰是最不可靠的字段， 在其上构建计量分析会把不确定性隐藏在看似精确的图表之后。

待抽样校验完成、错误率有据可依后，再评估哪些分析站得住。

M1750 is a selective filming: the NARA descriptive pamphlet states that the selection criteria are unknown, and the microfilm and the paper holdings do not fully overlap. What was not filmed remains, for the most part, among the paper records at NARA. The records are also simply incomplete: part of the archive was damaged or lost.

These are moreover the surveillance records of a colonial police force. Their characterisations of people, organisations and events carry the assumptions of the men who wrote them, and should be read as evidence rather than as statements of fact.

M1750 是选摄本——NARA 说明册明言其选择标准不为人知， 缩微胶卷与纸本收藏并不完全重合。未被拍摄的部分，大体仍在 NARA 的纸本档案里。 此外这批档案本身就不完整——有一部分已经损毁或散失。

此外，这是殖民地警察机构的政治监视记录，其中对人物、团体、 事件的描述带有书写者的立场与成见，应作为史料对待，而非事实陈述。