Corpus
389 issues from the second British presidency in India, machine-read for the first time. The scans have been public for years. The text has not, because the OCR was run in the wrong language.
The only copies of the Madras Courier online are the Tamil Virtual Academy scans on the Internet Archive. They are good scans. The text layer attached to them is unusable, and the reason is recorded in the item metadata itself: ocr_parameters: -l tam. Tesseract was run in Tamil on an English newspaper.
Measured on the issue of 21 December 1803: the text layer holds 122,769 characters, of which 57,659 are Tamil codepoints and none form an English word of three letters or more. The page underneath is four columns of English letterpress carrying Arthur Wellesley's despatches from the Second Anglo-Maratha War over his own signature, general orders from Camp at Cheeje Kair, casualty returns, and a table of European officers late in Dowlut Rao Scindiah's service listed by nationality.
So the paper was searchable in principle and unfindable in practice. We re-read it with a vision model.
The paper is now read end to end — 2,248 pages, 33,786 items. Of those: 25.4% articles, 15.9% advertisements, 13.8% notices, 12.8% shipping, 9.5% proclamations, 6.6% despatches, 2.0% prices current. A government news-paper with a commercial back half. Its money is the money of the Coromandel coast, counted in pagodas (4,819 mentions), star pagodas (1,651) and rupees (1,382).
Every item carries a dateline and every issue a print date, so the gap between them is measurable. Medians with interquartile range, from the issues read so far:
| origin | median | IQR | n |
|---|---|---|---|
| London | 168 d | 140–185 | 208 |
| Paris | 173 d | 164–207 | 88 |
| Calcutta | 22 d | 20–30 | 119 |
| Bombay | 18 d | 14–24 | 78 |
| Fort St George (minute to print) | 8 d | 4–15 | 489 |
Source images: Tamil Virtual Academy, via the Internet Archive. The works are public domain by age. Fetching was serial, one request at a time with a two to five second gap.
The transcription is machine-made and is not a critical edition. Names, places, ships and sums extracted from each page were checked mechanically against that page and dropped where they did not appear, because the extracting model invented roughly one name in a hundred. Spelling is left as printed, so variants of one person's name are not merged.