讀字典

Reading Bonet

How a dictionary printed in 1899 became searchable, one character at a time.

951scanned pages
7,090headwords
4,035characters found

1The book

In 1899 Jean Bonet published in Paris his Dictionnaire annamite-français: two volumes, 951 pages, every word of Vietnamese in the Latin script, with its character in Chữ Nôm and its meaning in French. It shows how Vietnamese was written and read at the end of the 19th century, when Chữ Nôm was still in use beside Quốc Ngữ.

On a page, a reader sees all this at once. A computer sees only an image: it does not know which character is printed where. Searching the book for a character was impossible.

2The challenge: which one is it?

This is what the computer had to decide for every headword. The crop shows the headword and its character; below are the five characters that are most alike. Which one is printed?

All twelve are places where the program's first guess () was wrong and the right character was among its five; they were settled by reading the scan. Old print, look-alike characters, and characters missing from today's sets made thousands of places like these.

3The way

Page 520 as the program sees it:

Cut the pages

Every page is split into its two columns; a line that starts at the margin is a headword, the dark shape beside it its character.

7,090headwords found on 951 pages

Use the alphabet

The book is alphabetical, so a character on a page must have a reading between the first and the last headword of that page. Of tens of thousands of characters only a handful remain.

Bonet does not sort as dictionaries do today: a, ă, â are one letter, as are o, ô, ơ and u, ư; y stands with i, and đ comes before d. So ăn comes between ầm and anh. The program sorts the way the book does.

Compare the shapes

The crop is compared with a drawing of every candidate, stroke direction by stroke direction, and checked once more against more than 35,000 known characters: is another one even closer?

2,949placed for sure by the program

Read the rest

Where the program was not sure, the crop was read: with the candidates beside it, and wider, with the headword, when the character was rare.

3,030read from the crop

4The result

Every page of the book

One square per page, the darker the more of its characters are found. Click a page to open it.

0 %100 %

How reliable?

100 places that the program had marked as sure, picked at random and checked by eye:

right wrong probably wrong unclear

5What Bonet writes differently

Next to each character the reading was taken from the book itself: 1,506 places where several readings were possible were decided by hand on the scan. 123 readings are not in the input method of chunom.org for that character; 29 of them are syllables that no character there has. Some are old spellings, some readings that are no longer used:

All the new readings (123)

The reading in Bonet, the readings known so far, and the page; a few may be a character that the index matched wrongly: the page link shows the place.

6People and machines

Neither alone would have got this far. The program did the bulk and knew where it was unsure; an AI read the doubtful crops; a person decided the readings and checked what looked wrong.

2,949characters placed by the program, with a cross-check
3,030characters read from the crop by an AI (a large language model)
1,506readings decided by hand on the scan

7What is still open