Reading a book in translation

This guide takes a scanned book from upload to a working side-by-side reader. Active time is a few minutes; the model work in between depends on how long the book is.

You will need a dataset with a PDF in it that has already been through extraction — if you have not done that yet, start with your first dataset.

Step 1 — Open the file as a book

In your dataset, go to Books. The page lists the books you already have and, below them, the files that could become one.

Pick the source language before you press Open as book. This matters more than it looks: the language determines which romanisation scheme is used, how sentences are split, which font the text renders in, and whether the reader lays out right to left. Getting it right up front saves regenerating the text later.

Opening a book does not re-read your scans. It projects the text the dataset already extracted into pages, so the book and the corpus can never disagree about what a page says.

Step 2 — Run the translation

The book page shows the tools available and, next to each, what running it would cost for this book — roughly two model calls per page for translation. A three-hundred-page book is around six hundred calls, and you should see that before pressing the button rather than afterwards.

Press Run on the translate tool. Progress appears per page. You can leave the page; the work continues.

For a language written in a non-Latin script, this produces two columns: a romanisation and a translation. For a French or German source there is nothing to romanise, so you get the translation only and the reader hides the empty column rather than showing you the source twice.

Step 3 — Read

The reader gives you four columns — the scan, the source, the romanisation and the translation — and you choose which to show. Hiding one widens the rest.

Four presets cover most work:

• Reading — scan, romanisation and a wide translation.
• Translating — scan, source and translation, equally weighted.
• Linguistic — all four.
• Scan only — just the image.

Hover a paragraph and the same paragraph rings in every other column. That works because each region of the page becomes exactly one paragraph everywhere, so the columns are aligned rather than merely adjacent.

Arrow keys turn pages.

Every derived column names the model that produced it, and the romanisation column names its scheme. If you are going to quote from this, quote with those.

Step 4 — Find the chapters and the people

Two more tools, and they run in order because the order matters.

Find chapters reads the table of contents and then locates each title by searching the body for where it actually begins. Printed page numbers are not trusted — front matter shifts things, so printed page 37 is rarely the thirty-seventh scan. Books without a table of contents fall back to checking pages directly.

Find people and places extracts entities per chapter. Doing it per chapter rather than per book is the point: in an anthology, the Rama of one story is not the Rama of another, and merging them gives you a count that describes nothing.

A model decides what the entities are. It is never asked where they appear — models asked for character positions produce plausible ones that are wrong — so every mention is located by exact search instead. Click one and you land on the passage. A name the model suggested that appears nowhere literally is discarded rather than shown to you.

You do not need to run these in sequence by hand. Asking for entities on a book that has not been translated will translate it, find its chapters, and then extract.

If the extraction was wrong

Fix it in the document review view, then press Reproject on the book. Only the pages whose text actually changed are rebuilt, and re-running the translate tool re-translates only those. The rest of the book — including anything you have annotated — is left alone.

What to be careful about

A translation is a machine’s reading of a text, not the text. It is labelled with its model so you can weigh it; see a translation is not the source.

For Chinese, Japanese and Thai, keyword search uses character n-grams because those scripts do not put spaces between words. It works, but it is not the word segmentation a dedicated tool for those languages would give you.