Reading an 1871 dictionary with three OCR engines

Hitchcock’s Interpreting Dictionary of Scripture Proper Names fills ten pages at the back of Hitchcock’s New and Complete Analysis of the Holy Bible (1871): three columns to a page, one name and its meaning to a line. Boaz, or Booz, in strength. The app uses it to explain the names you meet in the text.

There’s already a well-known e-text of it. I wanted our own copy, read from the printed page, that we could check line by line and stand behind.

Three readers

Each column is cut into halves, and each half is read three times:

• glm-ocr, a small vision model running locally on my Mac through Ollama;
• Apple Vision, the text recognition built into macOS, with its language correction turned off so it reports what it sees;
• the Internet Archive’s own OCR of the same scan.

At first the rule was simple. glm’s reading is the base, and wherever Apple and the Archive agree on something different, take their version.

Two readers, one mistake

That rule turned out to be wrong more often than right. Old type wears out. A semicolon’s tail fades, and both machines see a colon. A worn u prints like an n, so both read “house” as “honse”. A word split across two lines (“smit- … ing of his son”) gets rejoined by glm and left in pieces by the other two. In each case glm, which reads more like a person does, had it right, and the vote overruled it.

So the rule changed: a correction is only taken if it produces a real word. It’s the rule a human proofreader uses without thinking about it.

Checking against the page

After the change, 2,341 of the 2,390 names that appear in both our copy and the earlier e-text read exactly the same. I checked every remaining difference against the scan, zoomed in:

• Where we were wrong, glm misread a few names. Zibcon is Zibeon, and roof had been read as root.
• Where the print itself is damaged, the name is certain from the King James text. Chnza is Chuza (Luke 8:3) and Zcr is Zer.
• Where the print is simply wrong, we fix it and say so: “release; per-don”.
• Where the earlier e-text is wrong, the scan sides with us. Bethphage is “house of my mouth”, not “month”, and Habaziniah, Sanhedrim and Shaaph are spelled the way the page spells them.

Every correction is logged with what the page shows and a cropped image of it. One line, Mizar, little, was lost because the column was cut through the middle of it; the next version of the cutter looks for a blank row in the image before it cuts.

Next up: Nave’s Topical Bible, at two columns and about eight hundred pages.