Page MenuHomePhabricator

Request to enable "Google Cloud Vision OCR" to recognize additional diacritic letters on Wikisource
Closed, DeclinedPublicBUG REPORT

Description

Steps to replicate the issue (include links if applicable):

Dear Sir, here is the Index page.
of the Northern Min New Testament (language code is mnp)

When creating a page there, for example, this page

What happens?:
Several letters are not recognized in the "Google Cloud Vision OCR" ( nor on the other two OCRs, namely "Tesseract OCR", and "Transkribus OCR".

What should have happened instead?:
Could someone please help add the following fifty-eight letters to Google OCR tool:

Vowels with Double Macrons above: a̿ i̿ u̿ e̿ o̿ A̿ I̿ U̿ E̿ O̿

Vowels with Macron above and diaresis below: : ā̤ ē̤ ō̤ ṳ̄ Ā̤ Ē̤ Ō̤ Ṳ̄

Vowels with Acute accent above and diaresis below: á̤ é̤ ó̤ ṳ́ Á̤ É̤ Ó̤ Ṳ́

Vowels with Breve accent above and diaresis below: ă̤ ĕ̤ ṳ̆ ŏ̤ Ă̤ Ĕ̤ Ṳ̆ Ŏ̤

Vowels with Grave accent above and diaresis below: à̤ è̤ ṳ̀ ò̤ À̤ È̤ Ṳ̀ Ò̤

Vowels with Circumflex accent above and diaresis below: â̤ ê̤ ô̤ ṳ̂ Â̤ Ê̤ Ô̤ Ṳ̂

Vowels with Double Macron accent above and diaresis below: a̤̿ e̤̿ o̤̿ ṳ̿ A̤̿ E̤̿ O̤̿ Ṳ̿

Other information (browser name/version, screenshots, etc.):
Screenshot 1 showing Google OCR selecteed prior to transcribing:

Screenshot 1 Googlw OCR selected.PNG (1,328×641 px, 302 KB)

Screenshot 2 showing the transcribed text with errors:

Screenshot 2 showing transcribed text with errors.PNG (1,325×624 px, 370 KB)

Event Timeline

Could someone please help add the following fifty-eight letters to Google OCR tool

Do you have any information on how this can be done? As far as I'm aware there's not much we can do to influence how Google's OCR works.

PR: https://github.com/wikimedia/wikimedia-ocr/pull/157

It seems that we have previously switched from TEXT_DETECTION to DOCUMENT_TEXT_DETECTION, but for some reason went back again: T246944#5978037

Hi Samwilson,

Apparently there are actually 73 diacritic letters that are not currently recognizable by Google OCR. The full list highlighted in red can be found On this page.
Also, Northern Minuses both carons and breves which I have also shown on that page.

Carons used in Northern Min: ǎ ǐ ǔ ě ǒ Ǎ Ǐ Ǔ Ě Ǒ
Breves used in Northern Min: ă ĭ ŭ ĕ ŏ ă̤ ĕ̤ ṳ̆ ŏ̤ Ă Ĭ Ŭ Ĕ Ŏ Ă̤ Ĕ̤ Ṳ̆ Ŏ̤

The carons and breves in Northern Min are completely different pronunciations so they are not interchangeable for words.
For example, the following Northern Min words use carons (and it cannot be substituted with a breve):

  1. Iǎ-Sú (Jesus) (the "ǎ" in this word is always a caron)
  2. Gṳ̌ing (Son) (the "ṳ̌" in this word is always a caron)
  3. Mǎ-duōi (Matthew) (the "ǎ" in this word is always a caron)
  4. ě̤ng-iāu (glory) (the "ě̤" in this word is always a caron)
  5. sǎ̤-chṳ̄ (angel) (the "ǎ̤" in this word is always a caron)
  6. Bǐ-dǎ̤ (Peter) (the "ǐ" and "ǎ̤" in this word are both carons)
  7. biǎng (bread) (the "ǎ̤" in this word is always a caron)

The following Northern Min words use breves (and it cannot be substituted with a caron):

  1. ŭ-di̿ (there/that place) (the "ŭ" in this word is always a breve)
  2. dă̤-dí (know) (the "ă̤" in this word is always a breve)
  3. iŏ̤ng (this) (the "ŏ̤" in this word is always a breve)
  4. cì-ciă (an) (the "ă" in this word is always a breve)
  5. Bă-duōi-nǐ (Bethany) (the "ǎ" is a breve; the "ǐ" is a caron)
  6. hă-sáing (disciples) (the "ă" is a always breve )
  7. ciĕ-duòi (receive) (the "ĕ" is a always breve )

If you looked at one of the sample images, it should be relatively simple to detect whether the letter is a breve or a caron (Google OCR currently detects neither of them)

If you looked at one of the sample images, it should be relatively simple to detect whether the letter is a breve or a caron (Google OCR currently detects neither of them)

If Google OCR doesn't detect them, I'm not sure what we can do. What are you suggesting to change here?

Hi Samwilson, Google OCR is not currently programmed to detect them. That does not mean it would be unable to detect them in the future.

If Google OCR can detect 58 of the 73 letters then that would still be a very massive improvement from our perspective.

That's all true, but I don't think there's much we can we do about it, unfortunately. We just send images to the Google API and get text back. We don't have any say in their algorithm.

You could try asking on the Google Cloud Vision support forums: https://docs.cloud.google.com/vision/docs/getting-support

I don't think there's anything we can do here. (If we weren't using a proprietary service, things might be different.)