Community Wishlist Survey 2020/Wikisource/New OCR tool

New OCR tool

  • Problem: 1) Wikisource has to rely on external OCR tools. The most widely used one has been out of service for many months and all that time we are waiting, whether its creator appears and repairs it or not. The other external OCR tools do not work well (they either have extremely slow response, or generate bad quality text). None of these tools can also handle text divided into columns in magazine pages and they often have problems with non-English characters and diacritics, the OCR output needs to be improved.
    2) The tool hOCR is not working for wikisources based on non-Latin scripts. PheTool hOCR is creating a Tesseract OCR text layer for wikisources based on Latin script. E. g. for Indic Wikisource, there is a temporary Google OCR to do this, but integrating non-Latin scripts into our tool would be more useful.
  • Who would benefit: Wikisource contributors handling scanned texts which do not have an original OCR layer or whose original OCR layer is poor, and contributors to wikisources based on non-Latin scripts.
  • Proposed solution: Create an integral OCR tool that the Wikimedia programmers would be able to maintain without relying on help of one specific person. The tool should:
    • be quick
    • generate good quality OCR text
    • be able to handle text written in columns
    • be able to handle non-English characters of Latin script including diacritics
    • be able to handle non-Latin languages

Tesseract, which is an open source application, also has a specific procedure to training OCR which requires corrected text of a page and an image of the page itself. On the Wikisource side, pages that have been marked as proofread show books that have been transcribed and reviewed fully. So, what needs to be done is to strip formatting the text of these finished trascriptions, expand template transclusions and move references to the bottom. Then take the text along with an image of the page in question and run it through the Tesseracts procedure. The improvement would then be updated on ToolLabs. The better the OCR the easier the process is with each book, allowing Wikisource editors to become more productive, completing more pages than they could do previously. This would also motivate users on Wikisource.

Some concerns have appeared that WMF nearly always uses open source software, which excludes e. g. Abby Reader and Adobe, and that the problem with free OCR engines is their lack of language support, so they are never really going to replace Phe's tools fully. I do not know whether free OCR engines suffice for this task or not, but I hope the new tool to be as good or even better than Phe's tools and ideological reasons that would be an obstacle to quality should be put aside.

Discussion

I think this is the #1 biggest platform-related problem we are facing on English Wikisource at this time. —Beleg Tâl (talk) 15:09, 27 October 2019 (UTC)[reply]

Yeah. For some reason neither Google Cloud nor phetools support all of the languages of Tesseract. Tesseract in comparision to the wikisources is missing Anglo-Saxon, Faroese, Armenian, Limburgish, Neapolitan, Piedmontese, Sakha, Venetian and Min nan.--Snaevar (talk) 15:12, 27 October 2019 (UTC)[reply]

Note that you really don't want a tool that scans all pages for all languages as that is so compute-intensive that you'd wait minutes for every page you tried to OCR. Tesseract supports a boatload of languages and scripts, and can be trained for more, but you still need a sensible way to pick which ones are relevant on any given page. --Xover (talk) 07:27, 31 October 2019 (UTC)[reply]
I know. Both the Google Cloud and phetools gadgets pull the language from the language code of the wikisource that the button is pressed on and thus only uses one language. The same thing applies here. These languages are mentioned however so it is clear which wikisources this proposal could support, and witch ones it would not. P.S. I am not american, so I will never try to word things to cover all bases.--Snaevar (talk) 23:01, 2 November 2019 (UTC)[reply]

Even aside from the OCR aspect, being able to extract the formatting out of a PDF int wikitext would be highly valuable for converting pdfs (and other formats via pdf) into wikimarkup. T.Shafee(Evo﹠Evo)talk 11:19, 29 October 2019 (UTC)[reply]

I am not sure about formatting. Some scans or even originals are quite poor and in such cases the result of trying to identify italics or bold letters may be much worse than if the tool extracted just pure text. I would support adding such feature only if it were possible to be turned on and off. --Jan.Kamenicek (talk) 22:05, 30 October 2019 (UTC)[reply]

Many pages requires only simple automatic OCR. But there are pages with another font (italics, fraktur) or pages with mixed languages (e.g. Missal both in local language and latin), where would be usseful to have possibility of some recognizing options. This can be more easily made on local PC, but not everybody have this option. JAn Dudík (talk) 11:21, 31 October 2019 (UTC)[reply]

Would also be great to default the OCR formatting to match the MOS, rather than having to change it all to conform to the MOS manually. --YodinT 14:19, 25 November 2019 (UTC)[reply]

Voting