Some web pages are incorrectly divided into paragraphs

[mltony]

On some webpages NVDA divides content into paragraphs incorrectly. Two major examples of that are:

1. Certain emails, presumably composed in GMail basic HTML or any other text-based email client, such as Alpine. Here is an example - search for the phrase "Wishing you a merry Christmas" to find the beginning of email body that is split into paragraphs incorrectly. Specifically, every line ends up being a separate paragraph.
2. Virtually all Latex PDF files such as any scientific paper are also affected. Here is an example - search for "Foundation models" to find the beginning of the abstract, that is also divided into lines.
   This also poses as a great problem for sentence navigation (Feature request: Include Sentence Nav as part of NVDA #8518) as reconstructing sentences spanning across multiple paragraphs is pretty complicated from algorithmic point of view.
   So I would like to start a discussion to see if it's feasible to fix paragraph segmentation without causing too many side effects.

Currently SentenceNav add-on only tries to expand sentences across paragraphs if those paragraphs are compatible, where compatible is defined is:

1. They have the same horizontal offset - that is their textInfo.NVDAObjectAtStart.location[0] is the same.
2. They have the same style. Style defined as FonntStyle + fontSize + bold/italic, etc. We only consider style of the first character of the paragraph for simplicity.
   For SentenceNav these two conditions are enough. There will be false positives in a properly paragraphed web page, but then SentenceNav will still detect that almost any paragraph still ends with a period, so it will still understand not to merge sentences incorrectly only because paragraphs are compatible.

However, if we consider the task of identifying paragraphs on broken pages (see two examples above) - these two conditions won't be enough. Primarily because the same kind of false positives that are benign in sentence division case would be considered a major regression for paragraph navigation.
Take a look at a sample blog post, that is properly paragraphed. If we only apply two heuristics above to merge paragraphs, then will end up with the entire article being a single gigantic paragraph. That would be a major regression.
So we need to come up with another heuristic to add to two sentenceNav heuristics mentioned above. I can see a couple of options:

1. Try to measure vertical offsets between lines. In most PDFs paragraphs are visually separated by a greater gap between lines. I played with this a bit with a following snippet:

   t=focus.treeInterceptor.makeTextInfo('caret')
   def measure(offset):
       rect = t._getBoundingRectFromOffset(t._startOffset)
       y0 = rect[1] + rect[3]
       tt = t.copy()
       tt.move('line', offset)
       rect = tt._getBoundingRectFromOffset(tt._startOffset)
       y1 = rect[1] + rect[3]
       return y1-y0
   

   And this approach doesn't appear to be so good, because the gap between two consecutive lineds within a paragraph tends to change every time - I'm getting 22, 24, 28 pixels. The gap between paragraphs is more like 36 pixels. I would hesitate to hardcode these numbers.
2. Try to check if the entire paragraph appears visually as a single line. That might actually work for PDFs, but won't work for emails, since browsers will do word wrap automatically when the browser window width is too small.
3. Check for double newline character. Sometimes this can be a good hint. But in PDF example above in the abstract they appear to be useing double newline character to separate lines, not paragraphs.
4. Check that the number of characters (or words) in a paragraph is below certain limit. This would again add a hardcoded constant, which is not ideal. So far this appears to be the best option we have.

Therefore I don't see a good heuristic that would work correctly in all cases for reconstructing paragraphs, that would work for both use cases I mentioned above and not cause any major regressions on other websites.
But I would like to check if I'm missing anything obvious and perhaps there is an easy way to reconstruct paragraphs? @CyrilleB79, do you have any suggestions/thoughts?

[mltony]

Closing this as there appears to be no solution for this problem.