FR version is available. Content is displayed in original English for accuracy.
Advertisement
Advertisement
⚡ Community Insights
Discussion Sentiment
80% Positive
Analyzed from 966 words in the discussion.
Trending Topics
#arabic#text#rendering#article#read#left#right#algorithm#more#significant
Discussion (17 Comments)Read Original on HackerNews
Think of variable width characters and kerning and ligatures and hyphenation and justification. Imagine computers had been won by a CJK language, which have none of these problems. You could imagine a similar article about how exotic and difficult English layout is.
> The relevant rule, W2 of UAX #9, reclassifies a digit as an ARABIC NUMBER if any of the previous strong characters in the paragraph were Arabic letters, and as a EUROPEAN NUMBER otherwise. Both render their internal digits left-to-right, which is correct: numbers everywhere on Earth are read most-significant-first.
Does the author mean most-significant-on-the-left? The statement as written is a statement about the order in which one reads or perhaps thinks the number, whereas I think the author is discussing how numbers, including collections of numbers delimited by hyphens and such, should be laid out on the page.
on the other hand, in formal arabic, it's not unusual that numers are read in clusters from least significant to most significant (right to left). 1984 would be read : eighty four and nine hundred and a thousand. not sure if the author is aware of this
What does that even mean in this context? In a strictly LTR language, sure, you read left-to-right and the glyphs are rendered left-to-right. But the whole discussion is about bidirectional text, where the text is rendered by a complex algorithm. What is the “rendering direction”?
I know just enough about some RTL languages to know that one can absolutely intersperse RTL text with, say, and English phrase, and you still read the first (leftmost in the group) English sound first and so on :)
This part nearly had me chuckle audibly:
He says yes. The result is "Simplified Arabic": initial fused into medial, final into isolated, ligatures dropped. It conquers the Arab newsroom in a generation. Mrowa is assassinated at his desk eight years later, by an unrelated faction, in an unrelated dispute.
Also, it's depressing how hundreds of millions of people couldn't even get their language typeset on a computer, and our industry meanwhile was busy building AI-native AI for your groceries (have we mentioned it has AI btw?) and similar performative bullshit.
(i will permit myself to not explain in excruciating detail why i feel that way about this, as we have this discussion several times a day on this site)
> the Kashida section was contributed to this post from a talk in Arabic of Nawal Hadeed, which she translated and added to the post herself. Although I'm unsure of LLM usage in the translation process, looking at the original Arabic I felt some change in tone while editing the post. I could have either declined the translation and never have this documented, procrastinate in translating it myself (which has been ongoing for a while), or publish as it is. I found the last least damaging.
If you want a solution for this it has to happen in the rendering step, not the shaping (which is HarfBuzz's main task). The shaper has no information about the available space, but when rendering you could stretch individual glyphs to the desired width, similar to adjusting the width of whitespace in Latin, but more complex, because you actually have to modify the glyphs with a scale transform. I am not an expert on Arabic script by any means, but this should be possible IMO. It would at least be an interesting experiment. Of course the JSTF table would be the right way to do it, but there seems to be a lot of confusion around it. Maybe in the age of LLMs we can give it another shot.
As a practical matter, there’s an input length n and there is some upper bound B on a credible line length as measured in code points, so there are only at most n*B credible proposed lines to evaluate, which also limits the useful look back on the table to B positions, so I think the time complexity could be reduced to O(n*B^2) without making the results worse on reasonable inputs, and this is probably quite tolerable.
[0] Straightforward once you’ve implemented the whole Arabic rendering stack, anyway. I am certainly not qualified to calculate this function :)
The hard part is that typography, shaping, bidi behavior, font fallback, search, and the editor model all leak into each other.
You cannot fix one layer cleanly when the assumptions are wrong in all of them.