One thing I sometimes think about when I think about text layout problems is how the text we use also has a bunch of complexities that we can take for granted.
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.
Both Latin and Chinese have been modified by the technology used to write them.
When carved in stone the lines are much straighter. When written with brush or pen they became semi-cursive. When printing was introduced, they became grid-like and regular.
What westerners who are passingly familiar would think of as the standard Chinese typeface - the strict square grid with straight-line characters - arises in part from printing technology. Easy to carve that into wood blocks, and easy to line up the slots into a grid.
Latin was similarly morphed to fit into the realities of printing in the 1500s. And is still being morphed. Notice how numbers 123... are in-line and at the same height as the letters. That's a very modern convention, typewriter and computer influence on our orthography. Traditionally digits were more likely to appear as subscript, off-centre.
> 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.
I think he's talking about the rendering algorithm with regards to the stream of text. essentially saying rendering direction should follow reading convention.
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
> rendering direction should follow reading convention.
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 :)
Disclaimer: I’m not fluent in Arabic by any means, but the stretched out to both margins style looks very Quranic to me. I don’t think it looks appropriate for say, a message about my DoorDasher.
the entire article has llm tells all over it. i read it anyway, and i'm grateful for all the facts i learned (although i cannot trust all of them, for reasons aforementioned), but i genuinely (!) think it's a shame because the topic is an absolutely fascinating one
(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)
That might be because of translations from Arabic. The article was also posted on a different website where the author responded
> 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.
This article is wonderful. It's interesting, it's captivating, full with detail, and to think I never gave much thought about Arabic rendering before.
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.
Very interesting. I just implemented a text shaper and renderer from scratch with support for complex scripts like Arabic, Nastaliq and Indic (will soon post about it here on HN). Now that you write about it, the lack of stretching really is a deficiency in the OpenType spec.
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.
It does seem like a KP-like algorithm ought to be able to optimize the break positions without extreme algorithmic difficulty aside from the inputs being considerably more complex than for Latin block print: the cost function for a proposed line is a straightforward [0] calculable function of the contents of the line, and I think one could make a dynamic programming algorithm that tracks, for each input position, the cost of the optimal layout of all text up to that position with a break at that position. This gives an algorithm that takes cubic time. (For input length n, you need to fill in n values in the table. Each value scans the entire table before that position and does a calculation with complexity linear in the proposed line length.)
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 :)
> but when rendering you could stretch individual glyphs to the desired width
“Individual glyphs” :)
It’s Arabic, so you wouldn’t stretch a single glyph, id would have to e done after shaping so you can work out the next run (either a single Aleph or the joined characters) in order to know what is stretchable (then throw it to your layout step)
Internet Explorer 5.5 implements text-justify: kashida. For one brief, weird browser-quarter Microsoft is the only software vendor on earth that can justify Arabic correctly on a screen.
One thing I sometimes think about when I think about text layout problems is how the text we use also has a bunch of complexities that we can take for granted.
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.
Both Latin and Chinese have been modified by the technology used to write them.
When carved in stone the lines are much straighter. When written with brush or pen they became semi-cursive. When printing was introduced, they became grid-like and regular.
What westerners who are passingly familiar would think of as the standard Chinese typeface - the strict square grid with straight-line characters - arises in part from printing technology. Easy to carve that into wood blocks, and easy to line up the slots into a grid.
Latin was similarly morphed to fit into the realities of printing in the 1500s. And is still being morphed. Notice how numbers 123... are in-line and at the same height as the letters. That's a very modern convention, typewriter and computer influence on our orthography. Traditionally digits were more likely to appear as subscript, off-centre.
A more academic treatment of justifying Arabic-script text can be found in https://quod.lib.umich.edu/j/jep/3336451.0023.104?view=text;...
awesome article! i appreciated the alternatives and misc. workarounds to OT/unicode typesetting toward the end, very helpful :D (go blue!)
I'd like to see some more mainstream usage of disconnected fonts for Arabic, for example like these: https://www.arabaddigital.com/en/article/2100-Quarantining-o...
Is this a small typo?
> 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.
I think he's talking about the rendering algorithm with regards to the stream of text. essentially saying rendering direction should follow reading convention.
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
> rendering direction should follow reading convention.
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 :)
actually : four and eighty and nine hundred and a thousand
Disclaimer: I’m not fluent in Arabic by any means, but the stretched out to both margins style looks very Quranic to me. I don’t think it looks appropriate for say, a message about my DoorDasher.
But a message about your Door dasher also likely wouldn't be justified.
the entire article has llm tells all over it. i read it anyway, and i'm grateful for all the facts i learned (although i cannot trust all of them, for reasons aforementioned), but i genuinely (!) think it's a shame because the topic is an absolutely fascinating one
(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)
That might be because of translations from Arabic. The article was also posted on a different website where the author responded
> 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.
Same, and it is a shame. Reading it is wearisome.
This article is wonderful. It's interesting, it's captivating, full with detail, and to think I never gave much thought about Arabic rendering before.
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.
Very interesting. I just implemented a text shaper and renderer from scratch with support for complex scripts like Arabic, Nastaliq and Indic (will soon post about it here on HN). Now that you write about it, the lack of stretching really is a deficiency in the OpenType spec.
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.
It does seem like a KP-like algorithm ought to be able to optimize the break positions without extreme algorithmic difficulty aside from the inputs being considerably more complex than for Latin block print: the cost function for a proposed line is a straightforward [0] calculable function of the contents of the line, and I think one could make a dynamic programming algorithm that tracks, for each input position, the cost of the optimal layout of all text up to that position with a break at that position. This gives an algorithm that takes cubic time. (For input length n, you need to fill in n values in the table. Each value scans the entire table before that position and does a calculation with complexity linear in the proposed line length.)
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 :)
> but when rendering you could stretch individual glyphs to the desired width
“Individual glyphs” :)
It’s Arabic, so you wouldn’t stretch a single glyph, id would have to e done after shaping so you can work out the next run (either a single Aleph or the joined characters) in order to know what is stretchable (then throw it to your layout step)
very interesting, arabic is a good reminder that text rendering is mostly solved for the scripts that shaped the defaults.
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.