[BlindMath] Areas still to work on?

Mon Jun 1 07:56:25 UTC 2026

I knew it, AI would have to come in to the discussion. As with a lot of 
AI stuff I think "but is it really appropriate". In this case, if a 
encoding system (eg. UEB) is meant to have deterministic rules then is a 
non-deterministic technique the answer? If the rules are too complicated 
to code manually, I think it says something about the design of the 
encoding system.

Michael Whapples

On 31/05/2026 20:59, Neil Soiffer via BlindMath wrote:
> Two different replies in one...
>
> @Ken: I co-wrote a paper to be presented at ICCHP this summer where my
> co-authors and I examined the state of AI to both produce braille from
> MathMl and also to convert braille to MathML. We looked at both Nemeth and
> UEB. The good news is that the models are improving rapidly -- there was
> ~15% improvement from Gemini 2.5 to 3.1 in a 9 month period. GPT 5.4 was
> worse than Gemini by upto 20%. The open-weight models were significantly
> worse even when given hundreds of thousands of examples. Without lots of
> training examples (2000+) for the commercial models, the results weren't
> good (30% - 70% depending on the model/direction/code). In other words,
> their native knowledge of braille is poor. The best results with Gemini 3.1
> achieved 92% correctness over 200 tests. Some downsides are the slowness to
> get an answer, the flakiness of the answers (it's not deterministic), and
> also, at least for trying to run multiple tests at once, the flakiness of
> the AI to even give a response. And there is the cost also... I'm hoping to
> put the paper up on arXiv soon.
>
> Re: Ray's original question about what work can be done:
>
> In MathML 4, we've added the ability for authors to disambiguate speech.
> For example |x| could be absolute value, determinant, or cardinality. For
> higher level math, there are other interpretations. The only person who
> truly knows what a piece of math should be spoken as is the author. In the
> absence of author input, MathCAT and a few other systems make educated
> guesses (capital letter -> determinant, etc). But they are just guesses.
> Intent can also be used on operators such as "×" (times, cross-product, by,
> ...). With MathML 4, authors have a way of conveying the meaning to a
> screen reader. LaTeX now has a way for the author to indicate the speech
> and pass it along to MathML, but it is a bit awkward at the moment. I
> believe the plan is to fold that into current macro packages and maybe even
> come up with a speech macro package so that "abs" gets added and "det" has
> speech specified.
>
> MathML defines a core concept list
> <https://w3c.github.io/mathml-docs/intent-core-concepts/>. which is a list
> that all screen readers should know how to speak (and translate to other
> languages). MathML 4 is at the point where we are asking for feedback on
> what needs to change, so that list is a good place to start. Is there
> something missing or something wrong? In addition to this list, there is an
> "open" list where anyone can add a (non-conflicting) concept name. The open
> list *might* be adopted by a screen reader and is a source for potentially
> bringing names into the core concept list. We are still working on a
> mechanism so people can easily add their notation, but that's not ready for
> others to try at the moment. We currently have about 1,000 entries in that
> list pulled from Wikipedia and other sources.
>
> @Ray and others: it would be helpful to look at the core concept list and
> file a github issue for any problems you see. Once we have a good way to
> contribute to the open concept list (either corrections or additions), that
> would be another way to contribute. MathCAT supports this feature and the
> next release will support all of the core concepts. MathCAT is used by
> JAWS, NVDA, and Orca screen readers (among others), so improvements you
> make to those lists will show up in the screen readers many people use.
>
> Neil Soiffer
>
>
> On Sun, May 31, 2026 at 11:24 AM Ken Perry via BlindMath <
> blindmath at nfbnet.org> wrote:
>
>> I would agree and I wrote an expression parser once in python for fun.
>> With that said.  We are at a breaking point. The new AI models that are
>> popping up will be able to do much more better. True right now they are non
>> deterministic but the things you can make with them to do what you want now
>> and have them make it deterministic is almost amazing. Two years ago I
>> would not have said this but in less than 2 years we won't have any problem
>> producing 2D and 3D mathematics from formula's in either Nemeth or Ascii
>> math.  I feel like having to be a programmer to be a mathematician is just
>> extra work and that is coming from someone who currently works as a Senior
>> Software Engineer.
>>
>> Ken
>>
>> -----Original Message-----
>> From: BlindMath <blindmath-bounces at nfbnet.org> On Behalf Of Ray
>> McAllister via BlindMath
>> Sent: Sunday, May 31, 2026 2:11 PM
>> To: 'Blind Math list for those interested in mathematics' <
>> blindmath at nfbnet.org>
>> Cc: Ray McAllister <raymcal at att.net>
>> Subject: Re: [BlindMath] Areas still to work on?
>>
>> What about the idea of using existing math coding languages that we have
>> to use anyway with computers, like Sage, Wolfram, etc?  We already have to
>> know those, and after working out a long problem, we often have to test our
>> work to see if it is correct, so we have to code it into something to work
>> in a computer.  If we had the option to enter a long formula in
>> Safe/Wolfram/Pithon, whatever, code, and then the computer, which obviously
>> knows how to read such code, would convert that into 2d graphical display,
>> which we could peruse with a screen reader, that might give some options
>> not previously available, and that would be good even for sighted people.
>>
>> Ray.
>>
>>
>> -----Original Message-----
>> From: BlindMath <blindmath-bounces at nfbnet.org> On Behalf Of Ken Perry via
>> BlindMath
>> Sent: Sunday, May 31, 2026 1:59 PM
>> To: 'Blind Math list for those interested in mathematics' <
>> blindmath at nfbnet.org>
>> Cc: kperry at blinksoft.com
>> Subject: Re: [BlindMath] Areas still to work on?
>>
>> As a sighted person you may not know the speed in which your tools urn
>> matters.  Recorded speech is not the best.  A lot of the symbols already
>> speak fine with screen readers. The problem is more the graphical style of
>> math that many authors use.  We are so much farther along than we were even
>> 2 years ago.
>>
>> -----Original Message-----
>> From: BlindMath <blindmath-bounces at nfbnet.org> On Behalf Of Jonathan Fine
>> via BlindMath
>> Sent: Sunday, May 31, 2026 1:25 PM
>> To: Blind Math list for those interested in mathematics <
>> blindmath at nfbnet.org>
>> Cc: Jonathan Fine <jfine2358 at gmail.com>
>> Subject: Re: [BlindMath] Areas still to work on?
>>
>> Hi
>>
>> This is a very interesting thread. I'm sighted, and reading your comments
>> educates me. I'm trained as a pure mathematician.
>>
>> Here are two comments regarding accessibility and mathematics as a formal
>> and highly symbolic language.
>>
>> The first is that the symbols used depend on the context. I think having a
>> list of relevant symbols (and their meanings) will be a great help.
>>
>> The second is that we could use pre-recorded or pre-generated speech for
>> the most common symbols, instead of generic text to speech.
>>
>> To summarize, I'm suggesting that we consider creating a mode specially
>> designed for authoring and speaking mathematics. In this maths mode there
>> can be a context, which is a short list of common or allowed symbols.
>>
>> I hope this is clear. I welcome comments. Please you'd like some examples.
>>
>> with best wishes
>> Jonathan
>> _______________________________________________
>> BlindMath mailing list
>> BlindMath at nfbnet.org
>> http://nfbnet.org/mailman/listinfo/blindmath_nfbnet.org
>> To unsubscribe, change your list options or get your account info for
>> BlindMath:
>>
>> http://nfbnet.org/mailman/options/blindmath_nfbnet.org/kperry%40blinksoft.com
>> BlindMath Gems can be found at <
>> http://www.blindscience.org/blindmath-gems-home>
>>
>>
>> _______________________________________________
>> BlindMath mailing list
>> BlindMath at nfbnet.org
>> http://nfbnet.org/mailman/listinfo/blindmath_nfbnet.org
>> To unsubscribe, change your list options or get your account info for
>> BlindMath:
>> http://nfbnet.org/mailman/options/blindmath_nfbnet.org/raymcal%40att.net
>> BlindMath Gems can be found at <
>> http://www.blindscience.org/blindmath-gems-home>
>>
>>
>> _______________________________________________
>> BlindMath mailing list
>> BlindMath at nfbnet.org
>> http://nfbnet.org/mailman/listinfo/blindmath_nfbnet.org
>> To unsubscribe, change your list options or get your account info for
>> BlindMath:
>>
>> http://nfbnet.org/mailman/options/blindmath_nfbnet.org/kperry%40blinksoft.com
>> BlindMath Gems can be found at <
>> http://www.blindscience.org/blindmath-gems-home>
>>
>>
>> _______________________________________________
>> BlindMath mailing list
>> BlindMath at nfbnet.org
>> http://nfbnet.org/mailman/listinfo/blindmath_nfbnet.org
>> To unsubscribe, change your list options or get your account info for
>> BlindMath:
>>
>> http://nfbnet.org/mailman/options/blindmath_nfbnet.org/soiffer%40alum.mit.edu
>> BlindMath Gems can be found at <
>> http://www.blindscience.org/blindmath-gems-home>
>>
> _______________________________________________
> BlindMath mailing list
> BlindMath at nfbnet.org
> http://nfbnet.org/mailman/listinfo/blindmath_nfbnet.org
> To unsubscribe, change your list options or get your account info for BlindMath:
> http://nfbnet.org/mailman/options/blindmath_nfbnet.org/mwhapples%40aim.com
> BlindMath Gems can be found at <http://www.blindscience.org/blindmath-gems-home>