Minutes: MathML Full Meeting, 13 April, 2023

 Attendees:

   - Neil Soiffer
   - Louis Maher
   - David Farmer
   - Steve Noble
   - Bruce Miller
   - Bert Bos
   - David Carlisle
   - Deyan Ginev
   - Paul Libbrecht
   - Cary Supalo
   - Andy Smith (Vispero)
   - Andres Gonzalez (Apple)
   - Tommy McHugh (Apple)
   - Glen Gordon (Vispero)

<https://sandbox.cryptpad.info/code/inner.html?ver=5.2.2-0#cp-md-0-regrets>
Regrets
<https://sandbox.cryptpad.info/code/inner.html?ver=5.2.2-0#cp-md-0-agenda>
Agenda
<https://sandbox.cryptpad.info/code/inner.html?ver=5.2.2-0#cp-md-0-1-announcements-updates-progress-reports>1.
Announcements/Updates/Progress reports

NS: TPAC 11-15 Sept in Seville, Spain. Meetings Mon, Tues, Thurs, Fri

Anything that can be closed?

We had guests from Apple and Vispero to give us feedback on our ideas.
Everyone on the call introduced themselves.

   1. Discussion about AT and "intent"

The AT reps were given. this explainer about "intent".
<https://w3c.github.io/mathml-docs/intent-explainer/> They were also
pointed at. this issue with questions we hope to get answers to from them.
<https://github.com/w3c/mathml/issues/461>

NS: Using MathCAT, I have tried to show what does and does not work for
intent. MathCAT is just one implementation. Other people have other ideas.

NS: Does anybody need a recap of what the idea behind intent is?

No.

NS Our first question is, what do you think? Would you be interested in
implementing intent?

GG: Can you turn intent off in MathCAT? Can you restructure intent's
implementation in MathCAT?

NS: You can tell MathCAT what to do when MathCAT encounters an error.
MathCAT can crash and burn, or MathCAT can try to ignore the malfunctioning
intent.

NS: If you need to be able to turn off intent, I can do that.

NS: We have not yet come up with a list of known intents.

GG: Turning off intent was more of a request than a question.

GG: The person writing the intent may have different ideas about speaking
math than the listener of the math would have in hearing it.

GG: How would intent configurability work?

NS: It is easy to turn off intent processing in MathCAT completely, but
turning it off on a case-by-case basis would be complicated to specify.

GG: The more intent processing can be configured, the more GG would like to
work with it.

DC: MathCAT does have different verbosity levels.

NS: MathCAT can use Dr. Nemeth's Math Speak.

NS: Often Dr Nemeth wanted to hear exactly what was there so that he could
copy it directly onto his Braillewriter. I haven't implemented navigation
by intent arguments, but I think character level navigation will ignore
intent once I implement argument navigation of intent.

GG: Different readers want to hear speech in different styles.

PL: Sometimes tunability by the reader is desirable. But there's many, many
situations where it will ruin political scenarios. It will ruin the
literary style that one would typically expect so this is a single setting.
I see no problem with this.

PL: It's okay to warn authors that Intents could be ignored.

PL: I think that disabling intent is a good way to destroy quality.

PL: If this can be done with a single setting that users will know, then
that is OK.

TM: If you have a long math equation, where do you put the intent?

NS: Intent should be used mainly to resolve ambiguities like explaining
that (A, B) stands for a coordinate or an open or closed interval.

NS: Vispero is willing to implement intent.

NS: Is Apple willing to implement intent? I do realize that you cannot make
any definite statements about what Apple will do.

AG: I have another question before we go into that.

AG: You said that intent will not have much impact on Braille output. Can
you go into that?

NS: Braille reflects the same symbols that print does. Braille does not try
to interpret math; it just tries to copy the math. Intent is used to change
the output to make it clearer; therefore it would have no impact on the
actual math shown by Braille.

NS: There are however a few cases where braille requires inferring meaning.
Someone this week found what they thought was an error in MathCAT. They
used a WYWYG tool to create what could be interpreted as a binomial or as a
2x1 column vector. They look very similar. In TeX, there are some subtle
spacing differences. Although I implemented a binomial intent for speech, I
forgot to make the braille look at that.

NS: Intent could also help recognizing a Roman numeral. Nemeth code at
least has a few rules around Roman numerals.

NS: Intent could also help a reader distinguish between an absolute value
or a determinant. This would not matter to how Braille is embossed.

Ag: Apple is interested in intent. This is why we are here. Intent needs to
be a flexible mechanism that allows customization to the presentation to
the AT end user, depending on the user’s needs and preferences.

NS: DG is a big proponent of the need for authors to control the speech
because he has to deal with a great deal of higher-level math.

DG: Higher level math often publishes non-standard math written with
non-standard symbols. Often this math and these symbols cannot be expressed
in high school math terms. Regular AT will not catch up with this math for
many years.

DG: is worried about consistency. Various ATs can read the same text and
give different speech outputs.

DG: If you know of any threats we should know about, please let us know.

GG: Warned against consistency. Vispero wants to have the flexibility to do
things differently. Not all screen readers agree on how things should be
read. If you are a for-profit company then you might disagree with another
third party company. Mandates would make the AT developer uneasy.

DG: We can recommend potential default rules. These default rules would
have to have alignment with other companies.

NS: Some default rules are still undecided, like should x superscript 2 be
pronounced as x to the 2 power, or x squared.

GG: It would be good to have some default rules because we're not going to
get the intent right.

PL: It would be good for internationalization to have standard terms. We
would like to standardize not only in English, but to have standardized
terms in other languages as well.

SN: Pearson is an international company in various languages. It does
assessments in many languages. Pearson uses specific terms in each
language. Because Pearson does not want to rely on ATs, Pearson overrides
each MathML math expression with an alt tag. This is terribly slow and
expensive.

SN: Pearson would like test producers to get standard renderings from
MathML so that Pearson would not have to override Ats for each equation.

NS: When he worked with the Educational Testing Service (ETS), ETS worried
about having the AT give the test answer away. If you asked a question
about scientific notation, and the AT read a number in scientific notation,
then you just gave the answer away.

SN: If a state wants something read one way, he would like a way to make AT
do this.

PL: Intents will help in this effort.

CS: Speech Synthesis Markup Language (SSML) should fix this. NS: said it
would not.

BM: said that intent could not automatically read math equations without
giving away the answer. Each equation's intent component would have to be
written to obscure the equation to hide the test answer.

PL: Perhaps we should have standardized pronunciations of equations for
tests. This would be a special use of intent for math.

NS: There is no way AT systems will always speak math identically.
Therefore, intent cannot solve this problem.

NS: Is the way parentheses are pronounced a test standard? For example,
saying open parenthesis or left parenthesis?

SN: Yes.

NS: Is there flexibility in how Voiceover can speak math?

TM No. You can change how some things are pronounced. You cannot change how
math is spoken.

NS to SN: We cannot alter the speech to help test takers.

GG: Maybe there should be a testing mode.

ns: We could make a test override.

GG: If you are doing a standardized test, you need to give people the same
output. You also need to give them the same output that they would
encounter during their normal work. The tests should not output unusual
ways of pronouncing math during tests.

PL: said that when certain things are translated, the gender of the words
might change to feminine. He thought it would be at least ten years for AT
to gain this ability.

NS: How deep into math translation does Vispero get.

GG: We are as good as our localizers.

PL: There are an infinite number of language variations.

PL: You may need context to determine if things should be singular or
plural, masculine or feminine.

NS: MathPlayer can handle plurals and some gender changes, but not verb
changes. The translators could use phrases so that it was not a word for
word translation. MathPlayer used some context.

NS: said MUS said that Microsoft used word-by-word translation. They did
not use context for their translations.

NS: to Apple: Do you use context in your translations?

AG: We tried to provide context to the translators. Math is a special
problem.

AG: I'm not saying that we are doing this at the moment, or we would need
to give context for different types of mathematical constructs.

GG: What we do now does not matter because we are about to abandon it.

GG: Looking at what MathCAT does, and optimizing it, is going to be more
effective than looking at what we do now.

NS: MathCAT looks at the number in front of a word, and decides if singular
or plural words should be used.

BM: The intent that we formulated implies a functional way to write things.
but there is also a template way to write intents, which is like a sequence
of words to be pronounced sequentially. Which method is better?

AG: goes for flexibility. He prefers the functional approach.

NS: prefers the functional approach.

NS: thanked the speakers. We learned a lot. Feel free to come to other
meetings.