[lojban] Re: my opinion on why lojban isn't specifically well suited for human-computer interaction.1

["M@" <lojban-out@lojban.org>]

> > The
> > difference between MI and AI is as fundamental as the difference between
> > 1 and 0.  Thus, the two potential uses of a syntactically unambiguous
> > language (SUL for brevity) in a computer have no grey area between them.
> 
> Guessing what you mean by MI and AI, I think it is the grey area that
> you don't see that is precisely the area of interest.
> 
So you're saying that within the context of AI lojban could be a mechanism
to get incrementally better as an approximation?  That's possible depending
on what the goal is.  IMHO the goal should be MI, which would be an exact
solution because, after training, it would be able to speak and understand
language(s) fluently.  You could converse with a computer instead of using
it.  The problem is that it's impossible with classical computer hardware
(though building hardware that makes it possible is possible, if not at the
moment).

I guess what I'm saying is that even if an AI could pass a Turing test 99%
of the time it would still run into the fundamental problem that it has no
idea what it's talking about, and until it does it will never achieve the
last 1%.  

> > The short term use (dream) is to teach computers to utilize the
> > language?s unambiguous nature to provide an interface that can be read
> > and written to in a spoken language.
> 
> That could be done with FORTRAN, but the issue is "what is a 'language'"
> 
That question haunts me, I've never heard a good definition.  I harass every
linguist I come across about it. .a'o Do you have one?

> > The problem I have with this goal is that it?s a solution to a problem
> > that has been solved for decades.  There are many SULs that are already
> > written and in use as computer interfaces across the world.  My personal
> > favorite is bash.  To compare bash to lojban I would say that bash is
> > easier to learn (it is much smaller after all) and every bit as
> > versatile as a cross-programming-language interaction framework as
> > lojban could possibly be in a classical computing environment.
> 
> That's fine, if the only thing you want to do with a computer is write
> programs.  But computer languages are limited to ONLY writing computer
> programs.  One has to at least get to the versatility of SQL in order to
> be useful to users, and even there the ideal is to have a somewhat more
> natural-language interface.
> 
Bash isn't a programming language (though it can be used as a scripting
language, much like lojban could if you added a few keywords), it's a UI.
If I wanted to play a game I would type ut2004 and press enter and a game
would start (just as with the hypothetical lojban interface I might say mi
kelci ko).  The problem I'm pointing out is that the watered down version of
lojban that would be required to get even a very good AI to make something
useful out of it would seem just as unintuitive as bash or sql (if perhaps
more in depth by virtue of a spectacular AI), as you said, you would never
be fluent with it, even if you were fluent in spoken lojban.


> > Another way to say it, is no matter how hard you try classical computers
> > will never be able to figure anything, they rely on unambiguous meanings
> > to get things done, lojban doesn?t necessarily provide that, bash does.
> 
> Which is where AI comes in, because AI doesn't require unambiguous
> meanings.  It is significantly aided by unambiguous meanings when they
> are intended, known and can be conveyed, but an AI is ideally capable of
> asking a quasi-intelligent question to resolve an ambiguous meaning, to
> seek additional specificity for a vague meaning, or perhaps do a
> probabilistic or other analysis of user intent when these aren't possible.
> 
Even if the AI was smart enough to get answers it needs from the user it
still won't have the smarts to do anything it wasn't specifically coded to
do (AIs are really finite state machines (where the states are defined by
code), albeit complicated ones augmented with specialized tools to get
things done without state).  If I asked an AI about the temperature of the
sun it might be smart enough to go look that up on google and return the
results to me, but if I then ask it to calculate the temperature of the sun
pi*10^6 years ago given that the bread's in the oven at 451° there's no way
it will be able to answer because while it could do the math for how many
years that is, and it can do the math for the temperature of the sun that
many years ago it wouldn't associate those two numbers unless any more that
it would associate the bread as being 3.1 million years old, unless it was
specifically coded for the question type or it actually understood the
meaning.

> > The real beauty of lojban isn?t the fact that it?s a SUL.
> 
> By being SUL, it cuts out one chunk of the processing load in
> interpreting user input.  There is still speech processing, and semantic
> ambiguity, and syntactic/semantic errors as sources of processing load,
> but a complete solution to one problem should in theory make the total
> task of natural language processing easier.

That is, if you buy two key ideas.  First, that natural language processing
is a good enough substitute for natural language speaking that people will
have a use for it (impossible to know until we get there, but my belief is
no).  Second, that CPU time is the limiting reagent in the system (not the
case for hierarchy based MIs, or even AIs running on hierarchical hardware).
> 
>  >As with any real beauty there is always a converse beast, which
> > in this case is the fact that it?s silly to be more creative than the
> > person (or computer) you?re talking to.
> 
> They call this "art".
> 
Touche.

> > Of course, the other dream is to actually teach a MI how to speak and
> > understand lojban.  Assuming such an auto-associative system could be
> > made hardware wise, there would be absolutely nothing preventing it from
> > being a perfectly natural feeling interface in spoken or typed lojban.
> > You could ask it absurd questions and it could give you obviously
> > considered answers.  The problem I find with this is that once you?ve
> > built MI there?s no reason whatsoever you couldn?t teach it English as
> well.
> 
> English is a much more difficult problem, because it is not an SUL, and
> has evolved as a non-SUL so that it has extremely odd semantic
> relations.  If you ran your computer the way that run your dog or the
> way you run a lap on the track (as opposed to sitting on the lap of your
> girlfriend - I wouldn't want you to run her lap, or to have the runs on
> her lap - she wouldn't exactly lap it up).
> 
> And those are the obvious kinds of problems.  Much of the difficulty
> with English is that the problems are NOT obvious, and every structural
> analysis of English runs into some problems that violate any
> generalizations that have been otherwise made.  I was just reading a
> short discussion of the possible determiners in a noun phrase in
> English, and the restrictions are as clear as mud.
> 
That's the beauty of auto-associative systems, they're analog (not
technically if they're on digital hardware, but they approximate analog-ness
extremely quickly as you put more hardware into them (subject to the law of
diminishing returns of course)), so the absurdly specific cases of any
natural language won't pose any more of a difficulty than training it as you
would a child.  Probably less because MI computers don't have to have
emotions to combat.

The solution to linguistic computers is to have a great number fewer rules
instead of more rules (in other words AI is headed in exactly the wrong
direction), the only real problem is that we don't quite know what those
rules should be yet.  If we don't know most of the few rules in 20 years
I'll be shocked.

--M@



To unsubscribe from this list, send mail to lojban-list-request@lojban.org
with the subject unsubscribe, or go to http://www.lojban.org/lsg2/, or if
you're really stuck, send mail to secretary@lojban.org for help.