> I had expected to see more in this direction over the years. A famous
> (in its time and place) precursor with this approach was ROSIE, a
> very nice expert systems language done at RAND in the late 70s and
> early 80s. Have you seen it?

There was some by now classical research done on natural language interfaces
in the early 80s. Put simply, the general conclusion was that it is a bad
idea to make a computer mimic humans, because users will then attribute
human capabilities to it that it doesn't have. (Think ELIZA.) This is a
basic social principle that we need in order to interact with others (we
can't read their minds). When we then attribute too much to the computer,
the breakdown that follows is very harsh, much more costly than what is
gained by familiarity when it works.

It is much better to present the computer so that we will make reasonable
attributions. Math-like languages make us regard it as a math machine, which
is not that unreasonable.

But of course, this didn't affect AI researchers much. The real reason we
haven't heard about progress in AI is that there hasn't been any to speak
of. The small successes in restricted, well-chosen domains have never
generalized, be it natural language processing, computer vision, symbolic
induction, or whatever. The term "AI winter" was coined in the Lisp
community in the late eighties. Today it seems that Bill Gates is about the
only one who still thinks that AI will be "the next big breakthrough". For
the last 40 years, AI news stories have begun like this: "Soon they will be
here--computers that xxxx". "Don't hold your breath" is one of my favorite
American expressions.

> The trick in most of these systems is not how hard it is for the
> system to recognize a restricted English syntax, but how hard is it
> for a random human to learn to write in the restricted syntax.

Would another way to put this be that the supposed advantage, being easier
to learn than more standard programming syntaxes, is not there? "Restricted"
here means that it looks like it's ordinary language but it isn't. It
doesn't let people use ordinary language.

In fact, a main feature of natural language is not the syntax--but the fact
that most of the time we needn't be very precise at all with the syntax to
be understandable. _This_ is what a "natural" syntax suggests to a user. And
this is not what you want a user/programmer to belive, right?

Syntax is a big deal in programming languages but not real ones; this is a
point that CS researches miss all the time. In fact, eg. Chomsky's theories
apply much better to programming languages than real ones. With all the
research on syntax, Markov chains (statistics about what words occur close
to each other) are still better predictors of word order in "real" natural
language than any theory of syntax.

The hard problem in programming is figuring out what to do--you have to
understand the domain of available means (the language capabilities) as well
as well as the problem you are solving. Addressing syntax leaves this alone,
it only concerns how you express the solution once you have figured it out.
Programming languages are just meant to be means for expressing solutions.
What you need to address is the process of reaching solutions. The
interactive environment of Smalltalk is the most important advance that has
been made so far: a compiler merely processes the solution, whereas an
interactive environment supports you while working out a solution.

So to condense my point: we need to shift the focus from the form of the
expressed solution, to the process which produces it (and tools to support
this). But the nature of such cognitive tools has only begun to be addressed
in the last 5-10 years.


Oops, there I did that rant again.

Henrik