PEREIRA@SRI-AI.ARPA@sri-unix.UUCP (08/20/83)
I will also refrain from flaming, but not from taking to task
excessive claims.
I'll refrain from flaming about traditional (including
logic) grammars. I'm tired of people insisting on a
restricted view of language that claims that grammar rules
are the ultimate description of syntax (semantics being
irrelevant) and that idioms are irritating special cases. I
might note that we have basically solved the language
analysis problem (using a version of Berkeley's Phrase
Analysis that handles ambiguity) ...
I would love to test that "solution of the language analysis
problem"... As for the author being "tired of people insisting on a
restricted ...", he is just tired of his own straw people, because
there doesn't seem to be anybody around anymore claiming that
"semantics is irrelevant". Formal grammars (logic or otherwise) are
just a convenient mathematical technique for representing SOME
regularities in language in a modular and testable form. OF COURSE, a
formal grammar seen from the PROCEDURAL point of view can be replaced
by any arbitrary "ball of string" with the same operational semantics.
What this replacement does to modularity, testability and
reproducibility of results is sadly clear in the large amount of
published "research" in natural language analysis which is untestable
and irreproducible. The methodological failure of this approach
becomes obvious if one considers the analogous proposal of replacing
the principles and equations of some modern physical theory (general
relativity, say) by a computer program which computes "solutions" to
the equations for some unspecified subset of their domain, some of
these solutions being approximate or plain wrong for some (again
unspecified) set of cases. Even if such a program were "right" all the
time (in contradiction with all our experience so far), its sheer
opacity would make it useless as scientific explanation.
Furthermore, when mentioning "semantics", one better say which KIND of
semantics one means. For example, grammar rules fit very well with
various kinds of truth-theoretic and model-theoretic semantics, so the
comment above cannot be about that kind of semantics. Again, a theory
of semantics needs to be testable and reproducible, and, I would
claim, it only qualifies if it allows the representation of a
potential infinity of situation patterns in a finite way.
I don't recall a von Neumann bottleneck in AI programs, at
least not of the kind Backus was talking about. The main
bottleneck seems to be of a conceptual rather than a
hardware nature. After all, production systems are not
inherently bottlenecked, but nobody really knows how to make
them run concurrently, or exactly what to do with the
results (I have some ideas though).
The reason why nobody knows how to make production systems run
concurrently is simply because they use a global state and side
effects. This IS precisely the von Neumann bottleneck, as made clear
in Backus' article, and is a conceptual limitation with hardware
consequences rather than a purely hardware limitation. Otherwise, why
would Backus address the problem by proposing a new LANGUAGE (fp),
rather than a new computer architecture? If your AI program was
written in a language without side effects (such as PURE Prolog), the
opportunities for parallelism would be there. This would be
particularly welcome in natural language analysis with logic (or other
formal) grammars, because dealing with more and more complex subsets
of language needs an increasing number of grammar rules and rules of
inference, if the results are to be accurate and predictable.
Analysis times, even if they are polynomial on the size of the input,
may grow EXPONENTIALLY with the size of the grammar.
Fernando Pereira
AI Center
SRI International
pereira@sri-aists@ssc-vax.UUCP (Stanley T Shebs) (08/24/83)
Heh-heh. Thought that'd raise a few hackles (my boss didn't approve of the article; oh well. I tend to be a bit fiery around the edges). The claim is that we have "basically" solved the problem. Actually, we're not the only ones - the APE-II parser by Pazzani and others from the Schank school have also done the same thing. Our parser can handle arbitrarily ambiguous sentences, generating *all* the possible meanings, limited only by the size of its knowledge base. We have the capability to do any sort of idiom, and mix any number of natural languages. Our problems are really concerned with the acquisition of linguistic knowledge, either by having nonspecialists put it in by hand (*everyone* is an expert on the native language) or by having the machine acquire it automatically. We can mail out some details if anyone is interested. One advantage we had is starting from ground zero, so we had very few preconceptions about how language analysis ought to be done, and scanned the literature. It became apparent that since we were required to handle free-form input, any kind of grammar would eventually become less than useful and possibly a hindrance to analysis. Mr. Pereira admits as much when he says that grammars only reflect *some* aspects of language. Well, that's not good enough. Us folks in applied research can't always afford the luxury of theorizing about the most elegant methods. We need something that models human cognition closely enough to make sense to knowledge engineers and to users. So I'm sort of in the Schank camp (folks at SRI hate 'em) although I try to keep my thinking as independent as possible (hard when each camp is calling the other ones charlatans; I'll post something on that pernicious behavior eventually). Parallel production systems I'll save for another article... stan the leprechaun hacker ssc-vax!sts (soon utah-cs) ps I *did* read an article of Mr. Pereira's - couldn't understand the point. Sorry. (perhaps he would be so good as to explain?)