john@thelonius.mitre.org (John D. Burger) (06/22/91)
sharp@cpsc.ucalgary.ca (Maurice Sharp) writes:
The problem here is that people tend to assign a full range of
abilities based on the demonstration of only a few. That is, if the
system demonstrates an ability to understand NL, people will assume
it can understand normal english conversation. Since the state of
the art is not even close to this (except in very limited domains or
toy systems), the students will start using phrases that will not
work.
This is a real problem. It's exacerbated by the fact that everybody
thinks they're an expert in natural language (and, in some sense, they
are). For instance, I've heard users of NL systems say,
"Well, if X works, why doesn't X' work? After all, they're almost
exactly the same thing."
Unfortunately, X and X' have little to do with one another under the
surface, linguistically speaking. The result is a system that, as far
as the user is concerned, breaks randomly on 50 percent (or more) of
the user's input, so the user uses it 0 percent of the time.
As a general comment, the use of NL is overrated. In some highly
constrained domains it may make sense, but as a general user
interface tool for any community, the technology just is not there.
You are better off designing a system that meets the needs. In other
words, you are designing a tool to help get a task accomplished
(perhaps teaching addition). Find out a specification of that task,
and build a tool to support it.
It's certainly true that NL is not the right tool for most jobs, but I
think it can be an important part of an interface's arsenal of
modalities. NL generation, for instance, can be quite useful in
summarizing a database, or describing an event to a user, and the
technology is nearly good enough. NL input allows users to frame very
complex queries that would be nearly impossible to construct with a
menu interface, or even the proverbial SQL-like query language.
Unfortunately, the technology on the input side isn't quite there yet.
But the point is, an interface should give a user a variety of ways to
interact with a system, and allow her to choose the one she thinks is
most appropriate.
--
John Burger john@mitre.org
"You ever think about .signature files? I mean, do we really need them?"
- alt.andy.rooney