\]'he Commercial Application of
Natural Language \[nLerfaces
Harry Tennant
Computer Sciunee Center
Texas insLruments
Dallas,
! don't think that natural language
interffaces are a very good idea. By that \[
mean conventional natural language
interfaces -- the kind where the user" types
in a question and the system tries to
understand it. Oh sure, when (if?)
computers have world knowledge th,'~t is
comparable to what humans need to
communicate with eactr other, natural
language interfaces will be easy to build
and, depending on what else is available,
might be a good way to communicate with
computers. But today we are soooooo fat
away from having that much knowledge in a
system, conventional natural language
interfaces don't make sense.
There is something different that makes
more sense ---- NLMenu. It is a combination
of menu technology with natural language
understanding technology, and it e\]iminates
many of the deficiencies one finds with
conventional natural language interfaces
while retaining the important benefits.
~ince a eonvewLional natural language
interface invites the user" to type in
whatever he wants, it is based on the
assumption that it will be able to
understand just about anything that the
users are likely to say. All systems today
deal with limited domains of: discourse. I
am convinced that users are likely to type
in al\] kinds of things. The probability
high that users will not be understood.
They eou\]d be taught the limitations of
linguistic and concepLu~\] coverage of the
system, but: a major motivation of building
naJtural language interfaces is t:o allow
effective use without t~'aining or
retraining. So, it doesn't seem like a
very good idea.
is
The assumption behind NLMenu is the
opposite. It assumes that there are ali
kinds of th~ngs that the users would like
to ask but that the coverage is so limited
that i£ is best to reveal the coverage
(limitations) to the user'. He then (:an
find quickly that what he wanted £o ask
cannot be asked of this system, so he'll
give up quickly, minimizing his
frustration. Or he might find that what he
wanted to ask can be asked and the system
helps him ,ask it in the way it will
understand. There is another importarrL
advantage: there may be things that the
user did not imagine that he could ask
about. NLMenu revea\]s these to the user
encouraging him to make full use of sysLem
capabilities, Conventional natural
language systems do not.
NLNenu works by displaying a collection of
menus of words and phrases on the screen.
The user builds sentences by selecting
words and phrases from the menus. \]-he
menus are driven by a parallel parser which
maintains all parses of the sentence
fragment constructed so far. AFter each
word or phrase is selected, the parser
Texas
looks ahead in the grammar for all the
phrases that could come next for each
parse. A new set of menus is constructed
with these phrases, which are then
displayed £o the user for his next
selection. In this way, the NLHenu system
is constraining the user to constructing
only those ,'{ententes that are parsabie with
the grammar. By including semantic
constraints, one can also constrain the
user to the conceptual coverage o¢ the
system. In our" current implementations, we
use semantic grammars so syntactic and
semantic constraints are conveyed simply.
Many other techniques can be imagined.
Because the NLMenu technique rests on the
same technologies as conventional natural
language interfaces, they have 'Lhe same
expressive power --- one can say the same
range of things with either approach. But
since NLMenu is a system-initiated dialog,
the system always knows what the user is
trying to express. This makes it very easy
to mix natural lanquage (sentence building)
with other- interface techniques such as
graphical input, forrn fi\] ling and others.
For" example, in an application involving
airports, when the user was about to enter
the location, a map popped up, the user
pointed at the area of interest, then the
map went away and the coordinates of that
\]oeation were textually inserted into the
sentence. The user then continued building
the sentence. This allowed the use of text
where appropriate (specifying runway
lengths, location names, e~c.) and graphics
where appropriate (specifying locations).
It .seems £o me that there is much more that
user interfaces ~-an gain from natural
language research. For exeJmple,
cooperative response is a good idea
independent of whether an original query
was expressed in English or a formal query
language or through some other means.
Similarly, repeated reference is important
in any extended dialog. Discourse objects
should remain available for terse
reference. There is nothing that limits
repeated reference to natura\] language
dia\]ogs. Ideas based on focus and dialog
s~ruct:ure can be app\]ied to dia\]ogs
mediated throu!\]h all sorts of interface
languages. We seem to be concentrating on
reproducing the form of human communication
and ignoring the substance: large
vocabularies, concept creation through
reference, modificaJtion and analogy,
mechanisms I:hat use context to gain
terseness and allowing dia\]ogs to
accomodate the nonlir~ear characteristics of
human thought. Natural \]anguage researc\]h
has much more Lo offer" the world than
simply a means for interpreting typewritten
commands, yet we as a field have
accomplished little toward influencing the
other user interface technologies.
167