LINGUIST List 7.1130

Mon Aug 12 1996

Sum: Working Dialogue Systems

Editor for this issue: T. Daniel Seely <dseelyemunix.emich.edu>


Directory

  1. Nicholas Ostler, Working Dialogue Systems: Summary

Message 1: Working Dialogue Systems: Summary

Date: Fri, 09 Aug 1996 16:50:10 BST
From: Nicholas Ostler <nostlerchibcha.demon.co.uk>
Subject: Working Dialogue Systems: Summary
On 17 June I put out a request for information as follows:

>As part of my research for a lecture on the Commercial Reality of Dialogue
>Systems, I would like to ask for pointers to any systems on sale to-day which
>incorporate a speech or text dialogue front-end.
>
>Ironically, in the past (mid 1980s) these were the applications of language
>technology par excellence, usually giving access to SQL databases, and names
>such as Intellect, Q&A, Themis and Parlance are well known, though
>surprisingly difficult to find running these days. It is much harder to find
>modern instances of the species, despite the flourishing state of the
>underlying theoretical field of dialogue simulation, ...

>Partly perhaps, the ground has been taken over by multimodal interfaces.
>("This may be dialog, Captain, but not as we know it.") But even there, there
>does not seem to be a lot available in the stores as yet.
>
>If people would be kind enough to send me references (and indeed opinions or
>anecdotes) on new systems that are beyond the research stage (preferably on
>sale), or old systems that are still available, I should be happy to publish a
>summary.

I received about 12 responses, about half of them just expressing interest
in the question and any answers I received. Thank you to the following
people who materially helped me in putting together a response:

Pierre Nugues, Wlodek Zadrozny, Norman Fraser, Madeleine Bates, Ken
Purcell, Rene Collier, Ehud Reiter, Harald Aust, Susann LuperFoy, David
Nash

As I feared, the list of fully commercialized and still available dialogue
systems is rather short. The most lively in the respect is the following:

- -----------------------------------------------------------------------------
My company, Philips Dialogue Systems, was established about 18 months ago in
order to further develop, market, and sell natural language inquiry systems
for operation over the telephone. These systems are based on the technology
developed within Philips Research during the years 1992 - 1994, mostly by
my colleague Martin Oerder (speech recognition, speech output, system
integration) and myself (speech understanding, dialogue control). The two most
important references for this system are

H.Aust et al.: The Philips Automatic Train Timetable Information System.
Speech Communication 17, pp. 249-262, 1995.

H.Aust, M.Oerder: Dialogue Control in Automatic Inquiry Systems.
Proceedings of the ESCA Workshop on Spoken Dialogue Systems, pp. 121-124,
Aalborg, Denmark, 1995. Also in: Proceedings of TWLT9, pp. 45-49, Enschede,
The Netherlands.

Our demonstrator has been operational and publicly available since February
1994, in German language, under the telephone number +49 241 604020. Swiss
Railways commercially operates one of our systems under the number
+41 157 0222. A number of systems for other customers, in a variety of
languages, are close to their introduction.

Note also our dialogue description language HDDL.

Harald Aust
Philips GmbH Dialogue Systems

Kackertstr. 10 Tel. +49 241 8871 150
52072 Aachen Fax +49 241 8871 140
Germany E-mail austpfa.philips.de
- -----------------------------------------------------------------------------
- -----------------------------------------------------------------------------

Others suggested accepting a wider definition of "dialogue":

- -----------------------------------------------------------------------------
Norman M. Fraser, Vocalis Ltd:
Ah well, it all depends what you mean by dialogue...

In the world of spoken dialogue, most of the commercial action is
in connection with systems having fairly small vocabularies and
serving pretty mundane purposes. Telephone banking, for instance.
Vocalis has supplied lots of these systems around the world, mostly
offering DTMF only dialogues (customer conservatism, I'm afraid).
However, we've got some speech recognition systems out there, most
noteably with Abbey National in this country.

Then we've done a few interactive systems for British Airways. Most
of them are internal to BA (used for things like BA staff travel booking
and seat availability inquiries). For a time some calls to BA's general
flight information service overflowed to a Vocalis spoken dialogue
system, though that is not the case at the time of writing.

Then there is our Operetta `virtual operator' product which interrogates
callers to find out who they wish to speak to (and such things) and
acts appropriately. Call me on the number below if you want to have a go.

We sell a voice dialling product called SPEECHtel which supports a
certain amount of spoken dialogue in addition to the basic voice dialling
function.

I'm not aware of any really whizzy conversational spoken dialogue
systems which have been deployed in anger, with the exception of the
Philips train timetable system for Swiss Rail.

Norman M. Fraser Vocalis Ltd
Manager, Advanced Prototypes Chaston House
 Mill Court
Tel: +44 (0) 1223-846177 Great Shelford
Fax: +44 (0) 1223-846178 Cambridge CB2 5LD
Email: normanvocalis.com United Kingdom
- -----------------------------------------------------------------------------
David G Nash:
I'm not sure whether you would include as a dialogue system one in which I
press numbers on a phone pad to get info as digitised speech. If so, read
on... I use one called DECTALK to get aviation weather. It is run by the
Aussie Taxpayers' Airservices Australia, but the synthesised speech has an
American accent (male for info, female for readbacks), and I'm guessing it
is a commercial system. Pretty basic, but a good feature is that one can
interrupt at any point and move on to another query.

David.Nashanu.edu.au
- -----------------------------------------------------------------------------
- -----------------------------------------------------------------------------
There were no clear documentary responses from the producers of the classic
systems which have been in the commercial domain: Intellect, Q&A, Themis,
Parlance etc. However, I did have some enlightening responses on the
current commercial view:
- -----------------------------------------------------------------------------

Wlodek Zadrozny, IBM:
I could give you a demo of our systems.
They are not for sale yet, but work reasonably well.

 IBM T.J.Watson Research Center,
 P.O. BOX 704
 Yorktown Heights, NY 10598.
 tel. (914) 784 7835
 e-mail: WLODZ  WATSON.IBM.COM

- -----------------------------------------------------------------------------

Ehud Reiter, Aberdeen
I saw your note about commercial dialogue systems posted to various mailing
lists. I would be very interested in finding out about the results of your
survey, please let me know when you have this. For what its worth, here are
a few vague/random thoughts of mine on this issue:

1) Q&A is still around, indeed I saw it on sale in a software store a few
months ago. Of course, it is unclear how much the NL side of Q&A is really
used.

2) One might postulate that there's still a considerable amount of industrial
R&D in NL interfaces to databases, its just less visible than it used to be.
For example, I know that the Access group in Microsoft has some good NLP
people, but they keep very quiet about what they're doing for commercial
reasons. I also know people in other (US) industrial labs who are working
on adopting NL dialogue ideas to multimodal interfaces - but again, this
is commercially sensitive and they can't say much about this.

3) I must admit that my own experience (which is in NL Generation systems)
suggests that it sometimes is easiest to make a good business case for NLP in
document creation/management/processing/etc, which unfortunately does not
tend to involve much dialogue. I used to work for a company which
specialized in applied NLG systems, and generally found it much easier to
"sell" an automatic document creation tool than any other kind of system.

4) I looked at the Web page for the Twente Conference you mentioned in
your message, it seemed to mainly involve spoken language dialogue systems.
I gave an invited talk last week to a workshop on the Dutch "TST"
spoken dialogue system, and this seems like an interesting application area
with potential for the future - but I believe that its just becoming feasible
now to build such systems, so its a bit early to expect working commercial
products.
 Ehud Reiter
 email: ereitercsd.abdn.ac.uk

- -----------------------------------------------------------------------------
Madeleine Bates of BBN:
Parlance is not a currently supported product. We have turned our
attention to speech-based products, notably a large-vocabulary, real-time,
speaker-independent, continuous speech recognition system called Hark.
You can find out more about it on the BBN web page, http://www.bbn.com

Since the late 80's, much of the best research done in regard to dialogue
systems was done under the ARPA spoken language systems program. BBN was
one of the contractors, as were MIT, SRI, and CMU. Papers on all the
systems can be found in a series of workshop proceedings published by
Morgan Kaufmann. The common domain that was used by all the contractors
was ATIS (Air Travel Information System), and all of the systems that were
developed were quite powerful and impressive. That work hasn't quite made
its way into the commercial marketplace yet, but it is very close.

I don't know that I have any "received wisdom" about commercial natural
language interfaces, but I'll be glad to mention a few generalizations.
Certainly one truism is that marketing is at least as important, perhaps
more important, than technical capability. User acceptence depends
heavily on many factors that have nothing to do with the linguistic
capability of the system (for example, whether the time to produce an
answer is predictable, or how easy the system is to configure and update).

If you include speech systems in your report, you will find a lot to
choose from. There are many applications that can be improved by speech
recognition interfaces, without much in the way of understanding or
complex dialog capabilities to support them. I expect that as speech
interfaces become more common, more people will start demanding the kind
of capabilities that only an "understanding" system (as opposed to a
"recognition" system) can provide. It might be that this demand will be
the "user pull" that will be more successful than the "technology push" in
getting dialogue systems to the market.

Dr. Madeleine (Lyn) Bates
Assistant Manager, Speech & NL Dept.
BBN Systems & Technologies voice: 1 617 873 3634
70 Fawcett St. fax: 1 617 547 8918
Cambridge, MA 02138 USA email: batesbbn.com
- -----------------------------------------------------------------------------
- -----------------------------------------------------------------------------

The final list of relevant WEB SITES I gathered is

http://cs.uwp.edu:80/staff/haller/Activities/etce97.html Call for
Papers, Energy Week Conference and Exhibition, Session on Natural Language
in Human-Computer Interfaces, Houston, TX, January 28 - January 30, 1997

http://www-csli.stanford.edu/csli/projects/interface9495-srct.html
Social Responses to Communication Technologies / CSLI / Stanford University

http://www.ttt.bme.hu or else
http://www.cogsci.ed.ac.uk/elsnet/summerschool96.html The current
Budapest summer school

http://wwwis.cs.utwente.nl:8080/mars/ECAI96.html Dialogue Processing
in Spoken Language Systems, August 96 at ECAI, the European Conference on
Artificial Intelligence.

http://wwwseti.cs.utwente.nl/Parlevink/twlt/ University of Twente
Conference on Dialogue Management in Natural Language Systems, June 1996

- -----------------------------------------------------------------------------

These are mostly relevant to conferences but there are three sites where
ongoing work can be viewed, and in one case (CSLU) downloaded:

http://www.cse.ogi.edu/CSLU/toolkit/toolkit.html
http://svr-www.eng.cam.ac.uk/~kjp22/Personal.html
http://www.ensicaen.ismra.fr/~nugues

- -----------------------------------------------------------------------------
- -----------------------------------------------------------------------------
And my own two-penn'orth? Well, I incline to the view that economically,
Dialogue is a Luxury Good within language technology: it appeals to
institutions with a low marginal utility of money.
Specifically, this means
 "the rich": e.g. successful IT and communications companies
(Canon, Microsoft, British Telecom, NTT ...),
 military institutions (MITRE Corp.on behalf of US
Dept of Defense ...), and
 "the disinterested" (e.g. universities, and government funding
programmes - on a good day)

The problem with the systems which have already made it to the marketplace,
at least in text-based access to databases, seems to be that the back-end
systems they access have been updated, but the natural language front-end
has not: paradoxically, the NL part of the system is, over the long term,
the most rigid.

It seems that using a dialogue system is like J.L. Borges' view of
friendship between Englishmen: it begins by excluding confidences, and ends
by dispensing with conversation.


 Nicholas Ostler
Managing Director President
Linguacubun Ltd Foundation for Endangered Languages

 Batheaston Villa, 172 Bailbrook Lane
 Bath BA1 7AA England
 +44-1225-85-2865 fax +44-1225-85-9258
 nostlerchibcha.demon.co.uk
Mail to author|Respond to list|Read more issues|LINGUIST home page|Top of issue