SIGdial: Stray Thoughts from a Crossroads

SIGdial: Stray Thoughts from a Crossroads

By Hans Dybkjr, Prolog Development Center A/S, Brndby, Denmark
(in: Elsnews: The Newsletter of the European Network in Human Language Technologies) 10.3, Winter 2001, http://www.elsnet.org/elsnewsarchive.html)

Hans Dybkjr gives a commercial-eye-view of the 2nd SIGdial Workshop, held on September 1-2, 2001 in conjunction with Eurospeech 2001.

This years SIGdial workshop took place in Aalborg, Denmark, and I must say that it was very well organised.

A great thing about dialogue and discourse is that, on the one hand, it is deeply theoretical and philosophical, and on the other, it addresses practical, commercial-level problems of satisfying system users. Looking at the audience of this workshop, you see both theorists and engineers, and you know that over the years they will move back and forth across the boundaries of academia and industry.

Im one of them. Today I am "out there", building commercial information-service systems using the matured forms of the kind of tools and platforms we had in research five to ten years ago.

So, what am I after? Well, I need research and researchers to provide me with wider insights, concepts, and methodologies: hard evidence to steer my work. In this respect, the rhetorical structure theory tagging experiments of Carlson et al, are close to perfect: the theory itself seems conceptually appealing, and the careful investigation of the validity of this by extensive intercoder reliability experiments is something every body should learn from. Though we still fail to see dialogue control tools reflecting RST, this kind of experiment tells me that I can robustly rely on insights gained from RST.

As a working practitioner I am drawn to everything that tells me I can do things — and in a simple way — and to everything that describes the problems I can expect to meet. And advice needs to be given by people who approach problems at the same level as I do (although that is not to imply that I can solve them as elegantly). From this perspective I loved the reports from the Spanish railway information systems by San-Segundo et al, providing, for example, a cookbook on graceful degradation. And the dauntless practical systems presented by Rudnicky are just pure gold. But I also listened carefully to empirical work like Cavazzas recognition errors or to Aberdeens reports on the hopeless users in their experiments.

In my daily work, I not only use state-of-the-art tools, I also push them to the limits of their intended use. So, when Hirschberg convincingly told us that prosodic cues today can point out the cor rection awareness sites in practice, I immediately asked Philips (because I use their otherwise nice SpeechMania platform) if they could kindly provide me with that technology. And I am mathematically inclined to love the turn-minimisation criteria described by Yasuda et al, and to believe they could be incorporated into a dialogue control tool: so why arent they there?

This brings me to the issue of crossing the gap from theory to practice. The turn-minimisation is formalised and can be made operational. Dont forget that at the bottom of software engineering lies programming, which is nothing but applied, extremely formal logic. But what about the prosodic cues? Researchers may need to create the tools, not only as a necessary part of the research projects, but also to convince companies about the practical applicability of their work. And to mature the frameworks: the "plugnplay" ideas for grammars propounded by Lewin et al sound attractive and are, engineering-wise, desirable. I am, though, not convinced that they have got the abstractions right — the interfacing of their grammars makes me suspicious.

In gener al I will not use research-produced software. I love to try it out, and also enjoyed doing that at the workshop. However, at the end of the day I use software that has been made by professionals who know about basic engineering skills like regression tests and version control, and who are geared to support.

Summing up, the mix provided by this years SIGdial workshop was great, although I might have wished to see some more hard and weird theory and more of the theory subjected to practical and empirical scrutiny.

But at the end I just want to say, to the organisers and not least to the contributors: Keep up the good work!

For information

Hans Dybkjr is software developer and project leader at the Prolog Development Center A/S
Email: dybkjaer@pdc.dk
SIGdial Web site: http://www.sigdial.org/
The proceedings of the workshop, which include all papers referred to in the text, can be obtained in print from the ACL, and are also available from the SIGdial web site at www.sigdial.org/sigdialworkshop01/