Dialogue system juggles topics

By Eric Smalley, Technology Research News

There's more to speech than just words and sentences. There are half sentences, stops and restarts, changes of subject and corrections.

It took three decades of speech recognition research to produce systems good enough to use as interfaces between people and computers. One way to make speech interfaces better is to find ways for computers to recognize the context of a conversation.

Researchers from Edinburgh University in Scotland and Stanford University have built a dialogue management system that promises to improve verbal communication with computers by giving the machine a sense of the type of phrase a person is likely to say next.

The approach could be used in a wide variety of speech recognition systems including telephone-based information systems, interactive entertainment devices, robots, computer interfaces for the visually impaired, in-car dialogue applications, and speech interfaces for personal computers.

The Conversational Interface Architecture goes beyond the slot-filling dialogue systems commonly used for airline ticket booking systems by tracking multiple conversation threads, said Oliver Lemon, a senior research fellow at Edinburgh University. Slot-filling dialogue systems prompt users to provide topic-specific information and listen for keywords that determine the system's response to the user.

The software follows multithreaded conversations -- those that switch back and forth between several topics -- without having to be programmed, regulates particular topics, and uses this information to improve speech recognition rates, according to the researchers. It also recognizes corrective fragments -- phrases that correct something a user has just said -- and it allows users to initiate, extend and correct dialogue threads at any time.

The system accomplishes this by tracking different types of utterances, including yes or no answers; who, what, where answers; and corrections like "I meant the office" and "not the tree."

Tracking the context of a conversation simplifies the speech recognition task by limiting the range of words the system must attempt to recognize. Existing dialogue management systems constrain speech recognition choices by limiting them to a certain topic like city names. The researchers' approach works with any application, however, because it uses type of utterance rather than specific topic to constrain speech recognition, said Oliver Lemon, a senior research fellow at Edinburgh University in Scotland. "The system anticipates what type of dialogue move the user will say next, and uses this information to tune speech recognition," he said.

Tracking fine-grained dialogue elements lowered overall speech misrecognition by 11.5 percent and concept misrecognition by 13.4 percent, according to Lemon.

The researchers are working to improve their system with better dialogue modeling software and better ways to parse text, said Lemon. Another step forward is to use machine learning techniques to automatically optimize dialogue strategies so that conversational interface systems can adjust themselves based on experience, he said.

The context-sensitive component of the researchers' system could be applied to practical applications now, said Lemon. Multithreaded dialogue management could be used practically within two years, he said.

Lemon's research colleague was Alexander Gruenstein from Stanford University. The work appeared in the September, 2004 issue of the Association of Computing Machinery (ACM) Transactions on Computer-Human Interaction (TOCHI). The research was funded by the Wallenberg Foundation and the Edinburgh-Stanford Link Program.

Timeline:   2 years
Funding:   Private, University
TRN Categories:  Human-Computer Interaction
Story Type:   News
Related Elements:  Technical paper, "Multithreaded Context for Robust Conversational Interfaces: Context-Sensitive Speech Recognition and Interpretation of Corrective Fragments," Association of Computing Machinery (ACM) Transactions on Computer-Human Interaction (TOCHI), September, 2004


April 6/13, 2005

Page One

Programmed DNA forms fractal
Dialogue system juggles topics
Scheme reverses light pulses
View from the High Ground:
Joan Feigenbaum

Tough material gets functional
Water shifts rubber's shape
Interference scheme sharpens focus
System forms light necklace
Trapped light pulses interact
Optics demo does quantum logic
Strained material cleans up memory


Research News Roundup
Research Watch blog

View from the High Ground Q&A
How It Works

RSS Feeds:
News  | Blog  | Books 

Ad links:
Buy an ad link


Ad links: Clear History

Buy an ad link

Home     Archive     Resources    Feeds     Offline Publications     Glossary
TRN Finder     Research Dir.    Events Dir.      Researchers     Bookshelf
   Contribute      Under Development     T-shirts etc.     Classifieds
Forum    Comments    Feedback     About TRN

© Copyright Technology Research News, LLC 2000-2006. All rights reserved.