What Works,What Doesn’t --And What Needs to Work Lynette HirschmanInformation Technology Center The MITRE Corporation USA
What Works... • If we believe Eric Brill, we should just collect and annotate data... • Since data collection seems to work better than looking for new algorithms • What this really means is that data collection is more cost-effective than funding research • Similarly, we might conclude that waiting for the next chip is more cost-effective than creating faster algorithms So we should all stop doing researchand look for data and wait...
Conversational Interaction: A Case Study • Speech researchers have always said, there’s no data like more data • Many speech problems are, by definition, data-constrained: • Conversational interfaces require real(istic) data on what people will say to machines in the context of a specific application • Such application-specific data tends to be difficult (expensive) to collect • It requires simulation of interaction with a system, or a running system to collect data to build the system… • How do we collect millions of sentences of application specific data?
How to Collect Real Data Cheaply • Lesson from Victor Zue’s MIT Jupiter system: • Put something out there that people want to use: on-line weather information • This can be done by bootstrapping from a primitive system, using the collected data • MIT has been very successful in collecting data from real users; methodology now used by the DARPA Communicator program So to collect data to build a system,we need a system that works well enough for people to use it
Real Systems To Collect Real Data... • Building a usable system requires integration of multiple technologies: • We need ways to interface to real data sources • We need language understanding • We need intelligible generation and synthesis • We need dialogue management • We need ways to apply the techniques to a different problem domain (application portability) because otherwise, we have to do all this again for the next application • So collection of real data raises basic research issues
Error Rate Over Time in ATIS(Air Travel Info System) Understanding easier than transcription Limiting factor: understanding, not word error Sentence Transcription SL Error NL Error Word Error Error rate log scale Time (months)
Conclusion: What Needs to Work • So we can’t just wait for data -- we need to collect it • And to collect data, we need systems that work so that real users will use them; they must be: • Scalable to handle large amounts of data • Robust so they keep working • Fast, so people can stand to use them • Interactive and engaging, so people want to use them • And while we are at it, it would be nice if the systems not only supported data collection, but were able to learn interactively…