Broad Area Colloquium For AI-Geometry-Graphics-Robotics-Vision
(CS 528)

Collecting, err, Correcting Speech Errors

Mark Johnson, Brown University
November 15, 2004, 4:15PM
TCSeq 200


Disfluencies and errors abound in certain types of spontaneous speech, and cause problems for the computer speech applications such as speech recognition and question answering. This talk focuses on restarts and speech repairs, such as /you get, uh, you can get a system for $10,/ because these are particularly disruptive to our natural language parsing system. After explaining why these are so disruptive, we propose a noisy-channel model architecture for detecting and correcting speech transcripts which contain these kinds of errors. The channel model uses a novel Tree Adjoining Grammar model to describe the crossing dependencies that characterize these kinds of speech errors. By embedding these components in a machine-learning reranking framework we can adapt the model to new domains and further improve its performance.
Joint work with Eugene Charniak and Matt Lease.

About the Speaker

Mark Johnson received his PhD from the Stanford Linguistics department in 1987, was a Post-doc at MIT's Brain and Cognitive Sciences department, and became an Assistant Professor at Brown in 1988, where he has been on the faculty ever since. He has had sabbaticals and leaves of absence in lots of interesting places, the last of which was last year, when he was a visiting scientist at MIT's Computer Science and Artificial Intelligence Laboratory. His primary research area is Computational Linguistics, and he's especially interested in developing machine learning and statistical techniques to model the hidden compositional structures involved in natural language. He was President of the Association for Computational Linguistics in 2003.


Back to the Colloquium Page