Colloquium - Byrne

When ASR technology is embedded in complex information systems the overall system performance will be measured not by the number of words correctly recognized but by specific evaluation criteria that depend on the task. For example, a cellular phone user might want only to dial a particular number, while a person searching audio archives might wish to know how ASR errors influence search performance as evaluated by information retrieval measures such as precision and recall. In each of these tasks it is not necessary for the ASR component to recognize every word since much that is said will be ignored by subsequent processing steps. It is therefore desirable to create ASR systems that are tuned to find the words and phrases that are important for particular tasks. Minimum Bayes-risk (MBR) ASR attempts to achieve this by minimizing the empirical expected risk under loss functions that describe desired system behavior. The optimal form of these decoders is well-known from decision theory and approximate algorithms have been found that improve task-specific performance when compared to task-independent Maximum Likelihood decoders. The MBR approach will be presented and methods of building systems under particular loss functions will be discussed. Applications to ASR and other information processing problems such as statistical machine translation will be described.

Hosted by Daniel Jurafsky.
Refreshments will be served immediately following the talk in ECOT 831.