Date of Award

Winter 2008

Project Type


Program or Major

Engineering: Electrical

Degree Name

Doctor of Philosophy

First Advisor

W Thomas Miller


A spoken question answering system that recognizes questions as full sentences performs well when users ask one of the questions defined. A system that recognizes component words and finds an equivalent defined question might be more flexible, but is likely to have decreased speech recognition performance, leading to a loss in overall system success. The research described in this document compares the advantage in flexibility to the loss in recognition performance when using component recognition.

Questions posed by participants were processed by a system of each type. As expected, the component system made frequent recognition errors while detecting words (word error rate of 31%). In comparison, the full system made fewer errors while detecting full sentences (sentence error rate of 10%). Nevertheless, the component system succeeded in providing proper responses to 76% of the queries posed, while the full system responded properly to only 46%.

Four variations of the traditional tf-idf weighting method were compared as applied to the matching of short text strings (fewer than 10 words). It was found that the general approach was successful in finding matches, and that all four variations compensated for the loss in speech recognition performance to a similar degree. No significant difference due to the variations in weighting was detected in the results.