In spoken dialog systems it is often the case that the sentence produced by the decoder with the highest recognition probability may not be the best choice for extracting the intended concepts. Lower ranking hypotheses may present better alternatives. In this paper, we show how to integrate multiple knowledge sources for the decision of selecting one of these hypotheses. A scoring schema combining information from the recognizer output, the parser, an utterance type classifier and dialog context is used. The scaling weights of the combined scores are determined automatically by an optimization procedure. Finally, we show the results of testing this approach and its performance compared to the approach of selecting the best recognition hypothesis.
展开▼