Intention-based probabilistic phrase spotting for speech understanding

We present an approach towards probabilistic phrase spotting for evaluating a speech recognizer's utterance hypotheses for inferring the user's intention. The evaluation is done by mapping each word chain on each intention of the intention space. Therefore, we create an intention model for each intention as the basis for analysis. As the words of the speech recognizer's utterance hypotheses are assigned confidence levels, we treat these inputs as uncertain observations. We use Bayesian belief networks as a mathematical foundation for intention modelling and probability theory for evaluating such word chains. The algorithm considers syntactical and semantical relations between the words within a phrase, evaluating words regarding previously observed words of the current phrase.