BAYESIAN BELIEF NETWORKS AS A TOOL FOR STOCHASTIC PARSING

Bayesian Belief Networks are a powerful tool for combining different k nowledge sources with various degrees of uncertainty in a mathematical sound and computationally efficient way. Surprisingly they have not y et found their way into the speech processing field, despite the fact that in this science multiple unreliable information sources exist. Th e present paper shows how the theory can be utilized in for language m odeling. After providing an introduction to the theory of Bayesian Net works, we develop several extensions to the classic theory by describi ng mechanisms for dealing with statistical dependence among daughter n odes (usually assumed to be conditionally independent) and by providin g a learning algorithm based on the EM-algorithm with which the probab ilities of link matrices can be learned from example data. Using these extensions a language model for speech recognition based on a context -free framework is constructed. In this model, sentences are not parse d in their entirety, as is usual with grammatical description, but onl y ''locally'' on suitably located segments. The model was evaluated ov er a text data base. In terms of test set entropy the model performed at least as good as the bi/tri-gram models, while showing a good abili ty to generalize from training to test data.