LR parsing of probabilistic grammars with input uncertainty for speech recognition

Abstract A shift-reduce parser for probabilistic context-free grammars is described, based on the LR algorithm. Each of the standard types of LR parser generator has a probabilistic version and a Bayesian interpretation is advanced. A graph-structured stack permits action conflicts and allows the parser to be used with uncertain input, typical of speech recognition applications. The sentence uncertainty is measured using entropy and is found to be significantly lower for the grammar than for a derived first-order Markov model.