Third International Conference on Spoken Language Processing (ICSLP 94)
The dynamic probabilistic grammar (DPG) is a context-free grammar whose rule probabilities are dynamically controlled by a hidden Markov model (HMM). This HMM receives a rule number as an input symbol, and outputs the probability of the rule number sequence. The DPG parser generates plural rule sequences representing different syntactic structures. The parser calculates the probability for each rule sequence, and selects the rule sequence (i.e. syntactic structure) which achieves maximum probability. This disambiguation mechanism is also effective for grammar-based speech recognition systems. The number of candidate words (called perplexity) can be reduced effectively using this mechanism. The DPG provides a stochastic framework for CFG-class spoken language processing.
Bibliographic reference. Kawabata, Takeshi (1994): "Dynamic probabilistic grammar for spoken language disambiguation", In ICSLP-1994, 787-790.