An LR parser-based probabilistic language model for spoken language systems
David Goddeau · Medical Entomology and Zoology · 1993
Language models are an important component of modern spoken language systems. A major function of these language models is to provide constraints for the speech recognition process, reducing the recognition error rate. These constraints include both local word co-occurrence constraints and non-local syntactic constraints. In addition, language models produce semantic interpretations of spoken utterances for speech understanding. In many current systems, these functions are split among multiple language models; one model to apply local constraints during lexical access, another, used as a post-processor, to provide non-local constraints and generate semantic interpretations. The goal of this work is to develop a language model which can provide both local and long-distance language constraints and which can be efficiently integrated into lexical access search algorithms. This thesis presents a probabilistic language model, the PLR language model, which uses an LR parser to map sentence prefixes into equivalence classes. These classes are used to compute next-word probabilities for speech recognition. The model combines the ability to apply long-distance syntactic constraints and the ability to acquire co-occurrence information from a corpus of training data. The PLR model generalizes the familiar n-gram models and reduces to these models in the absence of grammatical knowledge. This thesis introduces the PLR model and investigates its relationship to other probabilistic language models. The model is evaluated in two different spoken language system domains. Among the issues addressed are: the performance of the PLR model in speech recognition and understanding tasks, the full integration of the PLR language model with lexical access, and the development of a language model with 100% coverage of the task domain from a grammar-based model with incomplete coverage. (Copies available exclusively from MIT Libraries, Rm. 14-0551, Cambridge, MA 02139-4307. Ph. 617-253-5668; Fax 617-253-1690.)