Third International Conference on Spoken Language Processing (ICSLP 94)
This paper describes a new syntactic/semantic grammar specification language and optimizing compiler that allows the user to write a single compact context-free grammar specification that unambiguously describes both the syntax and semantics for a given task. The compiler produces all of the necessary language processing components needed to efficiently couple speech to actions. The optimizations performed are specifically designed to significantly increase the efficiency of hypothesis scoring in recognizers. They include FSA determinization, minimization, transition consolidation and elimination of redundant epsilon-transitions as well as a transformation that eliminates left-recursive rules. While the primary target application is speech recognition, these tools are also being used for building stochastic language models for other applications like handwriting recognition. The grammars are written in a high-level language called GSL (Grammar Specification Language). The compiler generates both an optimized recursive transition network for syntactic constraints and a semantic processor that performs lexical semantic analysis. When driven by the results of the recognizer, the semantic processor dynamically constructs an evaluation tree that describes how semantic actions are to be performed. With the addition of a user defined library of semantic action functions or a simple interface to an existing application (e.g., DBMS), a complete limited-domain speech understanding system can quickly be constructed. Since the syntax and semantics are specified by the same language, any changes to the specification are consistently applied to the entire system.
Bibliographic reference. Brown, Michael K. / Buntschuh, Bruce M. (1994): "A context-free grammar compiler for speech understanding systems", In ICSLP-1994, 21-24.