Efficient Representation of Linguistic Knowledge for Continuous Speech Understanding

This paper describes a linguistic knowledge representation technique suitable for reducing analysis time and memory requirements in a parser for continuous speech. Parsing speech, having to process a lattice of word hypotheses instead of a string of words, involves a tremendous amount of search and the generation of a high number of phrase hypotheses. The aim is, while using powerful and flexible formalisms for syntax and semantics, to generate "compact" phrase hypotheses, each one accounting for many syntactic rules simultaneously. The proposed method is able to cope with, and to take advantage from, the fact that short words are often missing from the lattice. A detailed example is given to clarify this method. Finally experimental data arc presented and discussed, showing the effectiveness of the proposed technique.