A New Approach for Providing Natural-language Speech Access to Large Knowledge Bases

Constructing speech interfaces to large knowledge bases is difficult. Increasing the scope of the knowledge base often results in a decrease in speech-recognition accuracy. Re-engineering the input language to improve accuracy often necessitates non-trivial modification to the knowledge-base query processor. This problem is compounded if sophisticated techniques are used to analyze input to obtain context information to guide the speech recognizer. A partial solution to this problem has been developed. The knowledge base is divided into a collection of speech-accessible hyperlinked “sihlos” which are distributed over the Internet. Each sihlo has an associated grammar which is downloaded by the speech browser and used to configure the recognizer for that part of the knowledge base which is in scope. Sihlos are constructed as highly-modular executable specifications of attribute grammars. Re-engineering of the input language is still often required. However, concurrent modification to the sihlo query processor is now much easier. Experimentation with a prototype implementation has identified some linguistic issues which appear not to have been addressed elsewhere. The prototype provides a unique laboratory for the study of these issues.