Representation and Efficiency in a Production System for Speech Understanding

Abstract

he rc!>earch quc~.tion !.ludir.cl i!> whether it is prilclical to usc a production system (PS) architecture for a ver!>ion of the He;.~r~.;~y-11 (11511) !q">~:>ech understandi ng r.ystcm. The study focuses on adequacy of representation, space efficiency and time efficicnc.y . 1\ running PS architecture called HSP was der.igned and implemented and twelve HSII knowledge !.ourcc~ were tran!>latcd to HSP productions, with two of the twelve actually run under HSP. Detailed compari~. ons of HSII ;mel HSP produced the following results: (1) HSP is adequate to represent the HSII speech knowleclae, even thn11ch HSP ir. a comparatively simple PS architecture; (2) HSP suffers a moderate space penalty for repre~.entinp, deciMativr ~ HSII knowledge, which is serious since HSII contains a high proportion of such knowledge; (3) rcpre:;cnti1tion of 11511 dcclariltivc knowlcdp,c as HSP productions causes problems with multiple use of that knowledge; (4) lilck of a local working memory in the HSP architecture has serious consequences both for space and time t-fficiPncy; and (5) HSP ha s a time efficiency handicap of two to three-and-a-half ordElrs of magnitude, in spite of efficiency mechanisms which make HSP comparable in time efficiency to other PSs. The paper ends with a set of que~t ion !> that remain to be answered for a thorough evaluation of a PS version of HSII.