Abstract

this paper we will discuss some example sentences that illustrate the predictive/data-driven distinction and show how CIRCUS handles them. We further contend that any sentence analyzer which does not make a predictive/data-driven distinction must be either finessing a large class of problems, or trying to stretch a single processing mechanism farther than it can reasonably be expected to go. When viewed in terms of linguistic models that are not process-oriented, we will see that our predictive/data-driven distinction does not carve up the world along the same lines as a linguist might. For example, linguists often view the problem of prepositional phrase attachment (pp-attachment) as a problem that can be resolved in a purely syntactic manner (Frazier & Fodor, 1979). When semantic considerations are introduced, they are usually based on lexical preferences from verbs (Ford, Bresnan & Kaplan, 1981). In CIRCUS we will see that some pp-attachments are resolved by predictive semantics while others require data-driven semantics. In other words, some pp-attachment problems can be solved by standard symbolic methods, while others are best addressed using subsymbolic techniques. While we cannot hope to give a full technical description of CIRCUS in this paper, we will focus on those aspects of CIRCUS which are most interesting from the perspective of "high-level connectionism," and hope that our broad description is sufficient to convey a general sense of the overall system design. 0.2 Syntactic Processing

Citations