Principal Differences in Structure-Mapping.pdfVIP

Principal Differences in Structure-Mapping The Contribution of Non-Classical Models Tony Veale, Mark T. Keane Abstract: Recent research in metaphor and analogy, as variously embodied in such systems as Sapper, LISA, Copycat and TableTop, speak to the importance of three principles of cross-domain mapping that have received limited attention in, what might be termed, the classical analogy literature. These principles are that: (i) high-level analogies arise out of nascent, lower-level analogies automatically recognized by memory processes; (ii) analogy is memory-situated inasmuch as it occurs in situ within the vast interconnected tapestry of long-term semantic memory, and may potentially draw upon any knowledge fragment; and (iii), this memory-situatedness f requently makes analogy necessarily dependent on some form of attributive grounding to secure its analogical interpretations. In this paper we discuss various arguments, pro and con, for the computational and cognitive reality of these principles. 1. Introduction Over the last few years, we have been examining the computational capabilities of models of analogy (see Veale Keane, 1993, 1994, 1997; Veale et al., 1996). Some models of analogy, like the original version of the Structure-Mapping Engine (SME; Falkenhainer, Forbus Gentner, 1989), have been concerned with producing optimal solutions to the computational problems of structure mapping, although more recently, many models have adopted a more heuristic approach to improve performance at the expense of optimality; models like the Incremental Analogy Machine (IAM; Keane