The human semantic potential: Spatial language and constrained connectionism By Terry Regier (review)

Abstract

REVIEWS149 The human semantic potential: Spatial language and constrained connectionism. By Terry Regier. Cambridge, MA & London: MIT Press, 1996. Pp. xiv, 220. $37.50. Reviewed by Eugene H. Casad, Summer Institute of Linguistics This reworking of Terry Regier's doctoral dissertation is a substantive addition to the connectionist approach to modeling linguistic phenomena and to the analysis of spatial language.1 It is also only the second work of which I am aware that implements consistently Langacker's approach to cognitive linguistics. The other is Holmqvist 1993, which implements Langacker's approach from a rather different technical angle. Whereas R employs a bit-map based system, Holmqvist developed and implemented a model based on mathematical logic.2 R's multidisciplinary and cross-linguistic approach lends immediate credence to the entire study as does his repeated concern for motivating the distinct elements incorporated within the computational model that he employs (103-4, 119-20). Finally, R covers his bases very well and keeps his claims modest and realistic. By the 'human semantic potential', R means the capacity that all human beings have for perceiving and conceptualizing relationships between entities and the situations in which they are localized. This capacity allows speakers to extract meanings based on these conceptualizations (4, 152, 186, 193). These acquired meanings are conventionally established in systematic ways through this semantic potential; all such conventionalization implies more players on the stage than simply a single language learner. R's approach to characterizing the human semantic potential asks what kind of mechanism could adapt itself to the very different conceptualizations of space that we can observe in the world's languages. He seeks to show how it is possible that this is achieved by language learners even though basic research shows that children acquire proficiency in this area with only a minimum of negative feedback (59). Finally, R is interested in determining what an adequate model of the human semantic potential could tell us about semantic universals (152). R develops both a computational model and a model of spatial semantics. The computational model is one that R calls 'constrained connectionisms', a version of network modeling that incorporates a significant amount of information about specific domains into the basic kind of network associated with the notion of parallel distributed processing (PDP) (81). The extent to which information from all these domains coincides correlates with the degree to which a given usage of a term is prototypical or not (192). The computational device is a multilayered relatively unstructured pattern recognition device in which are embedded orientation combination structures , map comparison structures, and motion trajectory structures. These structures are jointly responsible for bringing the 'constrained' into R's version of Feldman's 'structured connectionism ' (119). There are two outside layers to this device, a current layer and a top layer. The third element of R's computational device is a training algorithm called Quickprop (cf. 140) which operates via back propagation under PDP networks. Back propagation refers to a continuous process of calculating weightings attached to distinct vectors of information that reflect specified evidence coming from a variety of sources of information that is needed to make a proper classification (87). The hoped-for result ofback propagation is an output classification of perceived relations that matches the relations in the training set. R's model incorporates general architectural principles ofhow linguistic systems can categorize and symbolize spatial concepts (81). Although R begins by confining his focus to objects that have convex surfaces, he includes objects with nonconvex surfaces in Ch. 7 (181ff.). Orientation combination refers to the melding together of several different kinds of evidence in order to describe how a trajector is oriented relative to a given landmark (89). Map comparison is used 1 My thanks go to Joe Grimes for comments made on a previous draft of this review, to Ron Langacker for some related discussion, and to Kenneth Holmqvist for a complimentary copy of his work. 2 Strictly speaking, Holmqvist did not implement Langacker's model but rather based his implementation on that model (Holmqvist 1993:3). Another related work is that ofGeorge Dunbar, who has drawn extensively on Langacker's approach for his work on the cognitive lexicon...