Selective Attention and the Acquisition of Spatial Semantics

/ l / / / / l Selective Attention and the Acquisition of Spatial Semantics James M. Hogan, Joachim Diederich and Gerard D. Finn NeuroComputing Research Centre, QUT, GPO Box 2434, Brisbane, Q, 4001. {hogan, j oachim, gerry}@f it. qut. edu. au Abstract The acquisition of the semantics of natural language spatial terms is considered within the cognitive framework introduced by (Lan- gncker, 1987), and the computational frame- work of the Berkeley L0 project (Feldman et. al., 1990). We describe a computational model which incorporates selective attention mech- anisms to facilitate the identification of sig- nificant objects within the visual field, and their consequent binding to linguistic rela- tional identifiers (for example, the trajector and landmark) according to the conventions of the input language. In contrast to previ- ous work in this area, the approach allows ex- tension of the system to more sophisticated (potentially cluttered and feature-laden) in- put scenes and referential linguistic phenom- ena, without a major redesign of the system. The application of the model to lexemes de- scribing static concepts such as the English above, below and in is discussed, as are exten- sions to dynamic concepts. 1 Introduction This paper is concerned with the acquisition of nat- ural language spatial semantics by a neurally plau- sible connectionist system. Within the cognitive framework introduced by (Langacker, 1987), elemen- tary spatial concepts (such as the English above) are characterised by locative relations between a poten- tially mobile object called the trajector (TR) and a static reference object called the landmark (LM). Previous computational investigations of this prob- lem (Regier, 1992), have relied upon highly struc- tured feature detection systems and the abstraction of object identification issues into the input data. While highly successful on their own terms, systems of this nature are not readily generalisable to prob- lems involving more sophisticated (especially clut- tered) input scenes and linguistic phenomena, and provide neither a conscious nor an autonomous se- lection mechanism through which such inputs may be successfully processed. The model discussed below resolves some of these issues through the use of mechanisms of selective visual attention, through abstraction of established models from computational neuroscience (Niebur and Koch, 1997) and extension to allow linguistic input to cue selection and scene parsing. While re- taining the overall computational philosophy of the Berkeley Lo project (see section 2), the present work does not rely upon feature pre-processing to the same extent as the Regier system - representations being based upon probabilistic receptive fields. In this way, 'prior knowledge' of limited specificity may be employed through higher level recruitment to rep- resent quite complex relations (see section 5.1 and (Hogan and Diederich, 1994), (Hogan and Diederich, 1995)). The computational philosophy of the Berkeley Lo project is introduced in the next section, followed by discussion of the Regier model and the importance of explicit object recognition in the light of evidence from early language acquisition. Section 2.4 relates this discussion to an accepted cognitive theory medi- ated through binding of representations at …

Content truncated to 500 words