Pragmatics, cognitive science and connectionism

Abstract

The article looks at two types of computer simulation of mental process: the so-called ‘classical’, serial architecture based on discrete input categories and a compositional syntax and semantics, and the ‘neural network’ approach based on connections that determine input-output relations. After giving an outline of the main features of the two types of architecture, we discuss the claims made on their behalf, and try to sum up their strong and weak points as candidates for duplicating human processes of understanding. A key point in the argument is the discussion of the ‘recognition of intention’ aspect of human understanding, where we argue that computer simulation can only handle intentions in terms of plans that can be taken as fixed in advance, while human intentions can only be understood as based on an assumption of freedom of action, implying notions like responsibility and free will. Having suggested how different aspects of language understanding fit into this picture, we conclude that no matter how far computer simulation proceeds, the inherent discrepancy between the status we attribute to a human subject (as part of our basic pragmatic competence as fellow subjects) and the status we attribute to a computational process (as part of an inherently controllable human plan) makes it contradictory to assume that a full understanding can take place by simulation. The contradiction,however, is not theoretical: it emerges from the ground rules of human practice.