ABSTRACT
The commitment to the concept of representation has been regarded as one defining feature of traditional cognitive science. In the symbol-processing approach internal representations have been conceived of in terms of syntactically structured symbols (Fodor, 1987; Fodor & Pylyshyn, 1988). Representational terminology is also widely used and explicitly acknowledged among connectionists, albeit in relation to dynamic networks of simple, neuron-like processing units (Rumelhart, McClelland, and the PDP Research Group, 1986; Smolensky, 1988). Despite this almost ubiquitous use of the term representation, it is far from settled what exactly is meant when semantic content or meaning is ascribed to internal states of symbolic or connectionist systems (cf. Cummins, 1989). Accordingly, it comes as no surprise that some of the most profound issues in the ongoing debate between the symbol processing and the parallel distributed processing approaches are related to the concept of representation. In the present chapter we discuss two questions that we consider of central importance for the development of a connectionist theory of internal representation:
The problem of semantic content. In what sense and in virtue of which properties can internal states in connectionist networks be said to have meaning or to be about something? That is, what is it that makes internal states in connectionist networks representations with a specific semantic content?
The problem of semantic compositionality. How can connectionist models account for the combination of simple representations into complex ones? In particular, how can connectionism explain the apparently productive and systematic nature of natural language?
