Where does linguistic meaning fit into cognition? In this section we consider three possible positions one could take on this question and some of the implications of these positions.
All models must deal with the obvious fact that different languages have different semantic systems, that they divide up the world in different ways. One important distinction between models concerns the way in which language-specific semantics is hooked up to the rest of the cognitive system in such a way that linguistic behavior is roughly appropriate to the non-linguistic context.
Probably the most widely-held view is one in which linguistic meaning is viewed as a symbolic system which maps onto a universal and symbolic conceptual system. This conceptual system may be innate or learned pre-linguistically. For example, this is the view of two very influential frameworks, those of Pinker (1994) and Jackendoff (1992) . We will refer to models that adhere to this view as symbolic models of linguistic behavior and acquisition. On the symbolic view, language acquisition involves learning about the particular language's semantic structures and how they map onto universal conceptual structure. Conceptual structure is in turn related to perception and action in ways which are usually left unspecified. An overview of such a perspective is shown in Figure 1.
If this position holds, then language acquisition can be studied and modeled without taking perception into account. The acquisition of semantics is a symbolic phenomenon, relating one developing symbolic system, the semantics of the target language, with another, the existing conceptual system. Furthermore, since basic conceptual structure is in place before language is learned, this position does not allow for any significant influence of linguistic categories on concepts. Finally, because semantic development in any language is driven by the same underlying conceptual system, there should be strong similarities in the developmental course of the learning of meaning across diverse languages.
A range of alternatives to this popular view have been set forth in recent years. These positions agree in assigning more significance to the interplay between linguistic and non-linguistic cognition. We will refer to them as grounded models of linguistic behavior and/or acquisition. Grounded models are associated with cognitive linguists [Lakoff, 1987, Langacker, 1987a] and with other cognitive scientists who seek to do away with mind-body distinction in one sense or another [Harnad, 1990, Johnson, 1987, Thelen and Smith, 1994, Varela et al., 1991].
We consider here two possible positions within the space of grounded models. The first is that of Regier (1996), whose computational model of the acquisition of spatial relation terms is one of a small number of serious attempts to actually implement the grounding idea. In Regier's model, linguistic meaning is learned directly via perception, and acquisition can only be studied in the context of a model of the vision system. The nature of this system constrains the kinds of possible meanings that languages can encode and the way in which these meanings are learned by children. However, Regier's model makes a clear division between vision and language and has no obvious place for spatial concepts. Presumably the acquisition of spatial terms has little or nothing to do with spatial reasoning, which in any case is not under the influence of linguistic categories. Furthermore, the model only runs in the production direction; it does not tell us how a child learns to comprehend spatial terms. A schematic of Regier's model is shown in Figure 2. The figure does not assign a place to non-linguistic concepts; presumably these would exist in a component of the system parallel to language.
A second, more radical, option within the space of grounded models also has linguistic meaning grounded in perception. The difference here is that there is no distinction between linguistic meaning and non-linguistic concepts and that the model runs in both directions, from language to vision as well as from vision to language. Particular meanings/concepts may be learned in three ways: through non-linguistic perceptual and motoric experience, through a combination of non-linguistic and linguistic experience, and through linguistic input alone. This type of model allows linguistic categories to influence concepts; that is, various forms of linguistic relativism [Gumperz and Levinson, 1996] are possible. Finally such a model predicts that the developmental course of the learning of meaning should depend on the categories inherent in the language being learned. A schematic of this sort of model is shown in Figure 3. This is the view that we favor and the one that is realized in Playpen, the connectionist model we present later in this report.