Abstract

Within a semantic perspective, when an individual gives a meaning to a configuration of symbols, he ascribes to it both a sense and a reference (Frege, 1892; Chierchia, McConnell-Ginet, 1993). Taking into consideration the psychological reactives, it is commonly accepted that an individual must assign a meaning to an item, before he physically responds to it. That is, the response an individual gives to an item depends both on the sense and the reference he assigned to it. On the other hand, an analysis of the manifest responses given by an individual to a reactive, may carry out only the reference that the individual assigned to the items (i.e. if they are true or false), but not their sense. In this way the individual meaning is only partly captured, and the structure of the individual representation of the meaning still remains obscure. The current report takes into consideration the possibility to develop a connectionist model that allows to study the individual representation of meaning within the two dimensions of sense and reference. A 3-layer feed-forward neural network was trained to approximate the function that relates a set of sentences (items) generated within a fragment of Italian language, to the truth values (responses) assigned to them by a simulated subject. It is underlined that, by means of this method, it is possible to change from an "extensional" description of the individual meaning (i.e. the set of item responses), to an "intensional" description of it (i.e. the function identified by the net). In a first study, the structure of the internal (hidden) representation of the network was investigated. An analysis of the hidden structure revealed the ability of the network to identify analogies and associations within the individual representation. These analogies did not appear in a direct inspection of the manifest responses. In a second study, the discriminant validity of the model was tested. In this case the model was employed as an inference device to identify the individual's omitted responses and/or the responses to a set of items that were never presented to the individual. Here, the task of the network consisted to guess the right individual responses. The goodness of the model was tested by means of the binomial test in a series of 1440 (20 x 72) independent experiments. The goodness was found to be significant in both of the experiments (p < .01). In conclusion some remarks as regards possibile extensions of the model are carried out.