Grounding compositional symbols: no composition without discrimination

The classical computational conception of meaning has been challenged by the idea that symbols must be grounded on sensorimotor processes. A difficult question arises from the fact that grounding representations cannot be symbolic themselves but, in order to support compositionality, should work as primitives. This implies that they should be precisely identifiable and strictly connected with discriminable perceptual features. Ideally, each representation should correspond to a single discriminable feature. The present study was aimed at exploring whether feature discrimination is a fundamental requisite for grounding compositional symbols. We studied this problem by using Integral stimuli, composed of two interacting and not separable features. Such stimuli were selected in Experiment 1 as pictures whose component features are easily or barely discriminable (Separable or Integral) on the basis of psychological distance metrics (City-block or Euclidean) computed from similarity judgments. In Experiment 2, either each feature was associated with one word of a two-word expression, or the whole stimulus with a single word. In Experiment 3, the procedure was reversed and words or expressions were associated with whole pictures or separate features. Results support the hypothesis that single words are best grounded by Integral stimuli and composite expressions by Separable stimuli, where a strict association of single words with discriminated features is possible.