From Action to Symbols and Back: Are There Action Symbol Systems?

We present an experimental and simulative method for exploring the properties of symbolic action representation. This work is related to the current debate about the compositional vs. holistic nature of such representation. We designed a new experimental setup to be used with human participants and a neural network model as well, in the context of an artificial language learning whose referents are actions. In different conditions, (a) actions are systematically connected with sentences made of arbitrary words, following a simple syntax, and consistently expressing features of involved movements; (b) actions as a whole are consistently associated with arbitrary words; (c) actions are consistently associated with other actions so that the former become symbols for the latter. In such tasks, results show a clear advantage for a holistic representation; there are reasons, however, to suppose this might be a first step towards a compositional representation. This analysis is supported by results with nets.