Abstract

Consider time as a total order among discrete events. If we look at human languages, a total order is always established among discrete events, which are the distinct pronunciations of the words (and morphemes they are formed by) in a sentence. Notice that this is not a necessary condition: if we think of the total order as a restriction imposed by the spoken modality, we could image a sign language in which, for instance, two arguments (e.g. «Eva» and «the apple») of a certain predicate (e.g. «eats») could be co-articulated exactly at the same time in space, one with one hand and another with the other hand in a temporally and spatially symmetrical situation. To our knowledge this possibility is limited by several restrictions in any sign language around the world. So we assume that «time» is an essential ingredient of any human language and must be included in a theory that aims at being cognitively adequate. More precisely, our hypothesis is that «time» induces an asymmetric (total) relation among spelled-out linguistic units and this is a necessary, though not sufficient, condition for the correct structural analysis of the sentence, hence for its correct interpretation. The other crucial property is hierarchy: implicit groups of words (phrase structures) that constraint temporal expectations and create complex (recursive) meaningful units. Our intent in this paper is to argue, by analyzing simple recurrent networks expectations and machine learning approaches to automatic classification on a very simple linguistic fact like anaphoric binding, that the first property (linear order) is neither a simple articulatory-perceptual restriction (also known as a «Phonetic Form effect», Fox & Pesetsky 2005) nor a sufficient distributive cue. On the other hand, we suggest that the second property (hierarchy) should be reconciled with a specific notion of time in order to solve fruitfully most of the anaphoric binding puzzles we discuss.