Bayesian interpretation and Optimality Theory

The paper explores the consequences of reinterpreting OT pragmatics of Zeevat (2009) as a Bayesian account of natural language interpretation, not unlike Bayesian accounts of vision. In such accounts, a model of the most probable interpretations in the context is combined with a model of NL production to give the most probable interpretation of a given form. It is argued that pragmatics can be equated with the model of probability maximisation of interpretations while “grammar” can be equated with the human capacity of mapping thoughts to utterances or any theoretical model of that capacity. The Bayesian model by itself does not give communicative success, it is merely a better model for estimating the most probable interpretation. It is essential that the speaker also estimates the most probable interpretation of her utterance, to see if the hearer will get her right. This allows alternative formulations with an increased probability of being understood as intended. One claim of this paper is that this self-monitoring is partially automatised and accounts for such phenomena as particle insertion and word order freezing. The simplest brain architecture is as two associative processes, one leading from forms to interpretations, the other from intentions to forms that can inhibit each other. A form that in the interpretation process does not assign the strongest activation to the speaker intention is inhibited, an interpretation that does not most strongly activate the form is inhibited. This dual inhibition model assigns a natural temporal structure to the development of linguistic skills. The production skills must be good before they can contribute to communicative success in their role of inhibiting interpretation. Similarly, the interpretational skills must be well developed before they can be useful role in production.