In an important article, Fei Xu and Joshua Tenenbaum propose a Bayesian theory of word sense acquisition (Xu & Tenenbaum, 2007). Their model assumes that the child has a vast space of hypotheses about possible word referents. Each hypothesis consists of a subset of objects to which a word can refer (e.g. " all living things ", " all dogs ", " all Dalmatians ", etc.). Each time the child hears a word in a given context, it updates the probability that each hypothesis is true, following Bayesian rules. Finally, a crucial assumption of the model is that likelihood varies as an inverse function of the size of the hypothesis considered.
Based on these axioms, the authors show that it is possible to account for a series of important empirical observations. The meaning of a word can be learned from a single example, or from a few. Positive examples suffice : the child needs no counter-examples. They can acquire a set of words for overlapping concepts. Inferences about the meaning of a word are gradual, with varying degrees of confidence. Finally, these inferences can be influenced by the learning context, particularly the speaker's attention, knowledge and intentions.
In particular, the Bayesian model accounts, without additional hypothesis, for a classical linguistic principle, the principle of exclusivity : each entity has only one name. From the age of sixteen months, in fact, when they hear a new name, children postulate that it refers to an object whose name they do not already know. This property simply follows from a hierarchical Bayesian model, assuming the child is capable of conversational inferences such as " if my interlocutor had wanted to name object X, he would have used word X ". Thus, lexical acquisition may require nothing more than a generic statistical inference algorithm that could be present in other animal species. In fact, the learning of several hundred words, respecting the principle of exclusivity, has been documented in a domestic dog.
