Iterated Bayesian Learning and Portuguese Clitics Qualifying Paper II

Diversity of language is a key part of our understanding of natural languages now and from the past. This diversity goes hand in hand with language change. Change is pervasive at every linguistic level. However, the space of existing languages does not appear to be unconstrained. In the modern generative tradition, this is governed by Universal Grammar (UG) (see, for example, Kroch (2000)). A number of computational models have been proposed to probe the nature of language change and role learning has within this (eg. Kirby (2001); Kirby et al. (2007); Nowak et al. (1999); Niyogi and Berwick (1998); Griffiths and Kalish (2005)). All of these studies can be grouped under the umbrella of iterated learning. Within these computational models, variation arises primarily from the individual learning mechanism and the varying input coming from the rest of the adult population. Different studies embrace different approaches to the problem. However, all of these approaches require the learner to select a grammar from a finite data set. However, the majority of computational models of language change have generally been frequentist in nature. In particular, Maximum Likelihood Estimation (MLE) is usually employed as the decision making criterion. In this context, the learner chooses the grammar which makes the input data most likely. In actuality, this amounts to assuming that grammars have specific and probabilities associated with particular syntactic constructions. Grammar acquisition is then reduced to matching observed frequencies to these grammar intrinsic probabilities. Whatever uneasiness one might have about embedding probabilities so deeply in the grammar we can immediately note another problem associated with this sort of estimation. That is, MLE requires us to conflate several factors: innate (prior) biases, social and communicative factors, and random noise. A natural candidate for solving such problems that has the ability of separating out these factors is Bayesian decision theory (c.f. Berger (1985)). That is, we set the learner the goal of choosing the grammar that will maximize their expected utility. This differs from the MLE based approaches mentioned above in that it takes a subjectivist view towards to the problem. Strong arguments have been made for the philosophical advantage of the Bayesian approach (Efron, 1986; Jaynes, 2003). The main benefit from our point of view is that the utility function explicitly takes into account issues of communicability and processing/production difficulties. This allows us to, at least theoretically, separate out such factors in a principled fashion. This paper considers these different approaches with respect to the case of clitic position change in European Portuguese. In the end, I will argue that, while MLE is the bread and butter of