Stochastic Typology and Optimality Theory

Elena Maslova
To be presented at Stanford Optimization/Variation/Change seminar, 11.09.01

Outline and reading

1. Introductory

The bottom line of "functional" cross-linguistic studies of the recent decades (disclaimer: not a received view in functional typology):
  1. There are no non-trivial absolute universals (apart from theory-internal ones), but there seem to be a lot of "strong tendencies".
  2. In particular, structural (morphosyntactic) categories (relations, functions) are language-specific (or theory-internal), but there are certain recurrent structural patterns (more or less common), which need to be compared and "explained" (in some sense) (Dryer, Croft)
There is no adequate theoretical response to these empirical challenges (so far); can (stochastic) OT provide such a response?

2. Distributional universals and transition rates

A model of "distributional" (statistical) universals is presented in (Maslova 2000). The major point is that we need to look at transition rates of type-shift processes in order to account for this type of phenomena. A possible way to do so empirically is outlined in (Maslova 2001), but this problem will not be discussed here. Both texts make use of a discrete time model of type-shift processes (since I assumed it would be easier to understand), but I actually have in mind a continuous time model. Similarly, this model assumes instantaneous type-shifts; stochastic OT might help to get rid of this oversimplification.

3. Stochastic OT as a model of typological space

Stochastic OT seems to give a straightforward model of "typological space" (= the space of all possible languages) -- constraints can be conceived of as "dimensions"; the weight assigned to a constraint, as the location of the grammar along that dimension. This approach presupposes that the weights are normalized in some sensible way (i.e. that their absolute values are given some typological sense); I will discuss some possible ways to do so. A more significant problem is that this approach imposes the following requirement on the constraint set: any language change should be conceivable as a minor modification of constraint weights (i.e. languages cannot hop between distant points). But may be this requirement is a reasonable one anyway?

In this model, an n-way typology divides the space in n domains; it appears that some distributional universals (=statistical tendencies) can be explained just in terms of various sizes of these domains: a "preferred" type just covers more points (note that the typologies for which strong tendencies have been claimed to exist are defined not in terms of OT-style constraints; rather, we are dealing with surface manifestations of certain re-rankings). This sort of entropy-based explanation is an extension of Antilla's approach to variation; from the typological perspective, it may reflect the intuition that a type is "preferred" if it gives a language more freedom of choice as far as other parameters of variation are concerned -- which would give a new sense to Greenberg's notion of "dominant" type. The good thing is that this approach, if empirically plausible, predicts transition rates as well as stationary synchronic distributions. In other words, grammatical change can be modeled in terms of some sort of "random walk" of constraints along the scale.

Another possibility is that there are some "genuine" preferences, that is, typological space may have its hills and valleys (Lass). To put it in a more scientific way, we will need some meta-evaluation for possible configurations of constraints. Since all actual descriptions of typological space are by necessity partial, it will be difficult to distinguish between entropy and meta-evaluation empirically (a valley may prove to represent a group of "lost" dimensions).

4. Problems of empirical adequacy, or whether and how each language evaluates all typological options

If the time allows, I will discuss two interrelated problems, which, in my view, may undermine typological plausibility of OT-style grammars.
  1. Grammars are supposed to evaluate analyses for which they have actually no output strings. This looks like a rather artificial way to make all constraints universal.
  2. It is assumed that each "real" output can be (uniquely) analyzed in terms of universal grammatical functions & features (cf. observation (2) above).
Is there a way to overcome these problems within OT? I will attempt to outline a positive answer...
Page maintained by Elena Maslova, Created: 11/2/01 at 10:08:18 Updated: 11/4/01 at 19:53:53