Carl Zimmer has a nice article in today's Science Times under the headline "Starlings' Listening Skills May Shed Light on Language Evolution", explaining clearly what the Gentner et al. experiments actually were, and what they might mean. Anyone who is even moderately interested in this topic ought to read his discussion.
For those who want to go into some of the issues at greater length, I'll reproduce here a recent email from Mark Seidenberg, who wrote that:
I am so much reminded of the older ape language research, in which it was possible to claim that chimpanzees could "name objects" if the task was defined in a restrictive enough way.
In reference to an earlier LL post on pattern-learning experiments on human infants and monkeys, which quoted from a letter to Science that Mark co-wrote with Jeff Elman, Mark observed:
I hope your earlier commentary helped clarify for people that there isn't much at stake if one is talking about rules and statistics in broad terms. We tried to make some similar points in the Seidenberg, MacDonald and Saffran commentary on a paper by Pena et al. that attempted to establish limits on "statistical learning". [Seidenberg, M.S., MacDonald, M.C., and Saffran, J.R. (2003). Are there limits to statistical learning? Science 300, 53-54. A subsequent exchange of letters with Gary Marcus and Iris Berent is here.]
Those references underline the important point that recent articles about monkey and starling pattern-learning have roots in earlier studies of human infants. To some extent this excuses the tendency towards overinterpretation -- similar tendencies can be seen in the earlier infant literature. On the other hand, most of the issues now being debated with respect to the performance of starlings and monkeys have already been debated with respect to the performance of human babies; so maybe there's that much less reason for unclarity now.
In any case, whether the research subjects are humans, monkeys or birds, I'd like to see less focus on over-simplified all-or-none hypotheses like "species X has (or doesn't have) recursion", and more focus on understanding what biological pattern processing really is and how it really works, in general and also the (perhaps different) specific cases of particular species dealing with particular kinds of patterns for particular purposes. This is a harder and less superficially glamorous program, but it's more likely to lead to durable progress, in my opinion.
Mark continues
I don't agree that rules and statistics are indistinguishable under all definitions of the terms. They are just indistinguishable in the vague way people have used the terms in certain contexts (e.g., the Marcus study to which we responded).
There are people who have tried to assign specific properties to rules, which differentiate them from "statistical" procedures. Certainly Pinker and Marcus come to mind. Rules are thought to be learned by different procedures than "statistics," represented in a different part of the brain, on a different developmental time course, etc. I don't think these claims turn out to be accurate, but they tried to identify unique properties of rules. Other people have also tried to do this; see Smith Langston & Nisbett (1992, Cognitive Science). I think the attempt is valid but the properties they assign to rules (e.g., not being sensitive to frequency or similarity) don't apply to people's behavior except under a very severe idealization that ignores this information (under "colorless green ideas" reasoning).
The term "statistics" on the other hand, as in "learning statistical patterns" has been completely mangled in the psychology literature. The influence of the original Saffran et al. study has been so great that researchers in child language have equated "statistics of the input" with "transition probabilities" between syllables, which is what they manipulated in the original study. So, to disprove the "statistical learning" hypothesis, you show that a species (humans, whomever) is able to learn patterns for which these particular statistics are uninformative.
The problem is that statistics always have to be over something, and things like syllables are not a priori but rather may themselves arise from statistical learning procedures. Statistics all the way down perhaps. The confusion about this is deep; there have been several exchanges in journals like Science, and others that could have occurred except that I decided to stop responding to every case. But, see recent papers from Jacques Mehler's group on this (e.g., Bonatti et al., Psychological Science, 2005, 451, and Pena et al., Science 2002, 298, 604, to which we did reply). I hope your postings on these issues are widely read; I am assigning them this week in my seminar.
For the convenience of readers who still haven't had enough, here are links to our recent starlings coverage:
The race to the bottom in science reporting (4/26/2006)
Starlings (4/27/2006)
Starlings linguists language loggers readers follow commented on the work of studied are damn smart! (4/28/2006)
Separating species with bullets (4/28/2006)
Wild? I was livid! (4/29/2006)
Can you speak in rhinoceros? (4/29/2006)
And links to earlier Language Log posts on related topics:
Language in Humans and Monkeys (01/16/2004)
Hi Lo, Hi Lo, it's off to formal language theory we go (1/17/2004)
Cotton-top tamarins: on the road to phonology as well as syntax? (02/09/2004)
Humans context-free, monkeys finite-state? Apparently not. (8/31/2004)
Homo hemingwayensis (01/09/2005)
JP versus FHC+CHF versus PJ versus HCF(08/25/2005)
Rhyme schemes, texture discrimination and monkey syntax (02/09/2006)
Learnable and unlearnable patterns -- of what? (02/25/2006) ]
[Carl Zimmer is (Language Log contributor) Ben Zimmer's brother]
Posted by Mark Liberman at May 2, 2006 01:27 PM