October 01, 2007

Finch phrase structure?

Ten days ago, I described the recently-published PLoS paper by Gang Li et al., "Accelerated FoxP2 Evolution in Echolocating Bats", which showed that "contrary to previous reports, [the] FoxP2 [gene] is not highly conserved across all nonhuman mammals but is extremely diverse in echolocating bats" ("Wherein I take the bait", 9/21/2007). I briefly mentioned the unfortunate fashion, at its height five or six years ago, for calling FoxP2 the "grammar gene" or the "language gene" -- see Geoff Pullum's 9/5/2005 post "The Continuing Misrepresentation of FoxP2 effects" for some details, and Alec MacAndrew's 3/1/2003 essay "FoxP2 and the Evolution of Language" for more. I expressed the hope that the new research would further discourage careless speculation on this point.

But just a few days later, Seed Magazine published an article (Juan Uriagereka, The Evolution of Language, 9/25/2007) asserting that "we can be confident of the fundamental role of FOXP2 in human language".

Uriagereka doesn't mention the bat results, but he does mention (without specific references) some bold speculations about the role of FoxP2 from his earlier work with M. Piatelli-Palmarini (The immune syntax: the evolution of the language virus, inVariation and universals in biolinguistics, 2004; and "The Evolution of the Narrow Faculty of Language: The skeptical view and a Reasonable Conjecture", Lingue e Linguaggio, 2005.). The Seed Magazine article goes beyond those highly speculative proposals, to an even more extreme suggestion about the role of FoxP2 in "what linguists call parsing: the integrated processes that allow you to reconstruct complex sentences as you hear or see them, to produce them, or to acquire the fundamental parameters of your language as you first experience language as a baby":

Because of the similarities in brain structure and in the syntax of their song, finches must also have this parser. They may not be using it to process complex thoughts because they lack the cortex to generate them, but their songs are certainly intricate enough, and reconstructing their subtle structure from a linear sequence of notes is a remarkable computational task. [...]

Chimps, and our other close relatives the apes, certainly have the hardware for some basic forms of meaning, but all indications are that Neanderthals also had meaningful thoughts, enough to bury their dead or control fire, without much of a language. What they don't have is a way to externalize their thoughts. I'd wager that chimps just lack the parser that FoxP2 regulates. Somehow humans, by contrast, were able to recruit an ancient gene with a relatively ancient function to help us squeeze our thoughts out into the airwaves, much as a finch does with his. We are just thinking apes, with a finch's ability to sing.

On the question of whether there's any evidence that the FoxP2 gene regulates a "parser", I'll refer you back to MacAndrew's essay. As for the vocal abilities of Neanderthals, Steven Mithen, in his 2006 book "The Singing Neanderthals", weighs the evidence and concludes that they did not have compositional language, but did have a highly evolved form of singing, such that "all modern humans are relatively limited in their musical abilities when compared with the Neanderthals". Whether or not this theory is true, it illustrates the lack of any scientific consensus that singing is an innovation of anatomically modern humans (though it must have developed in the hominin line at some point after we split from the other apes).

But it's the stuff about finches in Uriagereka's essay that is most deeply strange. He combines some important themes with what seem to me to be some bizarre fantasies.

Now, I'm a fan of Darwin's idea that human language developed out of love songs. And I'm a great admirer of songbirds, and have put in some time working with researchers interested in the structure of their songs. I think that the intersection of distributional, physiological and genetic studies of birdsong production and perception offers one of the most exciting opportunities for scientific research today. The learning, production and perception of birdsong may well share some principles with human speech and language. But I'm baffled by the idea that finches "must have this parser", where "this parser" is identified exactly with whatever allows humans to learn, produce and perceive linguistic structures.

I'm also baffled by the statement that finches perform a "remarkable computational task" by "reconstructing [the] subtle structure" of their songs "from a linear sequence of notes"

What is the syntax of birdsong actually like? Here's a description from Kazuo Okanoya, "The Bengalese Finch: A Window on the Behavioral Neurobiology of Birdsong Syntax", Annals of the New York Academy of Sciences, Volume 1016 Behavioral Neurobiology of Birdsong Page 724-735, June 2004:

Bengalese finches have been domesticated in Japan for 240 years. Comparing their song syntax with that of their wild ancestors, we found that the domesticated strain has highly complex, conspicuous songs with finite-state syntax, while the wild ancestor sang very stereotyped linear songs.


We may divide birdsongs into two types. When one song note is followed by another song note in deterministic fashion in a single song, or the order of song notes are fixed in each song of a multirepertoire bird, such songs may be identified as "linear" song. The most widely used oscine song system models (the zebra finch, white-crowned sparrow, song sparrow, and swamp sparrow) could all be identified as having linear song syntax. When there are some variations introduced in the ordering of song notes, such a song should be called as a non-deterministic song. Species with non-deterministic song repertoires include the nightingale, starling, willow warbler, and Bengalese finch.

Among these species, Bengalese finches are unique in that their songs are characterized by finite-state syntax. Finite-state syntax refers to a simple form of syntax in which finite numbers of state are interconnected by arrows and a string of letters is produced when state transition occurs. In Bengalese finches, 2 to 5 song notes are chunked together, each of these chunks are emitted at a particular state transition, and the pattern of chunk production follows finite-state syntax.

Here are Okanoya's grammars for the songs of a Bengalese finch and a white-rumped munia (representing the wild ancestors of Bengalese finches):

FIGURE 2. Example of sonograms and transition diagrams of a white-rumped munia song (upper) and a Bengalese finch song (lower).

Okanoya Fig. 2

When Okanoya writes that "Bengalese finches are unique in that their songs are characterized by finite-state syntax", I think he must mean to be claiming that they are the only songbirds whose song grammar involves a rich repertoire of repeated sub-sequences, or something of the sort, as created by the recurrent arcs in the sample song grammar that he provides. I can testify from observation that (e.g.) the syllable sequences in zebra finch songs are well described by a finite-state grammar in which some states can be skipped or repeated. Similar finite-state grammars  can also be inferred from other animal behaviors where discrete gestures are serially ordered in well-practiced ways, such as mouse grooming (See e.g. John C. Fentress, "Emergence of pattern in the development of mammalian movement sequences", Journal of Neurobiology, 23(10): 1529-1556, 1992; A.V. Kalueff and P. Tuohimaa, "Grooming analysis algorithm for neurobehavioural stress research", Brain Research Protocols, 13(3): 151-158, 2004.) And for more on individual and dialect variation in finch tweeting, see the sequence of Language Log posts on Vinkensport.

But I'm not aware of any evidence that birdsong grammars involve long-distance dependencies, or clausal structures in which subsequences belonging to specific lexical categories can be substituted in designated slots, or any number of other characteristic properties of human syntax; nor do I know of any evidence that birds "parse" these behavioral sequences, in any way beyond the neural equivalent of transition-matrix probabilities. Perhaps they do, but simple assertion doesn't make it so.

Working out the neural architecture of motor control that underlies such patterns is one of the most exciting research topics around these days, in my opinion. And it's plausible that motor-system patterns are somehow behind linguistic structures, and not only as a matter of evolutionary history -- the "Declarative/Procedural Model" of Ullman, Pinker and others suggests that "mental grammar involves procedural memory and is rooted in the frontal cortex and basal ganglia" (see Michael Ullman, "A Neurocognitive Perspective on Language: The Declarative/Procedural Model", Nature Reviews Neuroscience, 2:717-726, 2001).

But to reify human speech and language abilities as a "parser" located in the caudate nucleus, regulated by FoxP2 and shared with finches -- well, speculation is fun, but this is like the kind of too-specific science fiction that's out of date by the time it's published, and seems merely quaint within a few years.

Posted by Mark Liberman at October 1, 2007 08:20 AM