Starlings and Recursion

UPDATE: The actual linguists at Language Log have posted on the Gentner et al. paper, here and here, and they're less impressed with the results than I am. Since they're linguists, you should probably just read their post instead of this one.

UPDATE II: A really good discussion (much better than mine, and no, I'm not just being modest), and critical evaluation of the experiments can be found here. The one thing I wonder is whether the explanation there (simple counting) would account for the generalization to A³B³ sequences. It might, but I'd have to think about it a bit more.

A few years ago, Hauser, Chomsky, and Fitch wrote a paper in which they argued that the "only uniquely human component of the language faculty" is recursion, and on top of that, recursion is the only component of the language faculty that is not co-opted from other perceptual systems (though they believe that even recursion evolved for other purposes). It turns out that recursion recursion may not be uniquely human after all. A paper by Gentner et al. published in yesterday's issue of Nature shows that, in the authors' words:

European starlings (Sturnus vulgaris) accurately recognize acoustic patterns defined by a recursive, self-embedding, context-free grammar.

What does that mean? I'll try to explain. Since I'm not a linguist, I'll probably get something wrong. Feel free to point out my errors.

In Aspects of the Theory of Syntax, Chomsky specifies a hierarchy of grammars for formal languages, with those at the top being the most inclusive grammar, called Type-0, or unrestricted grammars, and at the bottom, the least inclusive, called Type-3 or finite-state grammars, which are grammars that can be decided by a finite-state machine. In essence, a finite-state grammar includes rules that allow you to add elements (words, morphemes, phrases, or whatever) either at the beginning of a string or at the end. For example, you could start with the word "Chris," and at each step add a new word after the last word in the sequence, perhaps based on a rule specifying which word you should choose based on the last word, and come up with a sentence, like say, "Chris should stick to psychology."

Two steps up from finite-state grammars in the hierarchy is the Type-2 grammar, or the context-free grammar (the next level up is context-sensitive, at Type-1). Context-free grammars can have rules that add elements (again, these can be any linguistic element) at the beginning and end of strings, but also in the middle of strings. So, a context-free grammar would allow you to produce the sentence, "Chris should stick to psychology," and then add, in the middle of the sentence, "Chris really should stick to psychology."

One of the important arguments in Aspects is that the grammars of natural languages are not finite-state grammars. In order to model natural languages, you need at least a Type-2, context-free grammar. It's upon that argument (and the work done based on it between 1956, when Aspects was published, and 2002) that Hauser, Chomsky, and Fitch are building. The type of recursion they're talking about is the type that is possible in context-free grammars. They are arguing that the only uniquely human component of the language faculty is the recursion of context-free grammars.

If that description doesn't make sense, then check out this graphic presentation from the Gentner et al. paper (Figure 1):


Gentner et al. used two elements of starlings' songs, rattles and warbles (here's a recording of a starling song, which also contain whistles not used in the stimuli), and produced two grammars, one finite-state grammar defined as ABⁿ in the figure above, and the other a context-free grammar, defined as AⁿBⁿ in the figure (an example sequence might be rattle-warble-rattle-warble). They then used operant conditioning to train half the birds to respond to the A²B² (e.g., rattle-rattle-warble-warble ) sequences, and half to the AB² (e.g., rattle-warble-rattle-warble) sequences. If the starlings could tell the difference between the two grammars, and thus can accurately respond to one or the other, it would mean that humans are not unique in this ability. And though it took the birds bit longer (about 3000 trials) than it usually takes for starlings to be trained to classify songs, nine out of the eleven birds they trained did learn to classify the two types of grammars. Furthermore, all nine of the birds who learned to classify the two types of grammars were also able to classify A³B³ (e.g., rattle-rattle-rattle-warble-warble-warble) and AB³ (e.g., rattle-warble-rattle-warble-rattle-warble) sequences, though not A⁴B⁴ or AB⁴. This implies that the birds who had learned to recognize the grammars were able to generalize that learning, and recognize longer strings. So there were memory limitations (most humans can recognize strings with more than 6 elements).

What does this mean for Hauser, Chomsky, and Fitch's claims? Well, here's what Gentner et al. concluded:

Although uniquely human syntactic processing capabilities, if any, may reflect more complex context-free grammars or higher levels in the Chomsky grammatical hierarchy, it may prove more useful to consider species differences as quantitative rather than qualitative distinctions in cognitive mechanisms. Such mechanisms (for example, memory capacity) need not map precisely onto strict formal grammars and automata theories. There might be no single property or processing capacity that marks the many ways in which the complexity and detail of human language differs from non-human communication systems.

In other words, the Hauser et al. claim that the recursive component is unique to the human language faculty needs to be qualified. The recursive component itself is not unique to humans, though the complexity of that component may be. This may mean that the language faculty (in Hauser et al.'s narrow sense) involves more than just recursion.

While the authors don't address it, I think their research might also say something about one of the other major claims of Hauser et al. It's unlikely that starlings use context-free recursion in their everyday singing and song recognition, as evidenced by the fact that it took them longer than usual to learn to recognize the context-free grammar sequences. So it's also unlikely that starlings evolved the ability to recognize recursion for the purposes of singing/song-recognition. Instead, the ability probably evolved for some other reason. Hauser et al. suggest that it may have evolved in humans for faculties other than language as well, and give as possible examples number computation, navigation, and social relations. I don't really know anything about European starling social relations, or their ability to compute numbers, but navigation certainly seems like a possibility for the evolution of recursion in birds. While it's not direct evidence, the fact that recursion may have evolved in birds for reasons other than language production and comprehension means that, even if it evolved separately in humans (and it likely did, as Fitch and Hauser found no evidence that cotton-top tamarins could recognize context-free grammar sequences), it could have evolved for use by faculties other than language in our evolutionary history too.