|
|
Greg Kochanski |
|
This page is a comment on "Computational Constraints on Syntactic Processing in a Nonhuman Primate", by W. Tecumseh Fitch and Marc D. Hauser, which appeared in Science on 16 January 2004 (volume 303), pages 377-380. Our original letter to Science is here; this is a modified version, after reading Fitch and Hauser's reply to the editors.
Fitch and Hauser [1] have shown that undergraduate students can learn patterns of syllable sequences of the form AABB, or AAABBB, whereas tamarin monkeys can not. Specifically, the humans could reliably decide whether auditory stimuli fit those patterns or not. The authors and Premack [2] interpret this to mean that humans can learn Phrase Structure Grammars (PSG), whereas the monkeys cannot; such grammars are claimed to be crucial to human language [3]. However, their interpretation suffers from a flaw: although the the patterns used were generated by a PSG, nothing in the experiment constrains how the subjects process the sequences. Specifically, the induction of a PSG is not required to learn them, and the experiment does not establish that the human subjects use any particular grammar when they decide which sequences fit the pattern. Fitch and Hauser have done nothing to investigate whether their human subjects actually used a centre-embedding parse to perform the task, so it would be premature to argue that the ability to execute a centre-embedding parse is the reason behind the difference in monkey and human language abilities. A crucial unanswered question is whether or not the humans parsed the stimuli like they would parse the sentences of a complex PSG.
We believe that the central flaw of the paper is that the authors assume the human or monkey who is solving the problem will use the same technique ( i.e. a PSG) as the experimenters who constructed the sequences. Without that assumption, describing the AABB, AAABBB patterns as examples of a phrase-structured grammar makes little sense. Yes, such patterns can indeed be made by a PSG, but they could also be made by any number of other algorithms. Equally true, the experimental subjects could be using mental processes that bear little or no relationship to PSGs.
If one were attempting to force a recursive parse of these strings, one might try to link the first A to the last B and so forth. For instance, if the strings were of the form A A A B B B it would be plausible to assume that a subject who is matching the pattern would need to keep a stack of the colours of the A's so that they could make sure that each B was of the appropriate color. However, that was not part of the experiment.
One can force a stack-like (recursive) interpretation of an AAABBB pattern by making the associations A="the [noun]" and B=[verb], then constructing a sentence like "The cats the dog the men walk chases run away." One then finds that the parsing of such a centre-embedded sentences is a non-trivial task [4] [5] [6] [7], close to (or sometimes beyond) the limits of human performance, especially so if one includes a time constraint. Untangling such a sentence is much harder if you only have a second or two before the next sentence comes along, as you would in a normal conversation or while reading. The fact that people could match Fitch and Hauser's AAABBB patterns so much more easily than they can parse the above sentence where semantic and syntactic constraints force a recursive parse suggests that the people in Fitch and Hauser's experiments were not doing a recursive parse.
After all, one might expect that "The cats the dog the men walk chases run away" might be even easier to recursively parse than an abstract pattern of nonsense sounds: our semantic knowledge should assist us. Men commonly walk; dogs commonly chase cats; cats commonly run from dogs. What could be more natural? Likewise, our syntactic knowledge might help us, too: men and walk are plural; the dog and chases are singular; and the cats and run away are plural. One might expect syntax and semantics to help connect each verb with the appropriate noun. If the abstract pattern proves to actually be easier to parse, perhaps it is because the subjects are using some simpler technique that does not involve recursion and the storage of so much information in short-term memory?
We note that Fitch and Hauser also gave their humans AAAABBBB patterns (this is in an on-line supplement to the paper, and their reply to our comment); the equivalent centre-embedded sentences are effectively impossible for humans to parse without pencil and paper. The fact that their patterns were so much easier than the equivalent sentences in actual language raises serious doubts that the experiment says anything about language.
We can think of another, simpler explanation for the observed difference between human and tamarin performance. Perhaps the human subjects determine the "grammaticality" of the stimuli by counting the number of syllables of type A and type B and checking that the numbers match. (We are not aware of any experimental evidence that counting to three requires anything approximating a PSG in the brain.) By counting, humans can judge the grammaticality of these recursively generated stimuli without a clear need for recursive mental structures.
While there is evidence that tamarins can compare numbers up to and including three, this experiment is a fairly complex task, with the added cognitive load of classifying stimuli into A and B classes. Tamarin number comparison could fail for a variety of reasons unrelated to PSGs. For instance, it is possible that the monkeys simply did not realize that the task could be solved by counting and comparison.
We note that, in the absence of an understanding of how phrase structured grammars could arise from interactions of nerve cells in the brain, the burden of proof is on Fitch and Hauser. It is their responsibility to show that no other simpler mechanisms are possible.
The existence of another interpretation of Fitch and Hauser's experiment shows that the case for recursive Phrase Structure Grammars as the crucial difference between humans and tamarin monkeys has not yet been made.
One could raise the objection that humans can parse PSGs, so that if Fitch and Hauser have established that monkeys can't parse PSGs, then they have indeed shown a crucial difference between the species.
Such an argument is imprecise. While humans can parse some sentences that are most elegantly described by a PSG, it is not an unrestricted ability. There are many sentences derived from PSGs that are unparseable, in the sense that no unaided human could decide on their grammaticality within any reasonable amount of time.
Centre-embeded sentences (the linguistic analog of Fitch and Hauser's patterns) are a good example. People can handle single centre-embeddings fairly easily. For instance "The man the dog growled at backed off" is not too hard to understand.
Double centre-embeddings are more difficult. A well crafted sentence in that form can be understandable, but such sentences are quite rare in most corpora. Certainly, some double centre-embedded sentences are better described as "puzzles" than means of communication. Triple centre-embeddings are (perhaps with a few exceptions) beyond human ability to generate and understand.
Consequently,since humans cannot parse every pattern derived from a PSG, Fitch and Hauser's identification of a pattern that monkeys cannot parse that is derived from a PSG becomes unsurprising. It's just another difference between monkeys and humans. Probably, it is more scientifically interesting than the fact that monkeys cannot do calculus, but there is nothing that picks it out as the single difference that explains everything, as the authors would like to claim.
Another comment on the same paper was written by Mark Liberman.
| [ Papers | kochanski.org | Phonetics Lab | Oxford ] | Last Modified Thu Apr 7 16:42:26 2005 | Greg Kochanski: [ Mail | Home ] |