January 10, 2005

? taR .rM, siht si egaugnal tahW

Various news organizations, like Al Jazeera and CNN, are running with a great new animal language story. You might prefer to get it from the horses mouth - full text of the Journal of Experimental Psychology article.

It turns out that Spanish rats can be trained to prefer synthesized pseudo-Japanese over synthesized pseudo-Dutch, or vice versa, more easily than Spanish rats can be trained to prefer backwards synthesized  pseudo-Japanese over backwards synthesized pseudo-Dutch, or vice versa. I'm not sure whether the rats actually consider themselves Spanish, as they reside in Barcelona. But they're hardly Catalan, as they come from a strain of rattus norwegicus which was originally cross-bred here in California. But as usual, I digress. Digression is so much easier than in the old days. How on earth did people manage to digress effectively before Google?

Anyhow, you can guess why the press is excited. A language log favorite. Talking animals. Indeed, the authors of the paper put their result in roughly the talking animal category, albeit in a much more finely nuanced way: animals are surprisingly well attuned to prosodic properties of language as against other physically similar stimuli. The researchers, Juan M. Toro, Josep B. Trobalon, and Nuria Sebastian-Galles, are sensible people, and do not take a Dolittlian inter-species communication or new age conclusion from this. Rather, they think it is evidence that in the development of human language, features already present in the mammalian auditory system were co-opted.

I am intrigued by the study, and I have the impression it was carried out carefully and effectively. But personally, I never had any doubt whatsoever that in the development of human language, features already present in the mammalian auditory system were co-opted. Moreover, I'm skeptical that Toro et al's study shows this. The problem is that Toro et al don't actually know which features of Japanese and Dutch were the ones that mattered, the relevant differences between the two languages that are more easily extracted forwards than backwards. And this leads me to some very general mathematical questions:

  1. Are there types of pattern recognizer such that those recognizers can differentiate between certain classes of pattern they are presented with in one order, but not differentiate between those classes of pattern when presented in the reverse order?
  2. Are there types of formal language recognizer (e.g. with a limited working memory, like just two states, or a limited stack, whatever) that can recognize classes of languages in one direction, but cannot recognize classes of language consisting of the same strings except in reverse?
  3. Are there types of learning algorithm such that these algorithms can learn to recognize certain classes of pattern presented in one direction but not learn to recognize the same classes of pattern presented in in the other direction?
  4. Are there types of  learning algorithm such that these classes can learn to recognize certain formal languages but not learn to recognize the languages which consist of the same set of strings except in reverse?

I'll put money on the answers being 4 * yes with even a quite modest definition of what a "type of pattern recognizer" etc is. In which case, the difference between Japanese and Dutch prosody might be intrinsically more learnable for a large class of abstract learning systems than is the difference between the reverse of these langagues. And this class of learning systems might well include every organic learning system that has ever muddied its feet, scales or other protuberances on our wonderful planet, not just mammals, and not just animals with auditory systems. In which case, all the results of Toro et al's study would show is that Dutch and Japanese evolved in such a way that they are potentially recognizable and learnable by some creature, whereas Hctud and Esenapaj did not.

Would this surprise us?

Posted by David Beaver at January 10, 2005 02:24 AM