Language Module

Yesterday’s Speech Research Laboratory presentation was by MIT’s Adam Albright, under the title Natural classes are not enough: Biased generalization in novel onset clusters. The link goes to a PDF of the paper on which the talk was based.

I’m not a speech researcher; I’m not really even interested in Phonology, let alone things like Phonetics and Auditory Perception. But I went to this one because it bears on a debate that everyone at IU is interested in by default - given the level of sheer idiocy with which the topic is often addressed here: the question of the existential status of symbol processing systems.

Basically, the problem is this. Linguistics has traditionally framed its explanations in terms of symbol processing systems. Which is to say, the fact that The books I shelved without reading were old is grammatical and I shelved the old books without reading them is also grammatical but *I shelved the old books without reading is, perhaps unexpectedly (given the first sentence), not, is usually explained with reference to absolute rules that operate over classes of words, rather than the words themselves. And the fact that while neither blick nor bnick nor bzick are actual words of English, native English speakers seem to feel that the first definitely could be, the second maaaaybe, but no way for the third, tends to be explained by reference to constraints that refer to classes of sounds, rather than to generated soundwaves directly. In the first case, the issue is that “reading” needs an object, and in the first sentence there is reason to believe that “books” is that object and has simply “moved” to a different position in the sentence than the one it “started out in,” where “them” is the object in the second, but in the final sentence the trouble is that there is a ban on “books” having moved from the object position and so “reading” has no object at all. This explanation does not rely on anything like the sound sequence of “books,” the frequency with which it is used by English speakers, at what volume the speaker is speaking, how quickly he speaks, what other sentences he has used leading up to this one, his relative state of health and nutrition when making the utterance. Rather, the hypothesis is that he has knowledge of a highly abstract set of grammar rules that determine which sentences are members of his language and which are not, and that he applies these like a computer to determine what makes sense in his language and what doesn’t. Likewise for the three nonwords. The traditional explanation isn’t too concerned with how many other words in English start with bl bn or bz sound combinations. Rather, it posits that sounds are members of feature-defined abstract classes that either do or don’t interact well with each other.

There is plenty of reason to think that the traditional symbol-based explanations are missing some important pieces of the puzzle. And in fact, as far as I can tell, virtually everyone does believe that the traditional symbol-based approaches are missing important pieces of the puzzle. Including, in fact, the people like Noam Chomsky and Morris Halle who were largely responsible for steering Linguistics toward the symbolic approach in the first place. Of course, a lot hinges on what you mean by “pieces,” and how big “the puzzle” is supposed to be.

Someone could easily write a multi-volume discussion on what the proper object of study of any field - Linguistics included - actually is, of course, which is why it always seems silly to me to get too bogged down in the debate about whether or not it is appropriate for Linguists to be framing their explanations as symbol systems or not. Just as it becomes difficult to talk about a defendant’s guilt or innocence of a particular crime if you have to first say what “the meaning of is is,” it’s hard to get any real work done when you’re constantly fighting with your colleagues over what “Linguistics” means.

So my personal solution has always been for peaceful cohabitation. More often than not, these “disputes” about whether symbols are allowed or not are beside the point, an artefact of the fact that some people study phenomena that operate at a higher level of abstraction than the phenomena that other people study. If you’re only interested in the fact that blick is acceptable while bzick is not, then a symbol-based statement of the generalization that accounts for it will do just fine. If you start to get interested in the fact that bnick, while unacceptable, is still somehow better than bzick, then while you don’t necessarily have to abandon symbols, you might do better filtering symbol sequences with constraints rather than banning them with rules. And if you’re interested in whether there is anything about the vocal tract or the configuration of the brain that accounts for any of this stuff, then the rules and constraints will be a good pointer to the kinds of tests you need to devise, but ultimately your explanation can’t have anything to do with symbols or rules. In other words, symbols and symbol systems are and always have been theoretical abstractions. The same way that Python mediates between the human user and the bit-and-byte operation of the computer, symbol systems allow linguists to talk about the generalizations they’re interested in without having to draw sound waves or pictures of neuron clusters. Phonologists are not interested in how the constraints of Optimality Theory are actually instantiated in the brain as that is a job for Psychologists (and, perhaps eventually, Neuroscientists).

The question of whether there is a grammatical “backbone” instantiated in the brain anywhere at all is more pertinent to Phonology than it is to Syntax. In the case of Syntax, it approaches the undeniable that there is such a system. Put differently, there is a very clean break between the algorithm and how that algorithm is actually “compiled into neurons.” Syntacticians study the algorithm. With Phonology, the field seems to get more and more suspicious that there could be anything like a “sound grammar” that can be studied apart from the soundwaves and articulators. But I think this in itself is revealing. Sound just so happens to be the kind of thing that is pretty amenable to quantitative study. It is comparatively easy to get people to agree to phonetic transcriptions of words in a language, such that reliable corpora can be built. There are spectrograms and such for measuring actual sound waves. And intuitions about whether strings of sounds are acceptable are generally uncomplicated by things like “on the present interpretation” and so on. This isn’t to say that syntax can’t be quantitatively studied, of course, just that it’s comparatively harder to do so - which is, if you think about it, something like equivalent to saying that abstractions are more appropriate to Syntax.

In any case, my impression of a lot of the attacks on formal Phonology of the kind that Bob Port and Adam Leary are fond of leveling is that to the extent that they invite discussion it is because (a) they are at least mildly successful at enticing people to subscribe to a straw man view of what “formal phonology” is, and/or (b) because there are reasonable inroads available into mapping the connections between the abstract level of formal Phonology and the concrete performance acts of speech perception and production, and the availability of these inroads (ironically) make it easy to blur the distinction between the two. The result is that we get drawn into a lot of misleading debate about whether Phonological models are valid, most of which involves simple haggling over the semantics of things like “formal” and “Linguistics” and “symbol” and “discrete.” What we get comparatively little of are demonstrations of how it is that Phonological categories are so persuasive if not real and what it is that accounts for the strange success of abstract Phonology if it is, in fact, an illusion.

Fortunately, there are some researchers out there (and of course Noah Silbert is about to grow up to be one of them) who aren’t so easily sidetracked. Which brings me back to Adam Albright. His talk yesterday was satisfying for precisely the reason that it had illuminating things to say about the mapping between performance and competence in Phonology without concluding that there was no Phonology. I’ll leave the details for the paper (linked above) - but the gist of it is that there are gradient judgements about the phonological acceptability of nonce words that cannot be easily accounted for by the usual appeals to statistical distribution. To the extent that you make an appeal to statistical distribution, it has to be filtered through a feature-based generalization over the data present. Finally, it turns out that there are some preferences that seem to defy the kinds of natural class generalizations that one might get automatically out of a corpus study of a particular language - but of course these are the kinds of things that Optimality Theory is tasked with noticing and writing down - i.e. “phonetic” constraints that are presumably true across all languages.

David Pisoni and Bob Port spent most of the talk smirking at each other like junior high school girls having “eye conversations.” What neither of them did was let any of the rest of us in on the joke by, say, asking Albright’s opinion of the apparently obvious alternative explanation for the data that they were privy to and the rest of us missed. What Pisoni did do at the end was make a funny joke about being on a “mission - like in the Blues Brothers” - but that in itself is a revealing formulation. In other words, Pisoni has made up his mind a priori to any actual evidence that only certain kinds of explanations are acceptable, and he will do the research it takes to convince everyone he’s right. Which is fair enough, really. I think we’re all aware that researchers start with personal biases - prejudices, really, in favor of some interpretation or the other of how the world works. Provided they do honest research to bolster their claims which they submit to due scrutiny by the community, the rest of us can at least try to evaluate the claims objectively, which is the name of the game in science. In fact, one of the best formulations of this idea that I’ve heard comes from an online criticism of Bob Port. Responding to this statement from (Port and Leary 2005):

There is only one route left to justify doing traditional generative phonology or for studying only the abstract sound structures of a language and deny the relevance of articulatory, acoustic and auditory details. It is to claim: We don’t care about linguistic behavior, only about linguistic knowledge. But there is no assurance that a coherent static description of knowledge exists just because that is what one wants to study.

Oostendorp writes the following:

For me the last sentence I quoted is a very important one. Of course there is no guarantee that we will be able to understand X just because we want to. That seems to me inherent in the nature of doing research — or of human existence, and it can hardly be a reason to give up.

Quite right. What is interesting is that Bob Port doesn’t seem to have any problem with the David Pisonis of the world taking exactly this attitude. It is only when people like van Oostendorp do it that it’s “wrong.” It is because Pisoni followed his hunch and did the grunt work that a lot of the “evidence” that makes Bob Port’s papers possible even exists. Perhaps that gives him some license to smirk like a schoolgirl rather than argue at talks that challenge his theories - I don’t know. But surely it’s preferable on both sides to simply amass evidence and discuss it dispassionately.

In any case, there can be no room for doubt that some level of abstraction is appropriate to any scientific inquiry. Science isn’t simply about “measuring things.” I found Albright’s talk satisfying because it presented the reality of phonological generalizations in terms that people like Pisoni and Port should have been able to appreciate, and should not have found objectionable. Indeed, it was a very useful study for precisely the reason that it gives these two camps a kind of level playing field where they can present their findings to each other in terms the other understands, accepts, and can challenge. In other words, it created a kind of framework for peaceful cohabitation.

Unfortunately, at least one of the two camps doesn’t seem interested.

A closing quote from van Oostendorp - just because I like it:

There is no ‘utter lack of evidence’ for the assumptions on which formal analysis of phonologies are based; there is plenty. Maybe we are not going to find it in the phonetics. But then, if we only would take phonological facts as evidence, there would not be a lot of evidence for many phonetic details: it would be a bit funny to conclude from this that these phonological facts put an unbearable empirical burden on phonetics.

Right.

In any case, Albright’s paper is to be recommended because it understands that one takes Phonology as far as it will go and no further. To the extent that it gives good explanations for things, it is useful. To the extent that findings fall outside its assumptions, appeals to other mechanisms (say, Phonetics) will have to be made. I don’t see what’s so hard to understand about that, or why it should generate controversy.