Wednesday, September 12, 2007

Steven Abney paper for 19 September

http://www.vinartus.net/spa/95c.pdf

Also in the locker here:
http://www.ling.umd.edu/locker/ComputationsOfLanguage/Abney95c.pdf

2 comments:

sarah a. goodman said...

Abney's illustration of ambiguity alleges extremely idiolectical readings of two run-of-the-mill constructions. Does this not illustrate the problem with considering, as Abney claims, one adult speaker an entire speech community? Put another way, has the Generative enterprise lost touch with language and grammar, containing itself instead to arcane chestnuts of so-called I-languages? Are I-languages even languages, or are they instead purely psychology? If we really want to know what English is, for example, shouldn't we as linguists look for more regularity than is offered by one's own grammaticality judgements? I know the answer to that question smacks of descriptive linguistics, but might not we use that as a precursor to describing internal grammatical rules, establishing some sort of regularity prior to delving into all things mental?

Abney further mentions the meat and bolts of the so-called competence versus performance distinction. In my syntactic experience, performance issues haven't really come up. So how much are we really sweeping under the rug with that device? If someone stutters, fine. Or if they repeat a word in hesitation or they change verb forms in mid-sentence, deciding to express something else. But, really, how often that come up when contemplating grammatical structure? Or when asking a native speaker if an expression is valid? If you're the one who's composing the sentence under examination, then you might say, 'excuse my pronounciation, my tongue's like a meat cleaver' or something. Performance, as I've mentioned it, might be worthwhile when describing a language, but, and some would say unfortunately, that's not what the Generative enterprise really does (see above). A field linguist in starving Zimbabwe wouldn't want to elicit a sentence with a stutter and mistake that for partial or whole word reduplication, hence forth putting it into a grammar for use at the UN, let's say. But for the Generativists? I can't see that performance really factors in much at all.

Tim Hunter said...

I think Abney makes a good case that there exist "graded", as opposed to strictly yes/no, properties of (PF,LF) pairs: there is some graded notion of accessibility or perceptibility, because when we are presented with the PF "John saw Mary", we find one of the possible ("John saw Mary",LF) pairs easier to access than the other possible one, and there are undoubtedly finer-grained distinctions than ungrammatical vs. grammatical. I don't think there's any denying that *there exist* graded properties of (PF,LF) pairs (I wonder how it could even be otherwise), but these graded properties don't see to correlate with each other, so I don't know that we really want to say that the linguistic status of a (PF,LF) pair includes some particular probability or score.

For example, in his examples like "John saw Mary" and "the a are of I", the pairings of these PFs with the unnoticed meanings would presumably need to have a low score, according to his idea of "specifying a constant C such that only structures whose cost is within distance C of the best structure are predicted to be perceived". But as he notes himself, once the alternative interpretations of these strings is pointed out, it's clear to us that these interpretations are also completely acceptable, so to model the other gradable property of "degrees of grammaticality" these (PF,LF) pairs would need to have a high score (probably whatever represents perfect, maybe 1). On the other hand, take strings like "Who do you wonder whether John likes?" or "We sleeps". The intended meanings of these are, I think, barely any harder than "perfect" sentences to perceive, if at all, so on the perceptibility/accessibility scale these (PF,LF) pairs need to have a very high score; but on the "degree of grammaticality" scale, they would need to have a pretty significantly degraded score.

All that seems to suggest to me that it's going to be hard to find a grammar which assigns a single number to each derived (PF,LF) pair, such that that number has any intrinsic connection to the (PF,LF) pair ... by which I mean, such that that number is relevant in most or all of the places that we need to say something about the (PF,LF) pair. For the question of how some (PF,LF) pairs seem to have higher "accessibility/perceptibility scores" than others, it seems quite possible that these numbers come from analog properties of the discourse context, because these observations of variable accessibility concern the outcome of computing what I like to think of as the function F: PF x D -> LF.

Turning to the question of how some (PF,LF) pairs seem to have a higher "degree of grammaticality" than others, it might seem like these graded scores make a more plausible candidate for an inherent or intrinsic property of a (PF,LF) pair, but I think this is over-simplifying. Sometimes two imperfect (PF,LF) pairs seem to have differing imperfect degrees of grammaticality, for example ECP violations compared with subjacency violations, but often they don't. For example, I don't have any intuitions about whether "Who do you wonder whether John likes?" (paired with the obvious meaning) sounds better or worse than "We sleeps" (paired with the obvious meaning). I know that that particular island violation sounds better than many others, but I don't know where "We sleeps" should go ... in between strong and weak islands somewhere, or worse than strong islands, or better than weak islands? This data seems to indicate that there might only be a partial ordering of degrees of grammaticality, whereas using numbers between 0 and 1 (or any other range) would impose a seemingly-unjustified total ordering.

As Norbert and Howard were saying towards the end of last week's class, I don't think anything syntacticians do is particularly inconsistent with this notion of degrees of grammaticality. In particular, they seem to use the partial ordering where one (PF,LF) pair sounds better than another iff it violates a proper subset of the constraints that the other violates. The classic picture where a (PF,LF) pair is either in or out adds the simplification that any (PF,LF) pair which violates *any* constraints at all is "out", whereas those which violate no constraints (and sound better than or equal to everything else) are "in". In general this works quite well, because it lets us say that "Who do you wonder whether John likes?" and "We sleeps" are both just "out", in the sense that they both violate something and we can't really compare them because they're just violations of different types.

On the other hand, if we compare "Who do you wonder whether John likes?" with "Who do you wonders whether John likes?", we do get one sounding better than the other, because the first one violates a proper subset of the constraints that the second one violates. And use of this kind of reasoning is very common, I think: often, when a syntactician presents data, he/she will present pairs of sentences where one (call it sentence A) sounds better than the other (call it sentence B), and then use this as evidence for a constraint which sentence B violates but sentence A doesn't, because A and B form a minimal pair. It doesn't matter "how good" on any global scale A or B sounds, and in particular it doesn't even matter whether A is perfect or not: just that A sounds better than B. Often A is somehow degraded, and the syntactician doesn't know exactly why, just that anything that A violates B must also violate, and that B seems to be worse, which implies that B must violate something extra that A doesn't. In this situation you might see sentence A presented with a question mark, and sentence B presented with a star.