Gibberish

Today in Syntax Reading Group we discussed Pesetsky and Torrego's 2004 paper The Syntax of Valuation and the Interpretability of Features. This paper has been quite influential and has even raised a number of eyebrows (apparently for being a direct challenge to one of Chomsky's theories?).

In an nutshell, the topic is something like this. Various syntactic phenomena are instances of what we might call feature sharing. So, for example, Latin adjectives agree in gender, number and case with the nouns they modify. In puella pulchra ("beautiful girl"), the adjective is "pulchra" (rather than "pulcher" or "pulchrae" or "pulchram") to show that it modifies a noun which is singular, feminine and nominative. Things like number and gender can be thought of a "features" of the noun.

Chomsky wrote some stuff in 2000 and 2001 spelling out a kind of typology for features. In particular, features vary on two oppositions. They are interpretable/uninterpretable and valued/unvalued. The idea with valued/unvalued is pretty clear: this is the position that the adjective "pulcher" isn't specified in the lexicon as to whether it's feminine, masculine or neuter, etc. It comes into the derivation in some indeterminate form, and then it agrees with the head noun. We believe this because adjectives seem to vary for the nouns they modify - not vice versa. (Evidence for this in English would be nouns like "scissors," which is inherently plural. We are aware of no adjectives which are inherently plural. The same phenomenon exists in Latin - there are nouns which are inherently feminine or plural, but adjectives seem to be able to combine freely with any noun, and to change their forms to fit that noun.) The idea, then, is that adjectives have "unvalued" features for person and gender that get valued when they come into contact with (are in the same syntactic domain as) a noun. Chomsky uses the somewhat misleading terms probe and goal for this. A "probe" has an unvalued feature and it looks for a "goal" (that invariably comes later in the sentence) which has a value for the feature. So in the example above, "pulcher" is the probe and "puella" the goal. Once the probe "pulcher" hits the goal "puella," it values its gender featuer as [feminine] and becomes "pulchra."

The interpretable/uninterpretable dichotomy is not so straightforward. What it seems to mean is "relevant to semantics." So, if a feature somehow affects the meaning of a sentence it is "interpretable." If not, if it is there for purely grammatical reasons, it is "uninterpretable." Gender features are presumably uninterpretable in most cases, whereas number is probably interpretable. Now, this is where Chomsky's theory gets a little weird. Chomsky's Minimalist Program has a principle of Full Interpretation, which says that uninterpretable features can't be passed on to the semantic module (called LF). The semantics only deals with things that are relevant to meaning. (It goes without saying that Chomsky is committed to a highly modularized model of grammaticality. Syntax and Semantics are separate things, as are syntax and pronounciation, etc.) So, if we get an agree relation related to an uninterpretable feature, Chomsky stipulates that this feature then simply deletes. (If you're like me and you have issues with the idea of things "deleting" when there is no possible empirical test that could verify that any such thing happens, then it's best to think of it this way: syntax is a kind of compiler. It transforms an underlying form into something that the interface can use. Just like in a computer program, anything that's in the program but is ultimately inconsequential to the execution stage will just be ignored by the compiler. I prefer to think of "deleting" as a shorthand for saying "this was here just for syntactic purposes and it will never be used again - so we just ignore it from here on out." Of course, if something with an unvalued/uninterpretable feature never gets that feature valued, then the sentence will be ungrammatical. But once its condition is satisfied then it ceases to play a further role.) It deletes because of Full Interpretation - becuase the Semantics doesn't know what to do with it, so we just stop considering it after we know it's satisfied. Chomsky's theory is also a bit weird in that it just sort of assumes that any unvalued feature will also be uninterpretable. The reasoning here seems to be that Syntax cannot possibly know what the Semantics will need to interpret - and yet it still needs to know which features to delete as part of its operations. So, it just deletes anything it values. Ergo, language is constructed in such a way that only the features that cannot possibly be meaningful to Semantics sometimes happen to be unvalued.

This is what's extremely annoying about Chomsky's theories. He gets carried away with the analogies he uses. There's no reason the Semantic module (and that's another thing - who says grammar is modular to this extent? Why can't Syntax and Semantics all be part of the same process - or at least happen simultaneously?) can't simply ignore the things that don't apply to it, no? I mean, even if we buy this modularity and say that Syntax can't know what the Semantics will need, it doesn't follow from that in any way that Syntax needs to delete anything - because for all we know the Semantic module might be capable of just ignoring whatever residue is left over from the Syntactic operations. Fine, you say, but does it matter? Well, yes, I think it does - becuase it leads Chomsky to goofy "generalizations" about language such as "only those features which don't apply to the Semantics can ever be unvalued." As though this were a useful or insightful thing to say! Why build machinery you don't need?

So Pesetsky and Torrego wrote this paper to try to weed out some of the unintuitive aspects of the theory. Fine, noble goal and all - but (and this is the point of this post) the solution they end up coming up with is more or less just HPSG plus a bunch of irrelevant stuff on top. What they end up saying is that, rather than one feature valuing another, what happens instead is that the features share values. Which is exactly HPSG. HPSG is a theory that builds up syntactic/semantic (there is no modularity) objects successively by adding on words. As words get added on, the whole object takes on the features of the new words. If there are conflicts, then the combination is not possible and the sentence is ungrammatical. It's a much simpler, more consistent system without all the extra baggage. And because it doesn't have the extra baggage, it avoids a lot of unintuitive stipulations. Like, for example, why is it that the "top" word is always the one with the unvalued feature (by "top" here we mean roughly "closer to the beginning of the sentence")? Why do we need this kind of "directionality" with "probes" and "goals" and all that jazz? Why can't it just be that one or the other has the valued feature, and one or the other is not valued, and whenever a word gets merged into the greater syntactic object it just feature shares - no directionality implied? I can't see that rigid adherence to this directionality buys you anything, and it seems to have the added danger of forcing you to conclusions you might not otherwise reach or need (we might, for example, have to at some point specify some movement operation just to make sure that the "probe" is ahead of the "goal" in some derivation. I can't think of an example, but it seems like a possible danger of writing this otherwise unnecessary directionality constraint into our grammar.). With HPSG, there is no directionality - just feature valuing, so no danger of later unnecessary stipulations (at least, not on this point.)

Pesetsky and Torrego seem proud of the fact that they've gotten rid of Chomsky's stipulation that all and only the unvalued features are uninterpretable. But this seems to me to come at a cost. Rather than getting rid of the extra baggage, they've added more on. Now we have a four-way distinction: rather than saying (as Chomsky says) that all unvalued features are also uninterpretable, Pesetsky and Torrego say it's possible to have unvalued-interpretable features too. And also valued-uninterpretable (from the lexicon - in Chomsky these delete as soon as they're valued). Great, so now we have four to deal with instead of just two (and those two were already pretty silly - because Chomsky goes to the trouble to say that there are two oppositions here - valued/unvalued and interpretable/uninterpretable - when actually we only needed one opposition. We could have just said "probe" and "goal" and not bothered explaining all this crap about valuation and interpretability, right? I mean, if there are only two kinds of features, why use four descriptive words to spell out their differences???). But the point is we don't need any of this save for the assertion that Syntax and Semantics can't talk to each other. That is, the only reason we're going into so much detail about different kinds of features in the first place is because we're assuming that there has to be some kind of operation that gets rid of all the stuff that Semantics doesn't need before Semantics applies. But no one has ever proven that Semantics needs to apply after Syntax. Why can't it just happen at the same time, like it does in HPSG? And even if we do believe in all this (completely unmotivated) modularity, why do we have to get so nitpicky about the features at all? Can't we just leave it up to Semantics to ignore the features it doesn't need? I mean - what the HELL is the motivation for stipulating that if we leave in a feature Semantics needs and Semantics happens to see it that the derivation will crash and the sentence will be ungrammatical? (???)

I dunno - maybe (probably) I'm mising something - but it just seems like a lot of handwaving over something that doesn't seem to have a whole lot of motivation. HPSG just works.

The reason I'm taking up so much time with this is because Chomsky is on record saying that HPSG is just a "notational variant" of Minimalism. What a crock. First of all, HPSG was talking about features long before Minimalism was a gleam in its daddy's eye. Back when HPSG was getting off the ground (in the 80s), Chomsky and his buddies were busy making functional heads for absolutely bleedin' everything and moving words willy-nilly all over the place to get the bottom line to work out. Hardly anyone's idea of an elegant or parsimonious theory, and not particularly intuitive to boot. But more importantly, if HPSG is a "notational variant," then it seems to be the superior notation and we should adopt it immediately. It doesn't commit us to things like four-way distinctions between features that we can't even study because they "delete" before the sentence gets produced!

What's annoying about Pesetsky and Torrego is that they actually cite some papers in the HPSG literature in their bibliography. So why on Earth, one wonders, didn't they just write something to the effect that HPSG gets a nicer solution than Chomksy? Oh yeah, and there isn't a scrap of data to be found in the whole paper. Aside from the Latin example given above and a couple of sentences which illustrate (but do not motivate) a side-point they were making about complementizers, there isn't a single motivating example anywhere in the paper. They're just "cleaning up" Chomsky's theory for him. Well thanks, but no thanks. HPSG has already done that well enough.

I think it isn't for nothing that my old syntax teacher said that "GiB-berish" is a technical term for the kind of Lingusitics that GB/Minimalism people talk about...

I have to give the presentation on the rest of this paper next week (Tossi did the first 8 pages this week), which means I'll need to read the paper very carefully. I'll be sure to write a correction post if anything I've said in this one turns out to have been unfair.