Friday, November 9, 2012

My Problem with Semantics

In her first Baggett lecture, Barbara Partee raised an interesting question: why does Chomsky seem ill disposed towards semantics?[1]  I’m going to address this question here, though obliquely. To avoid exegetical concerns, I will try to channel Chomsky and answer a related question: what are my problems with semantics?  You are not the only ones that find this bait and switch disappointing. Even I find the swap downward sloping.  However, maybe this will help: over the years, I have drenched myself with Chomsky’s writings on this issue and though I will almost certainly misrepresent his views what I will say, I believe, is Chomskyish if not 100% pure Chomsky.  It goes without saying that I hold him personally responsible for any missteps in the exposition that follows. I would also like to add that I talked about the stuff below with Paul Pietroski and Bill Idsardi and so whatever you don’t like that isn’t Chomsky’s fault is almost certainly theirs.  So with this much CYA, let’s get on with the post! Warning, it’s a bit long.

I have three reasons for being semantically cynical. 

First, semantics has an odd relation to what I take to be the central project of Generative Grammar; the investigation of and limning of the fine structure of UG (see here). This project takes the object of inquiry to be I-language and so is necessarily internalist (see here).  The first problem with semantics is that practitioners conceive of the discipline as necessarily not internalist. Two ur-texts for this enterprise are Lewis’s General Semantics and Language and Languages. These two texts define semantics as an externalist enterprise. In the first paper, Lewis excoriates “markarese” approaches to semantics precisely because they are internalist and eschew the semantic project of establishing referential dependencies between markarese features and mind external denotata.[2]  Lewis insists that semantics without language-world referential links is just not semantics, hence markarese, whatever else it might be cannot be semantics. In the second paper, Lewis defends the position that languages are more basic than I-languages (i.e. grammars) and that the features of the latter are ontologically secondary to those of the former (see here).  Grammars, being in the head, are not what semantics is about.  One virtue of Barbara’s lectures is that they make it clear that the Lewis perspective on the semantics enterprise is still a (the?) dominant conception, when semanticists think about these questions at all.[3]

In making his argument against markarese, Lewis was adopting a conventional view of what semantics is, a theory of the relation between representations and their external denotata. This adopts the well known tri-partite partition of syntax, semantics and pragmatics. At any rate, if semantics is the theory of referential relations between “words and objects” then internalist semantics is not semantics and if your interest is in UG and I-languages then this referential conception of semantics is not obviously relevant.[4]

Chomsky occasionally throws a bone to this kind of work and moots an internalist re-interpretation of the dominant model theoretic technology. However, there is also a line of argument where he suggests deeper problems with this kind of inquiry: the proposed idealization either misses the central facts or is technically superfluous.  Let me discuss these points in turn.

Wrong idealization: what distinguishes human language from other kinds of animal communication systems is precisely the looseness of the relation between lexical signs and their multiple open textured “denotata.” As Chomsky has repeatedly emphasized, it is not at all clear what terms in a natural language refer to. Does ‘London’ pick out a spatio-temporal local?  If so how we coherently consider the possibility that it burn down, it move 45 miles down the Thames and it get rebuilt.  Maybe, ‘London’ refers to some kind of functionally organized entity, say the organization of buroughs and towns that make up greater London.  But if this is what ‘London’ denotes how can London be 100 miles in circumference and densely populated.  What’s true about ‘London’ is true for books (musical compositions, essays etc.), which can be both physical objects (concreta) and notional ones (abstracta), and for terms like ‘average men,’ (which can have 2.5 children) ‘temperatures,’ (which can be 90 and rising) and almost any other word that one thinks of carefully. This vast polysemy marks natural language lexemes and distinguishes them from what we see in other animal communication systems, where in fact a crude kind of immediate referentiality is de rigeur.  Chomsky reinforces this point here by contrasting human and animal communication systems:

Maybe we don’t know the right things, but everything that is known about animal thought and animal minds is that the analogues to concepts…do happen to have a reference-like relation to things. So there is something like a word object relation.  Every particular monkey call is associated with a particular internal state, such as “hungry” or a particular external state, such as “There are leaves moving up there so run away.” [referring to a large survey by Gallistel Chomsky continues]…Animal communication is based on the principle that internal symbols have a one-to-one relation to some external even or an internal state.  But that is simply false for human languages-totally (25).

Conclusion: the loose fit between words/concepts and things in human linguistic systems as contrasted with the strong fit witnessed in animal systems speaks to the inappropriateness of notions like ‘reference’ in semantic theories interested in human linguistic practice. Referential theories abstract away from precisely those features of human systems that make them distinctively human, and that’s not good. So to the degree that we have a rough understanding of what a reference relation might look like, our use of language doesn’t display it!

There is a second line of argument, one that concentrates on the fact that it is of no obvious utility.  His argument here is by analogy with a model theoretic “phonology.”

Let us suppose that LI [linguistic item NH] has no I-sound but P-denotes some object that is external to the person; call it the phonetic value PV of LI…and suppose some computation on PVs yields the linguistic component of the sound of E, PV(E). PV could be something about the noises associated with the utterances…of E as circumstances vary...; a construction from motions of molecules, perhaps. The proposal could be elaborated by taking PV to be determined by social and physical factors of various kinds…
            The proposal leaves all problems where they where, adding a host of new ones. We understand nothing more about the relation of E to its external manifestations. The account of communication and other processes is worthless…

Chomsky then extends the analogy to the semantic interface and notes that it that it too doesn’t “advance” our understanding, as it “merely restate[s]” the original problems. Setting up this kind of referential story teaches us “nothing about how expressions are used and interpreted.” (177-8).  In sum, referential theories of meaning are either deeply misleading or of no utility, at least if one’s interest is in understanding how meaning works in I-languages or its roots (if any) in UG. 

Third: one answer to Chomsky’s polysemy examples relies on conceding that human lexical items are severely polysemous and to manage this fact by introducing disambiguating indices.  This too is unattractive for two reasons.

First, there appears to be no upper bound on the number and kinds of indices relevant to determining how we are to interpret an expression.  Aside from the simple indices like time, place and interlocutors, one’s hopes, dreams, ambitions, disappointments etc are all potentially relevant in getting a fix on what a term is supposed to refer to in a given context.  The problem is not that we cannot fix relevant parameters given a context, but there is no plausible suggestion of what the relevant parameters are and how to fix them across contexts.  What we find are not theories of indices but examples with no reason to think that the list of relevant indicial parameters is anywhere near complete or completeable.

Second, the polysemy problem goes to the very heart of the system once one considers that the same expression can bind distinct variables and give them different referents. Chomsky has provided many examples:  "The book that weighs twenty bounds is inspiring" (physical object in relative clause and abstract one in main clause), " The temperature which is now 90 is rising quickly" (value in relative, function is matrix), " John ate (some) lamb last night that was grass fed/slaughtered in the kosher style (mass in matrix, count in relative clause)."

In these examples the two variables relate to the same antecedent yet receive distinct denotations. It appears that the polysemy is not resolved but carries all the way to the interpretation of the various bound variables, a problem that assigning separate indices for each interpretation of the antecedent will exacerbate, not ameliorate.

I have reviewed three reasons that syntacticians like me are skeptical about the semantic enterprise as currently practiced. They all boil down to the same point: they don’t appear to reveal much about the structure of UG and actually presuppose a vision of the linguistic enterprise antithetical to the one that places the structure of UG and I-language as its object of inquiry.  Someone once suggested a verbal distinction to me that is useful here. Linguists are people who study FL/UG and the properties of I-langauge. Languists are people who study language.  Syntacticians are linguists, semanticists languists.  Linguists are skeptical that notions like 'language' pick out scientifically manageable objects of inquiry. Langusits think grammars are suspicious abstracta. Not surprisingly the two groups have trouble understanding each other. Thanks to Barbara Partee for making it clearer to me why this is so.

[1] Indeed, he considers the possibility that “natural language has only syntax and pragmatics” semantics existing “only in the sense of “the study of how this instrument whose formal structure and potentialities of expression are the subject of syntactic investigation, is actually put to use in a speech community…” More pointedly, there is no “provision for … “the central semantic fact about language,…that it is used to represent the world,” because it is not assumed that language is used to represent the world, in the intended sense.” (See here p. 132). In other words, Chomsky believes that the abstraction to a semantic level of analysis, which isolates a reference relation as the fundamental feature, is misguided.
[2] The specific target of criticism was a proposal developed by Fodor and Katz but their theory was stand-in for a broad range of non-denotational theories. I should add, the Fodor-Katz theory was, to my mind, not that terrific, but less because it was internalist than because postulating features ad hoc does not carry much explanatory oomph.
[3] Barbara pointed to the last chapter of Dowty’s Word Meaning and Montague Grammar as a representative example.
[4] Fillmore, quoted in Dowty’s book (375) makes a similar point that Dowty rejects: Fillmore said:  “…issues in semantics that have no conceivable application to the process of comprehension cannot be very important for semantic theory.”  Fillmore suggests this as a “relevance test” for evaluating research in semantics. Replacing ‘the process of comprehension’ with ‘the structure of UG’ in the above pretty well sums up my view as well.  Btw, Dowty agrees that the semantic enterprise “has in principle nothing whatsoever to do with what goes on in a person’s head” and thus has nothing whatsoever to do with the structure of UG.  Different strokes!


  1. Something that strikes me as odd about formal semantics textbooks (eg Chierchia&McConnel-Ginet) is that they start out with discussions of language and world, Cresswell's picture of an open door, etc, and then after this rather short introduction plunge into the details of how to define entailment, which can of course be interpreted 'internalistically' as one of semantic relationships of Katz 1972. In fact it really has to be to make sense of how formal semantics is actually done, because we're discerning the different entailments of things like 'the boys (each) carried a piano downstairs' inside our heads, not by rushing around organizing piano moving activities out there in the real world, and seeing what kinds of sentences observers used to report on what was going on.

    My hypothesis about the reason for this is that it's a confluence of two not-entirely-unrelated factors.

    One, it is indubitably the case that language has 'external significance'. Frank Jackson's version of one of the standard arguments for this is that the philosophers can discuss where to go for drinks after the seminar, go their separate ways, and reconvene 40 minutes later at the same pub. Therefore, in the eyes of ANU philosophers, anti-realism is insane (consensus view of the PhD students at a reading group I was attending a year or two ago).

    The other is that model theory, which sort of looks like a theory of how language can have external significance, is simply a far easier tool to use to define entailment and similar 'logico-semantic properties and relations' (term stolen from Larson&Segal's textbook) than proof theory, as Jerry Katz tried to to (without knowing any proof theory afaik, and trying to do it at a time when it was much harder to learn anythign about it than it is now).

    The reasons for this are not entirely clear to me, but one point to mention is that you can show that your account of entailment doesn't deliver A1,...,An |- B for all Ai, B by simply producing a countermodel, whereas to do this with proof theory you have to show that no sequence of deduction rules can deiver B from the Ai, which is likely to be harder. & that your model theoretic account produces some entailments (such as the ones that you want it to capture) can typically be shown by arguments at about the level of sophistication of sophomore abstract algebra, which ought to be manageable by people capable of doing any kind of linguistic theory.

    Also, the model theoretic accounts contain a lot less arbitrary detail than proof-theoretical systems, so that people seem to find them easier to discuss, and, finally, they do at least sort of model some of the cognitive processes that Byrne and Johnson-Laird have investigated in their studies of deduction.

    So, according to me, various factors add up to

    a) make model theory the only practical formal technique for people who want to study entailment and similar 'meaning-based properties and relations' (my preferred replacement for Larson and Segal's term).

    b) generate a very strong illusion that semantics is necessarily referential (so, I think Lewis was just wrong, which is an insanely arrogant thing for a linguist to say, but this is only a blog comment...IANAP).

  2. This comment has been removed by the author.

  3. Hello, and thanks for the blog (from a philosophy graduate student interested in linguistics). I am in a seminar this semester where precisely these issues are under discussion--with Rob Stainton over at UWO--and I am reading everything here with close attention...

    I've found Dr. Partee's overheads for these lectures, but I wonder if audio (or video) of the lectures is available?

  4. I hope to have something more than the handouts posted in the near future. These hopes may not see fruition, but we shall see. I'll note it when they are posted on the UMD website.

  5. A short note on Avery's comments: I agree with the diagnosis but would add two points. First, that for the longest time semantics was DEFINED as involving word-world relations. So if one identifies meaning with semantics and semantics with word-world relations then on the latter are really meaning theories. This was effectively Lewis's critique of Fodor & Katz in his 'General Semantics.' Weak argument but apparently very persuasive to some. K&F wre not doing semantics by definition. Second, word-world relations open a door that many are strongly insistent on pushing on, viz. environmentalism. There is a large literature out there that thinks of semantics as the beachhead against the rationalism of the Chomsky enterprise. Semantics is NOT internal hence Chomsky must be wrong. As Putnam said, meaning ain't in the head. The pull of empiricism should never be underestimated in my view. But, as with many intellectual movements, why some things caught on is overdetermined and your observations strike me as reasonable and maybe even right.

  6. I'd like to tone down my criticism of Lewis a bit, as follows. In some sense, at least some language use has real world significance, and is certainly the reason that we have it at all: language surely did not first develop for the purpose of aesthetic enjoyment of sharing surrealistic poetry.

    But, sez me, by the time the internal representations (perhaps formulated in some kind of merge of Jackendoff, Wierzbicka, Pustejovsky and any other internalists you happen to favor) cash out in terms of real world significance, such as the ANU philosophers successfully sitting around the same table in the correct pub instead of scattered across North Canberra, or somebody tossing their can of beer rather than taking a swig from it after being told that a wasp has just crawed in, so much other stuff of a non-linguistic nature has happened that it is perfect sensible for a linguist to disclaim any professional interest in it, ie, not care about the language-world connection as such.

    Of course, the technical success of model theoretic accounts of the MBPRs in certain areas, such as extensional, modal and other 'intensional' (in the restricted sense, as opposed to hyperintensional) contexts could be taken as an strong indication that the leading ideas of model theoretic NL semantics as pursed by people for the last 45 years or so have some basis in reality. (But the complete intractibility of belief/hyperintensional contexts also carries a lesson, I think.)

    & the observation that you can accept the technical apparatus without buying the entire philosophy is not at all original with me; Keenan and Faltz are very explicit about taking this point of view in their Boolean Semantic book.

  7. Yup, you can buy the technical stuff without buying into the referential perspective. However, it is then important to say what the technicalia mean; how it advances our understanding of semantic competence. More often than not the question of how to interpret the "results" is left unaddressed and this is too bad for being able to technically frame an issue is not the same as explaining what is going on. This said, I agree that much can be salvaged of interest.

    Where I am less clear about is whether the truth theoretic perspective illuminates anything at all. Paul Pietroski is writing on these issues a lot (including in his various posts here) and I have been largely convinced by his view that it adds virtually nothing. We can use language to refer and say true and false things, but whether this use is the central one in the sense of largely determining the structure of the apparatus, is far less obvious to me. There is an excellent old paper by Hartry Field ('Tarski's theory of truth') which goes into what would be needed for such accounts to explain semantic competence and observes that we have no even roughly persuasive account. I also found this convincing. So, yes, reinterpret as much as possible and drop the pretense that truth theories are self explanatory.

  8. My guess about what the truth-conditional approach is actually leading to is a useful intuition-pump, due to the fact that we really do seem to have facilities for assessing truth and falsity of a limited range of sentences against 'imagined situations' a.k.a. mental models in the sense of Byrne and Johnson-Laird, which are reasonably well represented by finite models. (Once upon I wrote something about this, in a 1993 BBS commentary on their book, touting Tableaux Logic as the right way to do semantics).

    It might be relevant that the standard logical inference rules appear to lack finite countermodels (my possibly distorted recollection of something that Barbara Partee once told me that Johann van Benthem and pointed out to her).

    Regardless, trying to do NL semantics with inference rules just doesn't seem to lead to very much, I conjecture perhaps because, in terms of Marr's levels, any concrete proposal you make (including with tableax) will contain too much in the way of arbitrary, impossible to justify algorithmic-level details which model theory lets you bypass.

    & of course, under just the right circumstances, with simple enough sentences, you can assess truth to a certain extent, although my wife wisely never believes anything I say about the nonexistence of some culinary substance in the refrigerator.

  9. I thought it was just my fridge.