I have recently stumbled upon a new metaphor may might help us think more clearly about different approaches in linguistics: the “moving-parts” metaphor that is sometimes used by generative linguistics.
It came up first in a Twitter conversation I had with Peter Jenks, which was originally about how we can tell affixes from non-affixed material, which Jenks said “can’t be answered a priori” (but must be determined by theorizing). I objected that generative linguists don’t keep empirical generalizations and theorizing separately, arguing that
“Ideally, theorizing would be like Ice-Age theory, or Periodic-Table theory: Someone suggests an idea, more data come in, it‘s confirmed, some skeptics object, then there‘s more confirmation, and finally, people move on because the original idea has become common knowledge.”
To which Jenks replied that this is actually how generative syntax works: “There are just so many moving pieces and the theoretical space is large enough that it’s hard for outsiders to keep track of.”
But is it only “outsiders” who have trouble understanding what’s going on? The more I read about MGG syntax (and I read a lot more recently, because the phenomena they discuss are the ones that interest me), the more I see the big differences in the assumptions that people make and the kinds of ideas that they find promising.
The moving-parts metaphor came up again recently on Norbert Hornstein’s blog, where Hornstein voiced some criticism of dependent-case theory (which he says has become mainstream in MGG). Tim Hunter says that he’s confused by a claim of Omer Preminger’s about a Hungarian sentence:
“If the theory of case features has this many moving parts, and makes the relation between the observable morphology and the features so indirect, then how do we know that the relevant Hungarian sentence is identical as far as case is concerned?”
I had a similar reaction to Mark Baker’s book “Case”, where he explains (his version of) dependent-case theory really well. In my review of the book, I noted that Baker’s notion of case is so abstract that he neither rules out case that is spelled out by zero, nor zero adpositions that assign overt case, nor overt adpositions that look like case affixes. So a seeming case affix can be either true dependent case, or case assigned by a zero adposition, or an adposition that has somehow “merged morphologically.” And a seeming lack of case-marking is compatible with case and with absence of case. With so many moving parts of the theory, what predictions are still made?
In the discussion on Hornstein’s blog, it becomes clear that Hornstein, Preminger, and Julie Anne Legate have three very different views of the matter – and some people like Tim Hunter are honest enough to admit that they are confused (so are probably many other readers). Is this sort of theoretical uncertainty necessary, or can we do better, as a field?
Then (in a private message) I asked Omer Preminger whether the many moving parts of his case theory do not make it impossible, in practice, to test the theory. Preminger replied:
“I’d take issue with your suggestion that moving parts preclude predictions. Given an observed (surface) case patterns, the premises of a particular case theory will generate particular entailments about what the underlying syntax of that languages can and cannot look like. We then have other tools at our disposal (I’m sure you’re familiar with them) to investigate whether the underlying syntax does or doesn’t look in that predicted way. I suspect that, at this juncture, you might point out that there are also “moving parts” at these other junctures, and this is true. That is normal, I think, for any scientific endeavor that seeks to explain the causal forces behind a certain observed pattern, and, as such, I don’t think it’s fundamentally different from a functionalist explanation, which will also have its own moving parts.”
A week after the exchange on Hornstein’s blog, Julie Anne Legate came to Leipzig for her Brugmann Lectures (as part of the IGRA programme), and the issue came up again: She argued that languages are actually more different than linguists often think, and she gave an example of how Dinka supposedly shows an Austronesian-like voice system (cf. van Urk 2015: 69; Erlewine et al. 2017), but in fact the differences are substantial: While voice is located on the verb in Western Austronesian, it is located on the auxiliary in Dinka. To which Gereon Müller replied (because he knows that the parts can move): “But they could still be the same underlyingly, and the divergent surface result could come about via movement. One could try to find arguments for such an analysis.” (The “parts” can “move” in two senses, but this is just a pun.)
I found these exchanges very instructive, because I have devoted a lot of energy over the last decade to a very different sort of enterprise: To highlighting the difference between presuppositions of an empirical claim and explanations for it (i.e. the difference between comparative concepts, on which typological comparisons are based, and UG categories that might explain the worldwide distributions), and more recently also to standardizing terminology for comparative concepts that most of us (including generative linguists) use in one way or another.
So is it the case that “any scientific endeavor” must have so many moving parts? I don’t think so – I think that many scientists who work with observed data actually agree on many basic methods, especially those having to do with measurements. Thus, climate scientists mostly agree on how to measure temperatures, air pressures and ocean currents, and astronomers agree on how to classify different kinds of stars, so they can compare their empirical results, even if they don’t agree on what it all means and how the phenomena could be explained.
Some fields of linguistics, such as comparative phonology (e.g. Gordon 2016) and some kinds of comparative lexical semantics, don’t seem to have this problem at all. Phonologists are very often happy to use IPA symbols as comparative concepts for segment types (even though different languages may have different features defining rather similar sounds). And comparative semanticists typically work with etic concepts such as colour chips and etic kinship grids (cf. Kemp et al. 2018 for some interesting recent work on worldwide lexical structures; also Levinson & Meira 2003 for earlier work using an etic grid for comparing spatial relations).
Thus, it seems to me that in general, we can distinguish three kinds of activities in linguistics:
– description of particular languages (ideally complete, conceivably using sopecialized descriptive concepts tailored to the language)
– comparison of a representative set of languages (using a special set of comparative concepts which apply to all languages equally, and which may be quite different from the descriptive categories)
– (functional-adaptive) explanation of the generalizations found through worldwide comparison.
Description is independent of comparison and explanation, and comparison is independent of explanation. A descriptive grammar of a language or a comprehensive dictionary keep their value, regardless of subsequent progress in comparison and explanation. Likewise, proposals for universals can be readily tested, regardless of the explanations that one might propose, and regardless of whether they hold up (for example, Greenberg 1963; 1966 proposed a wide range of generalizations that he made sense of in his 1966 book in terms of a notion of “markedness”; in Haspelmath (2006; 2008), I argued that these explanations were not adequate and proposed an alternative, but this did not affect Greenberg’s comparative discoveries).
In the functional-adaptive approach, there is of course some flexibility (or “movability of parts”) as well – descriptions of languages may turn out to be suboptimal (incomplete, or somewhat misleading) once comparative considerations are brought in, and depending on the explanations that one wants to test, one may want to adjust one’s comparative concepts. So Preminger is not completely wrong when he says that “it’s normal for any scientific endeavor that seeks to explain the causal forces behind a certain observed pattern” to have moving parts.
But there is still a big difference between generative syntax and the functional-adaptive approach: In generative syntax, there is basically no distinction between description and explanation, because both make use of the idea of a rich set of UG building blocks: The correct description (“analysis”) of a language-particular phenomenon presupposes that we have the right set of UG building blocks, and these building blocks also explain limits on cross-linguistic variability.
In an approach where description, comparison and explanation are logically separate activities, there are thus fewer moving parts, and it is more likely that robust results will be achieved.
Baker, Mark C. 2015. Case. Cambridge: Cambridge University Press.
Erlewine, Michael Yoshitaka, Theodore Levin & Coppe van Urk. 2017. Ergativity and Austronesian-type voice systems. In Jessica Coon, Diane Massam & Lisa Demena Travis (eds.), The Oxford handbook of ergativity. Oxford: Oxford University Press. doi:10.1093/oxfordhb/9780198739371.013.16
Gordon, Matthew Kelly. 2016. Phonological typology. Oxford: Oxford University Press.
Greenberg, Joseph H. 1963. Some universals of grammar with particular reference to the order of meaningful elements. In Joseph H. Greenberg (ed.), Universals of language, 73–113. Cambridge, MA: MIT Press.
Greenberg, Joseph H. 1966. Language universals: With special reference to feature hierarchies. The Hague: Mouton.
Haspelmath, Martin. 2006. Against markedness (and what to replace it with). Journal of Linguistics 42(1). 25–70. doi:10.1017/S0022226705003683.
Haspelmath, Martin. 2008. Frequency vs. iconicity in explaining grammatical asymmetries. Cognitive Linguistics 19(1). 1–33.
Haspelmath, Martin. 2010. Comparative concepts and descriptive categories in crosslinguistic studies. Language 86(3). 663–687.
Kemp, Charles, Yang Xu & Terry Regier. 2018. Semantic typology and efficient communication. Annual Review of Linguistics 4(1). 109–128. doi:10.1146/annurev-linguistics-011817-045406.
Levinson, Stephen C., Sergio Meira & Language and Cognition Group. 2003. “Natural concepts” in the spatial topological domain: Adpositional meanings in crosslinguistic perspective: an exercise in semantic typology. Language 79(3). 485–516.
I am not sure if that’s what Jenks meant, but I am reading the phrase “moving parts” different than you do. Your reading of the metaphor, if I understand you, is that many parts of the framework are changeable ad hoc, which makes it a poor theory. That is a valid criticism (which I myself agree with here). However, I read Jenks’s phrase “moving parts” as a metaphor for a complex machine with interlocked parts, like a clock. I read him as saying that MGG is a complex framework, with a steep learning curve, which makes it hard to approach. That view of MGG is one that’s both its practitioners and detractors would agree with, and that is what I thought Jenks had meant when I read his comment.
Following up on the previous comment, then, if my criticism of functionalism is that it sneakily presumes the universal in much that it claims to be language-particular, my criticism of MGG (and it’s yours as well) is that it sneakily presumes the language-particular in much that it claims to be universal.
I think this is perhaps revealed by Omer’s comment that you quote (though I’m sure he’d offer a more considered opinion given the chance) in which he talks about ‘predictions’ and ‘entailments’. I think there is a subtle difference here relating to the distinction between explanatory and descriptive adequacy, and I think a lot of theories in MGG pretend to reach explanatory adequacy by having a veneer of entailments that only masquerade as predictions. If I were to be super cynical about it – and this is really being unfair just to make the point – MGG analyses don’t predict anything at all about phenomena not yet observed; they merely entail (at best) that presently observed phenomena are not inconsistent with theory.
Going back to the planets for an analogy, at the time when only 7 planets were known, the outermost, Uranus, exhibited orbital timings that were inconsistent with predictions calculated with Newton’s law. There were two avenues offered at the time: either Newton’s law is taken as a basic principle and we extrapolate that there must be an 8th planetary body exerting an influence on Uranus, or Newton’s law is a good enough guideline that breaks down at long distances, so that we have to explain Uranus in its own terms, likely with the knock-on effect that it’s not much of a principled explanation of closer bodies either. History of course looks back kindly on the first proposal because that’s how we discovered Neptune and this is what predictions should do for you in any scientific theory but I don’t think MGG has anything remotely comparable to this.
The reason, to put it bluntly, is that MGG doesn’t have a coherent position on what our innate syntax *is* as a cognitive trait, so it doesn’t know what an explanatory theory would look like. All it supposes is that there *is* an innate syntax (and I agree with it in this) and that it is somehow responsible for the generation of sentences but, beyond that, it does not have a principled understanding of its properties. At the very least, it seems sensible to presume that there is a binary-branching Merge operation and that its function is to generate semantic compositions. But what beyond that? Are we going to have Agree? Are we going to have Reverse Agree? Are we going to have movement and which variaties? Are we going to have any derivational stages that are not motivated semantically? And so on.
The difficulty is that MGG doesn’t have any consolidated principles that constrain the UG toolkit, so instead of having a genuine idea of a universal toolkit that can be applied to language-particular phenomena, it merely tries to guess how to cram the language-particular into the universal. Thus, on observing various phenomena, it proceeds to invent universal tools that appear to satisfy descriptive demands, though its only metric for assessing the viability of these tools is whether they permit some degree of cross-linguistic consistency – it doesn’t have a way of finding out whether its tools are actually cognitively real. And to invoke your moving parts analogy, the extra trouble with permitting yourself to invent such tools is that if one tool works for one language but not for another, you can just make another tool for a second language to undo the work of the first. This is more Ptolemaic than it is Galilean.
The obvious charge I’m open to – and I accept it to some degree – is that this is all a fairly ordinary approach and so long as we cannot see directly into what is cognitively real, our best line of inquiry is to go on inventing these tools and to appeal to overall theoretical simplicity as a measure of what is most likely to be true. My issue with this is that I don’t believe MGG knows even what simplicity looks like, given how many fractured paradigms there are in competition. If you have two totally different theories that broadly cover the same data, how are you going to decide between them? This is not a trivial question in MGG because it has always billed itself as a cognitive theory, rather than as a bluntly linguistic theory, yet I think it has forgotten how to play that game.
Personally, I think a lot of important things have been discovered in MGG and remember that I am in league with it overall, but I criticise it this way because I think it assumes too much that it has not proved and it needs to ask some more fundamental questions than it is used to about what UG is for and what we can reasonably expect it to contain, particularly with regards to the logic of acquisition. Although MGG might not be the predominant means of linguistic description these days (I don’t know what the trends are statistically), I think its theoretical apparatus has become sufficiently entrenched that it is too often deployed without full consideration.
I suppose I comment as a representative of MGG syntax but I’m also sympathetic to some of your points of criticism.
Let me start by observing a tension, though. At one moment, you say, “description is independent of comparison and explanation” and shortly afterwards, “descriptions of languages may turn out to be suboptimal … once comparative considerations are brought in.” Now, you might want to rephrase things but, put this way, this is surely a contradiction and I think it is because, at least to my mind, you haven’t quite solved the Boasian problem of comparing the incommensurable.
I think there would be more consensus on this with a carefully drawn distinction between levels of analysis, in the sense that we can BOTH say that description should be independent of explanation AND that there should be no distinction at all, this being a matter of perspective. Why? Because this just encapsulates good old scientific reductionism; it’s the dissolution of “why?” questions into “how?” questions. To form an analogy, we can perfectly reasonably describe the orbits of the planets superficially, finding no congruence in their shapes, but at some point we want to be able to describe them as a function of Newton’s inverse square law and then what is incommensurable at one level of analysis becomes identical at another, though both are valid.
With this in mind, however, I think there are holes to pick on both functionalist and generative sides of the debate. For example, regarding language-particular description, if it is to be independent of comparison and explanation, we really must not have our cake and eat it. For example, in your discussion notes paper, you say:
“The argument for categorial particularism is simple: the criteria used for identifying categories such as cases, word classes, and grammatical relations are themselves language-particular. Subjects in English are identified by a set of criteria that is only partially comparable to the set of criteria that might be used to identify subjects in Tagalog (cf. Schachter 1976) … [and] there is no principled way of deciding which criteria are relevant.”
Now, I’m on board with the idea that the criteria for the English subject category should be constitutive only of a definition of English subjects, and the same for Tagalog, without presuming that these criteria must be drawn from a universal set that is constitutive of a universal subject category. However, the moment that we recognise subjecthood *itself* as a common property, we commit ourselves to the enterprise of determining its universal basis, as we have repudiated its identity as purely language-specific.
One might try to evade universality by claiming that there’s just an awful lot of typological coincidence and that what we call “subjects” in English and Tagalog are really totally distinct things with curiously overlapping features. But while this is arguably sensible in categorial terms, there is an extremely clear justification for treating language-specific subjects as instantiations of a universal *semantic* type. That is, the reason why we have the compulsion to label these language-specific entities in the same way is because they have similar properties with respect to the semantics of subject-predicate propositions.
Crucially, this is not a syntactic criterion and this is why I think the kinds of comparative concepts you describe are better written off either as useless or uninteresting. To the extent that common structural phenomena seem rooted in semantics, they should be accounted for in semantic terms, as it is semantics that is truly universal. To the extent that phenomena seem instead not to have a semantic grounding (like agreement), I think we plainly just don’t understand how these should be characterised at a comparative level yet because the past fifty years of MGG has prioritised semantic compositionality as the basis of syntactic structure.
I’ll turn to the associated problems with MGG in a separate reply so as to avoid mixing them up.