Martin: Many thanks, Gillian, for contributing a substantive comment on a recent blogpost about describing and comparing languages and framework-free theory. Instead of leaving your comment simply as it, here are some reactions of mine in a dialogue form (and thanks for adding a few more points, marked in italics below).
Gillian Ramchand: It seems logically obvious that IF there are things common to all languages, then every individual language will manifest them in some way. Can we agree on this first obvious point?
Martin Haspelmath: I’m not sure what exactly you have in mind, but it seems to me that this is not necessary. I think that implicational universals may be seen as properties of Human Language, but they may be irrelevant to many languages where the protasis does not hold (e.g. the implication that if a language makes a gender distinction in the first person, it also makes it in the third person looks like a deep regularity, but it is manifested only in languages with gender). So I don’t see g-linguistic regularities very closely connected to p-linguistic regularities.
Gillian: Fine. But I did not say that everything should be manifested, only that the manifestation has to be rooted both in what is necessary and what is possible, by definition. To be sure, in the absence of a deep understanding of all possible human languages simultaneously, it is rather difficult to know, when looking at that one language, which aspects are necessary and which contingent. The real question is the right way to proceed. Looking shallowly at many languages can often miss the internal organic genius of the individual exemplars.
Martin: I’m absolutely with you there! At least since Sapir’s time, fieldworkers have been emphasizing this. I think it was Sapir who first talked about the “genius” of a language, and around the same time in Europe, Mathesius talked about “characterological” study of individual languages.
Gillian: Every time I have seen broad swathe typological work on a language I happen to know rather well, I find even the crudest descriptive facts startlingly wrong. So, second and obvious point that I think we can all agree on is that we need many different linguists working on achieving deep understanding of many different languages from different language families.
Martin: Yes, of course. But we need to make sure that we don’t confuse comparison (a kind of g-linguistics) with p-linguistic description. For example, I have often heard people say that Matthew Dryer is wrong to say (in his WALS chapter on clause-level word order) that German and Dutch don’t have SOV order, but by his criteria, they don’t. He is applying special comparative concepts, and not trying to describe these languages in the most appropriate terms. There is no conflict with a description of German in terms of verb movement to second position and Dryer’s classification. The two are simply different views on the same facts.
Gillian: Now comes the tricky part. What methodology do the individual linguistic researchers apply to their study of their languages? Do they (i) take categories from received wisdom, collected abstractions from many other languages, linguists and theoretical frameworks, or do they (ii) try to find them emergent from the data? Can we not agree that in practice, all working linguists use a combination of the two? Doing the first thing is a crude caricature, in practice good linguists avoid this, but everyone is human and we all know it is hard to see outside our own unexamined assumptions at time.
Martin: I think that many linguists are quite explicit in that they strive to describe a language in terms of universal and innate categories that were hypothesized to be part of UG on the basis of completely different languages. For example, the idea that Chinese makes a finite/non-finite distinction wouldn’t have occurred to linguists if they hadn’t seen European languages and if they hadn’t hypothesized that some categories of European languages such as (non)finiteness are part of universal grammar. So I don’t think that all linguists want to avoid “taking categories from received wisdom” – in fact, many people see this as their central goal.
Gillian: I would say that the second strategy (finding categories emergent in the data) is also an unrealistic caricature. This is because categories cannot in fact be constructed purely inductively, but in a constant dialectic with hypothesis formation. Hypothesis formation in turn is informed by the linguists own expectations (formed via combination of experience and theoretical intuition).
Martin: Yes, I fully agree that one needs to set up hypotheses and test them, and of course, the kinds of hypotheses one tests are informed by experience and intuition.
Gillian: It is simply a mistake to think that it is possible to proceed purely inductively to create the individual p-grammars. Also, it may be a methodological mistake to attempt to do so. As an experienced linguist/fieldworker one has a lot of implicit expert knowledge and this informs hypothesis formation. If p-grammars were constructable purely by induction, we could let computers do it for us.
Martin: Here I disagree. I don’t know what computers can do, but I think that we can create grammars “purely inductively”, in the sense that we don’t need the model of any other language, at least in principle. Of course, such models help in practice, because languages are so similar in so many ways. But languages are also sometimes idiosyncratic, and in such cases, the models of other languages (or even more general g-linguistic frameworks) do not help. Take German strong and weak adjective declension. This causes non-native learners a lot of trouble, because it is so idiosyncratic and senseless. To state the generalizations correctly, one needs German-specific categories like “Definite Determiner” and “Indefinitie/Possessive Determiner”. There is no general framework that helps here. Languages are full of such idiosyncrasies, but many linguists wrongly try to subsume them under more general regularities, because they have the expectation that all grammars of all languages must be based on the same UG categories.
Gillian: I think it is simply a caricature (there’s that word again) of generative linguists’ methods to say that we import analyses wholesale from English. Here’s an example that _I_ think is interesting. The GB (Government-Binding) framework in the 1980s (Chomsky 1981) proposed the category of INFL as the label for the functional position that every assertable main clause seemed to require, later decomposed by Pollock (1989) into T and Agr. This was widely accepted at the time, and although Agr has now fallen out of favour as an independent functional head, T has remained. In the 2000s however, cue Ritter & Wiltschko (2014), who noticed, looking at Salishan languages and then at Blackfoot, that the obligatory category necessary to make an assertable main clause is not temporal but locational in the case of Salishan, and person in the case of Blackfoot (an Algonquian language). They then made the proposal that what was important in all of these languages was a piece of functional information connecting the event description to context (by which I mean the narrow Kaplanian context, c, consisting of the speaker, hearer, time, location and world of the current utterance). Their proposal was to rethink the category T as a more abstract category (which for historical reasons one could actually return to calling INFL if one chose), which performs an anchoring function. I found it an exciting proposal and one which now leads me to look in a different way at languages without tense inflection to see if there is anything else there that we might be missing. I do think that it is possible for anchoring information to be pragmatically negotiated, instead of explicitly represented in the grammar, but it does seem to be a necessary semantic ingredient that is required for purposes of inference in actual communication. So, it is possible in my notation/implementation for INFL to be contextually negotiated, but it is also possible for it to be negotiated by non-temporal contextual information. This is an example of where work on non-Indo-European languages successfully achieved a modification of the theoretical framework in a rather deep and enlightening way. I expect that kind of experience in my field, and I do not think that being a “minimalist” syn-sem person makes those kinds of things impossible. Quite the contrary, in the case of Ritter & Wiltschko, being specific about what they expected and not finding it didn’t just lead them to abandon the model and write a p-grammar for what they saw – they managed to see a higher level abstract pattern which I personally find quite insightful.
Martin: Maybe the importation was not “wholesale”, but certainly many ideas that were based on English were then applied to other languages, like finiteness to Chinese, or DP (i.e. determiners) to Russian. It may also happen the other way round, with an “applicative” notion imported from Bantu and applied to English. In all these cases, descriptions that focus on the “genius” of each particular language would be simpler, but generative linguists have typically not even considered such p-linguistic solutions. (Ritter & Wiltschko may not have framed their innovation in p-linguistic terms, but their proposals didn’t have much of an effect on the field of generative g-linguistics, and it’s not clear to me if it is more than simply a weakening of the general theory.)
Gillian: The guiding principles we are assuming are common are in fact extremely abstract – like for example, we are always looking for sensitivity to hierarchy, and we have an expectation that pronominal systems will be sensitive to participant vs. non-participant in the act of speaking. No linguist I know has a particular g-grammar in their head, since we do not know yet what the most general human language principles are. In my opinion it is a category mistake – g-level things are a completely different level of abstraction. Conversely, nearly every linguist I know who is trying to come up with p-grammars for particular languages is doing some combination of g-grammar study and p-grammar study, because all p-grammars are particular embodiments of whatever it is sits at the g-level.
Martin: Yes, as you say: “We do not know yet what the most general human language principles are” – this is why I find it so surprising (and disturbing) that many linguists think that they should teach the latest MGG ideas to students, as if they were somehow an established achievement that one can build on. So what MGG syntacticians do is to put a particular idea of g-grammar into students’ heads, and the students then find it difficult to think in more straightforward terms about the languages they study (apart from the fact that they realize that this would not be prestigious). Many younger-generation linguists seem to think that the MGG notation is merely a descriptive tool (as I noted earlier), whereas in fact it makes sense to use this highly specific metalanguage only if one claims that it is innate.
Gillian: Again, I don’t think the things we carry over to our understanding of a new language are as low-level and specific as you seem to think. And the minimalist tool box is much more open-ended and usable for many more different things than you seem to assume. It is also a much more heterodox theory, encompassing many more different views on what level of thing might be innate, than you seem to assume.
Martin: Yes, I realize that there are many different views among a group of people who look rather monolithic to outsiders because of their use of a certain notation and jargon that comes from Chomsky’s technical proposals (“vP”, “feature checking” etc.). You described this in an interesting way in your post “On being a generative linguist”.
Gillian: So are our disagreements substantive? Methodological? Or crudely ideological?
Martin: I think they are primarily methodological, because generative linguists think that they can use any kind of diagnostic to argue for any kind of category. For example, at the DGfS conference in Bremen earlier this year, Martin Salzmann and Benjamin Bruening both gave talks on the DP vs. NP issue (I mentioned this in an earlier post), but they applied very different diagnostics for different languages. A still different sort of diagnostic was chosen by David Erschler in his recent paper on Ossetic (mentioned earlier here). I think all of these arguments make sense (as long as one realizes that the categories one works with are innate), but they don’t seem to be resolvable with our current means, because there are too many possibilities. That’s why I recommended the more robust framework-free approach in my 2010 paper, which provides much simpler p-linguistic descriptions, making descriptions independent of possible further explanations (including perhaps in terms if innateness).
Gillian: In terms of methodology, I would not dream of trying to build an analysis of Norwegian adjectival definiteness inflection along the model of German. There are tons of distributional patterns and natural classes found uniquely in different languages. I am using my knowledge of highly abstract commonalities I see to guide me in hypothesis formation, and I try to express the particular in terms of the general, so that other linguists can follow the description up one level if they choose. One thing is sure: If everyone invents their own labels there is no hope down the line for comparison or commensurability of primes. So I like to know what people mean when they use a term. This means I have to belong to a framework, even if it is in some ways a moving target. So yes, I choose to teach my students a certain toolbox, but I also tell them to question everything.
In addition, I make them read everything in other frameworks in the empirical areas they are working on. My impression of certain people who style themselves as non-Chomskians is that they read philosophy of linguistics, but don’t actually read very many actual linguists working on stuff in my field. Some of it is rubbish theory-internal jargon, and other stuff is really quite excellent. Just the same as any other theory in fact. I think you overstate the innateness thing. Nothing I actually do in practice would change if I discovered tomorrow that everything that all languages have in common is highly abstract, cognitive and not specific to language.
Martin: Many thanks, Gillian, for this interesting converstion!
References
OpenEdition suggests that you cite this post as follows:
Martin Haspelmath (June 29, 2019). A conversation between Gillian Ramchand and Martin Haspelmath, on different perspectives in linguistics. Diversity Linguistics Comment. Retrieved October 4, 2024 from https://doi.org/10.58079/nsv4
I find it hard to see any real difference between practical versions of ‘framework-free’ linguistic description and what might be called ‘non-framework-bound’ description, where by the latter I mean using ideas from a framework as organizational tools but not trying to hammer too hard when the data does not quite fit, and making adaptions as needed (when you can think of them). In fact it seems to me exactly what Boas did, where the framework was traditional grammar as developed in the standard handbooks for Greek and Latin, and the main adaptation was dropping assumptions about what the inflectional categories and their values had to be, but retaining the basic organizational idea that there are in any language a certain number of word-classes (Parts of Speech; I don’t see him noticing new ones for open classes), and, in many, a discrete number of ‘inflectional dimensions’ with a discrete number of values. I think this is evident from the discussion in the Introduction to the handbook, as well as many of the grammars, where a great deal of traditional terminology is used (transitive, intransitive, subordinate, conjunction, declarative etc.). It seems clear to ne that Boas did not make up all this stuff himself, but modified the traditional framework to produce sensible results for the language he was looking.
But he did not get very far with syntax; the criteria for identifying subjects and objects etc. seem to be morphological; syntactic phenomena such as syntactic vs morphological ergativity were not detected. Indeed, I don’t think that the present apparatus of descriptive syntax could have arisen without the more aggressive attempts in generative grammar to make things fit into formalized frameworks. Without transformational grammar or similar, I don’t think anybody could have worked out the independence of case-marking and grammatical relations as found in Icelandic and various South Asian languages, or the existence of two subject-like notions functioning independently in Philippine languages or Mixed ergative languages. If you analyse things within some kind of framework, you open up the possibility of noticing things that don’t fit, whereas, in syntax at least, perhaps because of the inherent complexity of the material, without a framework, people tend to notnotice much of anything at all, or take a much longer time to do it (our trad. grammar framework goes back to Aristotle at least, and is a millenia-old work of many hands from many cultures).
So one reason for acting framework-bound, to some extent, is to notice data that embarrases the framework, and then try to improve it, a side-effect being greater insight into what is going on. Another reason is that without a definite formalized framework, there can be no substantive theory of language-learning, since the learner needs a hypothesis space to sift through. There are then lots of possibilities for what the nature of the space and the grammars in it might be. It might have a rather informative ‘prior’, and produce human-intelligible results (the ideal of generative grammar), or, a rather uninformative one, and produce results that are completely incomprehensible to people (mainstream deep learning, it seems to me), with numerous possibilities in between. Wherever the debate and most of the money goes, I strongly suspect that there will be a group of people who want formalized grammars that they can understand, to some extent, at least.
Otoh I would agree with Martin’s thesis that we should be wary of claims about restrictiveness, in the absence of demonstrations that they are actually helping to predict data that is too scarce to be plausibly learned from data that is reasonably plentiful (something that we can’t actually do in an adequately rigorous way until we have some idea of the amount and quality of data actually used for language learning by humans). Interesting, Ron Kaplan in 1985 gave a talk titled ‘Three Seductions of Computational Linguistics’ (written version in the 1995 ‘Formal Issues in Lexical-Functional Grammar’ volume) in which he rejected the common practice of slapping conceptual arbitrary restrictions onto your formalisms in the name of restrictiveness (I was told by somebody who was there that the talk was not well-received at the time, but I think that Ron was probably right). On this view, clean descriptions of the generalizations that you are actually finding is the main priority for framework development, restrictiveness an possible extra benefit, if it turns out the help explain learnability.
These are interesting thoughts, and I agree with the general thrust. But I think that the main innovation of generative grammar was not the idea that there is a universal framework that is expected to work for all languages, but that one should *complete the syntactic paradigms* (see https://dlc.hypotheses.org/1712). This idea was not around earlier – even though the Structuralists had questioned the universality of the “word” notion, it did not occur to them that just as we have complete inflectional paradigms, we also want complete syntactic paradigms. Once you add that desideratum, you get a large number of new phenomena, questions, and possible answers.
For me this passage was insightful (since the fragment is rather long I leave out some parts):
“Here’s an example that _I_ think is interesting. The GB (Government-Binding) framework in the 1980s (Chomsky 1981) proposed the category of INFL as the label for the functional position that every assertable main clause seemed to require, later decomposed by Pollock (1989) into T and Agr. This was widely accepted at the time, and although Agr has now fallen out of favour as an independent functional head, T has remained. In the 2000s however, cue Ritter & Wiltschko (2014), who noticed, looking at Salishan languages and then at Blackfoot, that the obligatory category necessary to make an assertable main clause is not temporal but locational in the case of Salishan, and person in the case of Blackfoot (an Algonquian language). (…) Their proposal was to rethink the category T as a more abstract category (which for historical reasons one could actually return to calling INFL if one chose), which performs an anchoring function. (…) This is an example of where work on non-Indo-European languages successfully achieved a modification of the theoretical framework in a rather deep and enlightening way.
If I understand this correctly, the theory uses theoretical terms such as T, Agr, INFL, independent functional head, etc. which are based on knowledge of all known languages. Relevant insights from one language (in this case Salishan languages and Blackfoot) have implications for other languages (in this case English). The reason for this is that at some level of abstraction, the theoretical terms (in this case T) refer to phenomena which are the same in all languages. Only by looking at more languages, we can gain more insight into the actual status of these concepts.
One might, however, also argue that in order to explain the data from English, one has to use insights from English, and to explain the data from Blackfoot, one has to refer to data from Blackfoot. Why should languages a priori behave in the same way at some level of abstraction? One could argue that a concept like T has no actual meaning in an individual language, and does not refer to anything ‘real’ (i.e. it has no ontological status). According to this point of view, theoretical terms like T won’t be successful in explaining the data, and will probably make things just more complicated. Since one cannot determine just by simple observation what a ‘functional head’ is, and whether it is present, there is a danger that linguists will posit a T for a language, where it is not helpful to better understand the structure of the language.
The question therefore is; how to determine whether there exists something like T as a universal category (even if it is not manifest in all languages)? And what is the ontological status of a theoretical term such as T? Perhaps the history of Agr that ‘has now fallen out of favour as an independent functional head’ could provide insight into this?
I definitely agree that there are definitely linguists with the assumption that there are universal categories, but in the case of Chinese finiteness, there is at least one linguist who staunchly opposes UG and generative grammar, but supports the view of a finiteness distinction in Chinese (Shi, 2001). I translate his last paragraph here, so that his view on universals of language are clear:
> In the study of Chinese grammar, there are two manifestations of an ‘Indo-European prejudice’. One is to take formal criteria applying to other languages to study Chinese questions. For example, other languages use tense as a standard to distinguish between finite and nonfinite verbs; Chinese has no morphology, so [they] judge that Chinese has no finite/non-finite distinction. The second type is to deny similarities between languages simply because some grammatical categories are realised in different ways in different categories, to believe that Chinese is a language of a completely different type, and even to advocate for adopting methodologies that are ‘only suitable for Chinese’. Actually, if we can brush aside these enigmatic and widely divergent forms, we can see that there are all types of similarities between languages. The same grammatical category may only be formally expressed in different ways in different languages.
I think what he’s saying here is that while there are no bounds on how forms are realised across languages, similarities in function still exist, and finiteness in IE languages and Chinese is one of those cases.
Shi, Yuzhi. 2001. Hanyu de xianding dongci he fei xianding dongci zhi bie [The distinction between finite and non-finite verbs in Chinese]. *Shijie Hanyu Jiaoxue*, 2, 23-27.
Thanks for sharing this! Yes, similarities of function can be found in a wide variety of ways, and linguists sometimes use form-derived terms for what really refers to function (e.g. “the genitive nominal”, when in fact they mean the possessor nominal). I think that in order to determine how similar or different languages are, we need to study forms and functions independently, and be clear what our terms refer to.