Since the early 20th century, linguists have generally recognized that different languages are different not only historically (with different genealogical origins) and culturally (with different words reflecting their speakers’ cultural needs), but also structurally: The meanings of words cut up reality in different ways, and grammatical categories in different languages do not map straightforwardly onto each other. Phonological systems make use of phonetic possibilities in different ways in different languages. More generally, each language is structurally unique (Haspelmath 2021).
It is an important task of linguistic analysis to find unified descriptions for phenomena that might look disparate at first sight. This might be called the Structuralist Consensus – and the fact that the term “structuralist” is not understood well by many linguists nowadays merely underlines the extremely broad consensus for this view of differences between languages (an excellent example of this is Nick Evans‘s work, which is both highly sensitive to structural uniqueness and greatly interested in evolutionary questions and approaches (e.g. Evans 2016)).
Against this background, I have asked how systematic comparison of languages is possible (Haspelmath 2009), and I have spelled out an answer that has been implicit in “language typology” since Greenberg (1963), and often in earlier work as well: We use observer-made comparative concepts whose reality in language systems we do not presuppose. One language might have a word for “lower arm plus hand” and another word for “upper arm”, while another language might contrast “hand” with “arm” (see Brown 2005). In order to compare these languages, we just look at the way they refer to “fingers plus palm” and the “arm above the elbow”. These comparative concepts need not play a role in a language in order for it to enter the comparison. More generally, comparative concepts are observer-made concepts that are not thought of as part of a language – and comparative concepts like picture stimuli or Bible passages make it even clearer that they are instruments of research, and not research results.
But many salient comparative concepts have names that are similar to (and historically derived from) language-particular grammatical (or phonological) categories, e.g. “k” or “dative” or “serial verb construction”. This makes it somewhat difficult to see the difference between comparative concepts and language-particular categories in some cases. Some people have objected that I have not drawn the right distinctions, or that my proposals are creating a problem. As far as I can see, these objections are mostly based on misunderstanding (though occasionally, the critics make some good points that had not occurred to me).
In a recent paper, Matthew Spike (2020) criticizes my (2018a) paper in great detail and ends up arguing that instead of making a fundamental descriptive/comparative distinction, “we should appreciate linguistic variation as a key component of our explanations rather than a problem to be dealt with” [from the abstract]. But these two views are not in opposition to each other, and I did not say that “linguistic variation is a problem”. The problem that I identified is the presupposition that all linguistic categories must be innate elements of a universal grammar blueprint, and I’m sure that Spike agrees with this. Quite generally, Spike and I agree on all the crucial points, and we differ primarily in emphasis and terminology. (From Spike’s rhetoric, one would think otherwise, so I feel that this blogpost may be needed.)
Spike cites some philosophical literature about the nature of categories, and he emphasizes the ubiquity of variation in biology, which he says makes it difficult to see biological species in essentialist terms (something that I had done for reasons of simplification, to drive home the point that observer-made concepts, historically unique social/linguistic categories, and essentialist categories as in physics and chemistry have very different properties). Spike gives few concrete examples, but in his §5-6 he basically argues that there is a lot of variability everywhere and that the “fuzziness” that we see in linguistics is not a reason to draw a fundamental distinction between comparative concepts and historically particular categories. He suggests that if cross-linguistic categories have a history of “being used in successful inferences”, that vindicates them. Concretely, he suggests (as in Dahl 2016) that linguistic categories may be “homeostatic property clusters” (Boyd 1999).
Concerning the latter suggestion, there is the well-known example of tense-aspect categories that cluster in intriguing ways, as shown by Dahl (1985), and which led Dahl to refer to these clusters as “gram types” (e.g. perfective, progressive, habitual, perfect). And of course, we see many similarities elsewhere, and everyone would generally agree with Evans & Levinson’s (2009: 446) vague statement that languages are “characterized by clusters around alternative architectural solutions” (also cited by Spike, §5). These clusters may be seen as analogous to organisms that are generally assigned to species (cf. Spike’s discussion in §4), because species are variable and problematic as well, as shown in a rich philosophical literature. Biologists are just doing fine, even with their philosophical problems, so don’t let’s worry too much about categories in linguistics – this seems to be Spike’s message.
So why do I worry? Because it isn’t so clear that linguistics is doing fine. We know that biologists have been successful, especially since the 1950s, because their insights have been applied to concrete problems in a wide variety of ways. Comparative linguistics has very little to show in this respect – applications are mostly limited to contrastive studies for second language acquisition, and there have not been any breakthroughs here. (What linguists knew 150 years ago would mostly be sufficient for these purposes.) On the contrary, more and more syntacticians seem to be adopting a weird way of representing syntactic relationships, with word-class labels and abstract features interspersed, held together by a bewildering variety of movement and checking operations. How can linguists go astray in such a drastic way? My explanation for this is that some of us mistake linguistic categories for innate parts of the human mind (“natural kinds”). Spike does not consider this option seriously, and he is not interested in explaining the strange behaviour of many syntacticians (of the generative tradition). But the distinctions that I draw can explain it: Some syntacticians posit these strange theoretical elements because there are no limits on what they can posit. They can claim that all languages have a DP and a vP (in addition to a VP) because they use different criteria for different languages. They have many moving parts in their theories, so it would be difficult (or impossible) to show that they are wrong. And because of the traditional prestige of Chomsky and MIT, the tradition continues (with many interesting low-level discoveries along the way, of course).
Spike seems to be concerned with something entirely different, which I understand only partially. He thinks that the comparative concepts that I take as the basis for comparison “are not ultimately meaningful” (§2), and he even attributes to me the view “that meaningful cross-linguistic comparison is impossible, and that any scientific inference on the basis of such comparison is invalid” (§2). He finds this “terrible news”, but it seems that he is not aware of the many papers (plus my two 1997 books) in which I have engaged in worldwide cross-linguistic comparison – in a “meaningful” way, I hope.
Apparently Spike has misunderstood what I say about inferences from one language to another language (the General Category Fallacy, which he mentions in §2). While identifying something as a natural kind provides definite information (e.g. the hypothetical finding that an exoplanet contains plenty of water, or that a population of surviving kiwi birds have been discovered in a remote location of New Zealand), identifying something as a “serial verb construction” in a newly discovered language does not provide the same kind of information if different criteria can be applied in different languages. But many linguists act as if this were the case, and this is what I have been criticizing.
I admit that this may be a matter of degree: Indeed, if variability in species were very great, some skeptics might say that the newly discovered kiwi population is “not really kiwi”, and maybe even the exoplanet water might be said to be “not really water”, given its isotope properties. However, the differences in degree are HUGE – and a philosopher like Spike may not be able to appreciate this, for lack of experience with cross-linguistic phenomena. But it should be sufficient to read two dozen of the chapters of the handbook of polysynthesis (cf. Haspelmath 2018b) to realize that this is probably not a clear cluster, or to look at phoneme systems in a few hundred languages (cf. Mielke 2008) to realize that languages have many different kinds of feature systems. As a result, there is very little agreement among linguists what the universal categories are, beyond the category labels inherited from the 19th century. The problem is not the lack of evidence for the Structuralist Consensus – the problem seems to be rather the stereotype that languages are generally so similar that the main difference is in all the different words (“All languages have basically the same grammar, and they differ only in their lexicon” – this was the view of 19th century European linguists who only knew the languages of Europe, and it is still the view of some prominent generative linguists who are not seriously interested in linguistic diversity.) Yes, everyone knows that the reality is a bit more complex, but it is apparently possible to focus one’s attention on the stereotypical similarities. If biologists were confronted with the same kind of situation, there couldn’t be any laws prohibiting the shooting of wolves (because we cannot really distinguish wolves from deer), and we wouldn’t be able to treat tuberculosis by antibiotics (because we cannot really distinguish Mycobacterium from Escherichia).
So I don’t want to insist that the difference is necessarily “ontological” (not being a philosopher, I have no strong opinion on what actually exists and what doesn’t). But in practice, there is a huge difference between categories of phenomena that exist in the world before we start observing them, and concepts that we as observers use in order to classify phenomena that do not fall into clear categories for the time being.
Spike distinguishes two kinds of scientific inquiry, statistical and causal-explanatory, and he seems to say that (possibly arbitrary) comparative concepts are sufficient for the first type of research – while “naturalness” of categories is critical for the second type (§3). He mixes this up with his discussion of natural kinds, so I’m not sure I understand what he means by “naturalness”, but maybe what he means is that the comparative concepts that we use must have something to do with our causal theories in order for the comparison to be useful. There is of course no question about this. Take an example from economics, where the comparative concept SME (small/medium sized enterprise) is often used for comparisons across countries – where presumably the size of these companies plays a crucial explanatory role (e.g. they may be more flexible, more vulnerable, etc. because of their size). If the causal question concerned the climate footprint, it would be absurd to only consider the size of the company and not also its activities (use of primary materials, pollution, etc.). So clearly, comparative concepts must be of the right kind to answer causal questions, and in this sense they must be “natural” – but this does not mean that they cannot be designed by observers (in the way the concept of SME was clearly designed by economists).
If the Principles-and-Paremeters generativists (like Cinque 1999; Baker 2001) were right, then there would be no methodological problem: The comparative concepts would be simultaneously the explanatory concepts (because they are innate categories – a biocognitive explanation), and these same categories would also be used for description, because all languages are made of the same building blocks (the way all creatures are made of the same DNA code). Spike keeps saying that I “insist that linguistic categories are never natural kinds” (§2), but this is actually not the case: I don’t insist on this, but I observe that we have not found the natural kinds yet, so that at the present stage of our knowledge, we cannot work with them. This reflects my pessimism about the success of generative grammar, not my principled opposition to these ideas (which Baker 2001 presents in a very beautiful and seductive way).
So “do cross-linguistic categories have a history of being used in successful inferences”? (§6) This is the question that I have been worrying about for two decades: How much of our endeavors have been successful? As for the explanatory stories that we tell, we may not have been successful – there is no fully independent way of assessing success, as far as I can see. So what I have been concentrating on is the question whether at least our cross-linguistic generalizations are true. And I don’t see how we can find out if we do not have objective definitions of our concepts. Spike keeps talking about “fuzzy” categories (and how they are normal in many fields), but the degree of precision is not the issue. The question is whether our categories can be applied objectively. In my experience, there is a large amount of subjectiveness in linguistic analyses, which forces linguists to join particular subcommunities (“theoretical frameworks” or research traditions). It is this problem that I think we need to tackle as a discipline. (Spike criticizes me, in §5, for saying that natural kinds can be identified by symptoms, which he says is not the case in the natural sciences. If so, then this widespread practice in linguistics looks even less justifiable – and I was trying to be charitable toward generative linguistics.)
As noted earlier, Dahl (1985) has discovered very interesting “category clusters” in the domain of tense and aspect, and Bybee et al. (1994) expanded on these insights. But how did they find these clusters? Crucially, this was NOT by hypothesizing that a particular set of categories are found everywhere and using different criteria in different languages (as is explicitly allowed in generative linguistics; cf. Chomsky’s 2001: 2 “Uniformity Principle”). Dahl (1985) used a translation questionnaire to get objective comparative data, and Bybee and colleagues used a set of semantic comparative concepts that they identified before examining the grammatical descriptions – crucially, these semantic concepts were not taken from the grammars, but were designed by the typologists. This is precisely what made this research so successful. There may be more clusters of this type elsewhere, but we won’t know unless we apply the same kind of rigorous comparative-concept approach that Dahl and Bybee used.
What I like about Spike’s paper is that in addition to natural sciences, it also has some discussion of social categories, because few linguists compare their rules and categories to the concepts that anthropologists work with. Spike does recognize that anthropologists have similar problems, for example with the term “clan”, but he says that such comparative concepts should not be “dismissed out of hand” (§5). I wholeheartedly agree, but I would add that to increase their usefulness, they should be given objective definitions, which are not based on culture-specific ingredients (for example, a general definition of “clan” should not mention tartan patterns). Spike mentions that cultural categories are “negotiable and a matter of agreement”, which is perhaps less visible with linguistic categories (as they are rarely discussed consciously). But his one example, “French culture”, nicely illustrates my point: Nobody would doubt that French culture exists independently of an observer, even if its precise limits may be somewhat in question. This is what cultural categories have in common with natural kinds: Once we find them, we have made a discovery. If Australian ships had discovered Europe in the 14th century, their anthropologists would soon have discovered that there is a difference between Spanish and French culture, just as European scientists in Australasia discovered in that there was a difference between Maori and Aboriginal culture. By contrast, comparative concepts like “clan” are not discoveries, but instruments for research.
Spike ends with a comment on “the ubiquity of variation and the intractability of nicely-defined linguistic categories”, and how this should not be seen as a mark of “inadequacy”. I quite agree with the latter point, but thus does not mean that linguistic categories are “intractable” or that defining them “nicely” is somehow difficult. The issue for me has never been the existing variation of categories – what I am reacting to is the nonchalance with which many linguists resort to subjective definitions which make use of different criteria in different languages. And I do not think that this “opportunism” (Croft 2009) is a symptom of sloppy scholarship – much of this work is very conscientious and sophisticated. My analysis is that these linguists are either unaware of the crucial distinction between comparative concepts (as instruments of research) and language-particular descriptive categories (as discoveries), or that they consciously adopt the strategy of regarding the relevant elements as natural kinds (given in advance by our biology, as part of a domain-specific cognitive module), which are the building blocks both of particular languages and of explanatory general theories.
Baker, Mark C. 2001. The atoms of language. New York: Basic Books.
Brown, Cecil H. 2005. Hand and arm. In Haspelmath, Martin & Dryer, Matthew S & Gil, David & Comrie, Bernard (eds.), The world atlas of language structures, 522–525. Oxford University Press. ((http://wals.info/chapter/s6))
Bybee, Joan L. & Perkins, Revere & Pagliuca, William. 1994. The evolution of grammar: Tense, aspect and modality in the languages of the world. Chicago: The University of Chicago Press.
Chomsky, Noam A. 2001. Derivation by phase. In Kenstowicz, Michael (ed.), Ken Hale: A life in language, 1–52. Cambridge, MA: MIT Press.
Cinque, Guglielmo. 1999. Adverbs and functional heads: A cross-linguistic approach. New York: Oxford University Press.
Croft, William. 2009. Methods for finding universals in syntax. In Scalise, Sergio & Magni, Elisabetta & Bisetto, Antonietta (eds.), Universals of language today, 145–164. Dordrecht: Springer.
Dahl, Östen. 1985. Tense and aspect systems. Oxford: Blackwell.
Dahl, Östen. 2016. Thoughts on language-specific and crosslinguistic entities. Linguistic Typology 20(2). 427–437. (doi:10.1515/lingty-2016-0016)
Evans, Nicholas & Levinson, Stephen C. 2009. The myth of language universals: Language diversity and its importance for cognitive science. Behavioral and Brain Sciences 32(5). 429–448. (doi:10.1017/S0140525X0999094X)
Greenberg, Joseph H. 1963. Some universals of grammar with particular reference to the order of meaningful elements. In Greenberg, Joseph H. (ed.), Universals of language, 73–113. Cambridge, MA: MIT Press.
Haspelmath, Martin. 2009. Pourquoi la typologie des langues est-elle possible? Bulletin de la Société de Linguistique de Paris 104(1). 17–38.
Haspelmath, Martin. 2018b. The last word on polysynthesis: A review article. Linguistic Typology 22(2). 307–326. (doi:10.1515/lingty-2018-0011)
Haspelmath, Martin. 2021. The structural uniqueness of languages and the value of comparison for description. Asian Languages and Linguistics (to appear).
Mielke, Jeff. 2008. The emergence of distinctive features. Oxford: Oxford University Press.
Spike, Matthew. 2020. Fifty shades of grue: Indeterminate categories and induction in and out of the language sciences. Linguistic Typology. De Gruyter Mouton 1(ahead-of-print). (doi:10.1515/lingty-2020-2061)
I am not sure I agree with your description of the methodology Bybee et al. and I have been using. You say about Dahl (1985) that I “used a translation questionnaire to get objective comparative data”, and about Bybee et al. that they used “a set of semantic comparative concepts that they identified before examining the grammatical descriptions”. Then you say that what made “this research” so successful was that the semantic concepts were not taken from the grammars but were designed by the typologists. It is unclear if “this research” includes my work or if it only refers to Bybee’s group. In any case, I did not start out by designing comparative concepts. Rather, my method (which Bernhard Wälchli and I have later applied on parallel corpora) was to start out with “seed grams” (e.g. the English perfect or progressive) for which I tried to find counterparts with similar distributions in other languages. Although this method would normally involve a preliminary understanding of the semantics of the seed gram, the semantic analysis of a gram type could not be finalized until the data has been collected. As for Bybee et al. (1994), it is actually not clear from the account in section 2.5.1 of that book that the authors did have a final list of “meaning labels” before starting to read grammars. Rather, one gets the impression that the coding process was in fact to a large extent guided by what they found in the grammars — I think it would be unwise to finalize the list of labels in advance, since one never knows what can turn up.
It may be true that comparative concepts are not discoveries, as you say, but they tend to depend on discoveries of phenomena. If no clans had ever been observed, social scientists would hardly go to the trouble to invent the concept of a clan.
Thanks, Östen! We may talk about these matters a bit differently, but I don’t think there’s any disagreement. Your translation questionnaire and Bybee’s “meaning labels” are objectively defined comparative concepts, not elements of Human Language that are thought to exist independently of observers and can be identified in a variety of ways (like other natural kinds). Now of course, they were not arrived at in a random way, and they were not only instruments for research, but also results of research in the sense that you couldn’t have used these measurements if you didn’t already know (or suspect) that they would yield interesting further results. What is really crucial here is that they were applied to all languages in the same way: You wouldn’t have said that slightly different approaches to translating the questionnaire are required depending on the language’s other properties, and Bybee wouldn’t have said that a meaning component such as “generic” or “habitual” can be recognized in different ways in different languages. So there was no *presupposition* in this work that you would arrive at gram types. The gram types were a research result based on the object comparative concepts, and they can be verified objectively, regardless of the specofic analysis. This is crucially different from generative grammar, where you first need to find the right analysis before you can start comparing languages. – The gram types that you found may have the status of natural kinds, but your work has never presupposed this, the way that the existence of natural kinds like “vP” or “c-command” is presupposed in much work in generative grammar.