The General Category Fallacy: Why grammatical category-assignment does not give us more insights

Describing a language means finding recurring elements in texts: not only recurring phonemes and words, but also recurring constructions – and to describe a construction, one needs to have classes (= categories) of forms that can go into a constructional slot. Everyone knows this, so where is the problem, and what is the General Category Fallacy (described in my 2018 paper, in §1 and §7)?

By General Category Fallacy, I described the widespread (though often implicit) idea that in order to describe a language, one must make use of general categories that were originally proposed for a different language (in other words, that one must use a pre-established framework, rather than working in a framework-free way). But when saying that a form (say, the English word ago) is a member of some general class (say, the class of adpositions), we do not learn anything about it that we didn’t already know. In fact, the English adposition ago is quite idiosyncratic – not only in that it is a postposition but also in that it can have a postposed deictic modifier (as in three years ago today; Kortmann & König (1992: 678) even argue that it is in fact a preposition, paralleling from in three years from now). This may be an unusual example, but the general point remains valid: language-particular description can be complete and fully clear without making reference to general categories. Language-particular categories are sufficient. (As Edward Sapir noted in 1921 (p. 118): “Each language has its own scheme [of parts of speech]. Everything depends on the formal demarcations which it recognizes.”)

Why do linguists often think that the language-particular (p-)categories that they set up must be linked to general categories? On the on hand, they may not be fully aware of the important distinction between p-categories and g-categories (because the same labels are generally used for both). But on the other hand, they may want to make their work more generally relevant. For example, although I only have data on a single p-language (e.g. Lezgian), I may want to make general (g-)claims, because general linguistics is more prestigious. (In fact, some linguistics journals make “theoretical relevance” a requirement for submission, by which they seem to mean “general relevance”.)

So for example, I may want to dispute Dixon’s (1979: 64) claim that “all morphologically ergative languages are in fact split ergative”, by citing Lezgian, which has no split ergativity (in its grammatical marking), but is fully and gloriously ergative (its ergative suffixes like -a/-u/di occur with nominals of all types). Since Dixon is a famous typologist, I may be successful in having an abstract accepted – and then I can use my description of Lezgian to question a prominent general claim. I can thus make a contribution to general linguistics, even though I only study one language.

What is the (potential) problem here?

Languages are structurally (somewhat) unique – they not only have different phonemes and words, but also different constructions and different classes (= categories) of forms. Such classes are defined with respect to the constructions they occur in (as emphasized especially by Croft, e.g. 2001). So how do I know that what I call “Lezgian ergative construction” is in fact an ergative construction in the same sense as what Dixon calls “ergative”? Mel’čuk (1988) claimed that Lezgian does not have an ergative construction at all, so what if he is right? (see Haspelmath 1991; 2020 for more discussion of Lezgian and ergativity)

And maybe even Dixon’s Dyirbal does not have an ergative construction – if one adopts the perspective of Baker (2015: 22-25) (who follows Legate 2008), then Dyirbal and other Pama-Nyungan languages really have tripartite rather than ergative alignment, even though the ergative/accusative exponents are often zero.

So whether there is an ergative case in Lezgian is unclear (and depend on one’s perspective), but this has no effect on my description of Lezgian. (I called the -a/-u/di suffix set Ergative case in my 1993 grammar, using upper case for a language-particular category.) Description does not need general categories – and comparison can be based on clearly defined, objectively applicable comparative concepts which do not depend on a particular perspective or theoretical assumptions.

By contrast, the General Category Fallacy is built on the presumption that p-categories must be linked to cross-linguistic (g-)categories because these have an independent existence, probably as natural-kind categories of the innate grammar bueprint. If this were the case, then we would indeed have to find out what the “true nature” of a category of a language is before we can compare it with other languages – it would matter whether English ago is a preposition (rather than a postposition), because it would be mentally represented as such by English speakers. Likewise, it would matter whether the Lezgian suffixes are indeed ergative in Mel’čuk’s or in Baker’s senses. And if we did successfully identify a p-category with a g-category, then we would gain additional insights, because everything we know about the g-category would necessarily be true of the p-category (e.g. we could make further predictions about English ago, which go beyond its distributional characteristics that we already know).

But there is very little evidence for natural-kind categories, so identifying a p-category with a g-category is not a sensible goal of p-linguistics at this stage. For example, when confronted with a case marker, it is hardly helpful to limit one’s activities to asking whether it is an “oblique marker” or an “ergative marker” etc. – this would be a sensible research question only if these categories were innate. But since they are not innate (as virtually everyone now agrees – even Chomsky seems to have given up the idea of a rich innate grammar blueprint), the task of linguists is different: P-linguists must describe each language exhaustively, i.e. examine all the constructions which may be relevant to its distribution. Which label they use in the end is a purely practical question.

Comparison is then a second step, making use of concepts that allow us precise measurements. Thus, for the study of ergativity, we must provide a definition of “ergative case marker”, and then see whether Lezgian ergative case markers and Dyirbal case markers match this definition (they do, on my favourite definition, which follows Comrie 1978; Bickel 2011; and much related work).

By classifying a Lezgian Ergative marker as an “ergative flag” in a general-comparative sense, we have engaged in the activity of comparing languages, but we have not thereby learned anything about Lezgian. Thinking that we learn something about a p-category by classifying it as an instance of a g-category is committing the General Category Fallacy.

In a recent philosophical paper about grammatical comparison, Spike (2019) criticizes my description of what good comparative linguistics looks like. His own proposals (in his §6) are fully compatible with what I think of as good comparative linguistics, so there is no disagreement. But he seems to misunderstand what I mean by “comparative concept”, because he insists that comparative concepts need to be “natural” to allow induction (which they are in my understanding if they lead to interesting results – though we do not always know this in advance of the comparison).  At the same time, Spike does not understand my point that observer-made concepts do not lead to additional insights. I noted in the 2018 paper that “mountain” or “wing” or “respiratory disease” are like linguistic comparative concepts (e.g. “ergative flag”) in that they serve to classify for comparison, while successful classification does not in itself provide additional knowledge (this is different with natural kinds, where, for instance, classification of a disease as tuberculosis can save a life). Spike says (in his §5):

“despite Haspelmath’s pronouncement that a geologist gains no information upon learning that some landform is a mountain, there is a great deal of knowledge this could impart: mountains are formed via a limited number of processes of plate tectonics and subsequent erosion, they have predictable effects on the climate and environment, they foster similar patterns of ecology, and certain types of rocks and minerals are more likely to be brought to the surface depending on the context of their formation and region. There is a whole science dedicated to orogeny, the process of mountain formation. The label ‘mountain’ allows us to make any number of reasonable inferences about the history, makeup, and interactions of a specific mountain”

But how do we recognize a mountain? Wikipedia does not cite any authoritative orologist, but merely the definition used by the UN Environmental Programme, which presumably gets its authority from the “UN” label. According to this definition, a mountain of at least 2500 m, or 1500 m if the slope is greater than 2 degrees, or 1000 m if the slope is greater then 5 degrees. This definition is thus quite arbitrary, and merely a simple operationalization of the OED’s definition: “a natural elevation of the earth surface .. [which is] is impressive or notable.” Clearly, knowing that an elevation is 2505 m (and thus officially a mountain) high or merely 2495 m (and thus not a mountain) does not give us “a great deal of knowledge”. Mountains (like meteor craters, discussed on this blog in October 2018) are comparative concepts, not natural entities which by definition all have the same causal origin (like chemical elements or physical particles). The UN definition gives us a precise yardstick, but this yardstick is essentially arbitrary (like many other definitions that the UN provides, e.g. a “small or medium sized enterprise”, with less than 250 staff). Just like other measurements, comparative-concept classifications are essential for comparisons, but they do not give insights beyond the classification.

Getting back to our ergatives: Employing the standard definition, Lezgian has an ergative pattern, and the fact that it is not split is counterevidence to Dixon’s claim that all ergative languages are split in some way. This is not a great advance of the science of language, but it is something that we know for sure (and something that may get me an accepted abstract). If an ergative pattern could not be observed directly, but had to be inferred in a complex way using different criteria for different languages (as seems to be the case for Baker 2015, cf. the discussion in Haspelmath 2018b; 2020), then we would not even know this. We would be permanently locked in eternal category-assignment controversies (Haspelmath 2007: §3.3), with no exit in sight. Fortunately, comparative linguistics is better off, as it has very often managed successfully to avoid the General Category Fallacy.


Baker, Mark C. 2015. Case: Its principles and parameters. Cambridge: Cambridge University Press.

Bickel, Balthasar. 2011. Grammatical relations typology. In Jae Jung Song (ed.), The Oxford handbook of linguistic typology, 399–444. Oxford: Oxford University Press.

Comrie, Bernard. 1978. Ergativity. In Winfred P. Lehmann (ed.), Syntactic typology: Studies in the phenomenology of language, 329–394. Austin: University of Texas Press.

Croft, William. 2001. Radical construction grammar: Syntactic theory in typological perspective. Oxford: Oxford University Press.

Dixon, R.M.W. 1979. Ergativity. Language 55. 59–138.

Haspelmath, Martin. 1991. On the question of deep ergativity: The evidence from Lezgian. 44/45(1–2). 5–27. doi:10.5281/zenodo.225289.

Haspelmath, Martin. 2007. Pre-established categories don’t exist: Consequences for language description and typology. Linguistic Typology 11(1). 119–132.

Haspelmath, Martin. 2018. How comparative concepts and descriptive linguistic categories are different. In Daniël Van Olmen, Tanja Mortelmans & Frank Brisard (eds.), Aspects of linguistic variation: Studies in honor of Johan van der Auwera, 83–113. Berlin: De Gruyter Mouton.

Haspelmath, Martin. 2018b. Review of “Baker, Mark. 2015. Case. Cambridge: Cambridge University Press.” Studies in Language 42(2). 474–486.

Haspelmath, Martin. 2020. Ergativity and depth of analysis. Rhema.

Kortmann, Bernd & Ekkehard König. 2009. Categorial reanalysis: The case of deverbal prepositions. Linguistics 30(4). 671–698. doi:10.1515/ling.1992.30.4.671.

Legate, Julie Anne. 2008. Morphological and abstract case. Linguistic Inquiry 39(1). 55–101.

Melʹčuk, Igorʹ A. 1988. Dependency syntax: Theory and practice. Albany: State University Press of New York.

Sapir, Edward. 1921. Language: An introduction to the study of speech. New York: Harcourt, Brace & Co.

Spike, Matthew. 2019. Fifty shades of grue: Indeterminate categories and induction in and out of the language sciences.


OpenEdition suggests that you cite this post as follows:
Martin Haspelmath (December 11, 2019). The General Category Fallacy: Why grammatical category-assignment does not give us more insights. Diversity Linguistics Comment. Retrieved July 19, 2024 from

3 thoughts on “The General Category Fallacy: Why grammatical category-assignment does not give us more insights

  1. Thanks for writing this Martin, I think I understand your position a little better now! I have a number of observations, but for now I’ll stick to the General Category Fallacy.

    I’ll start with my attempt to reconstruct your argument (so please tell me if I’m wrong!):

    1) There is a widespread assumption (in some quarters) that (at least some) linguistic categories in particular languages are tokens of general cross linguistic types. For example, the things which we call nouns in Swahili and Mandarin and Inuktitut are all, despite all their many grammatical differences, real instances of a more general cross-linguistic type, i.e. noun.

    2) If general cross-linguistic types were real/natural and useful, then a claim that some category in a particular language (e.g. nouns in Inuktitut) is a token of a general type (e.g. the cross-linguistic noun) allows for *deductive* inference: “everything we know about the g-category would necessarily be true of the p-category”.

    3) However:
    i) There are no features of any general categories which are necessarily true of all language particular categories: e.g. a noun in any given language is not guaranteed to have any particular feature.
    ii) Almost all (or all) language-particular categories assigned to the same cross-linguistic-category have a different set of features: e.g. there are no nouns with the same features in any two languages.
    iii) Assigning a category in a particular language to a cross-linguistic category is very often quite difficult, because it’s not clear where the *boundaries* lie between any two cross-linguistic categories: e.g., a noun in language X can have some features which fall under the cross-linguistic category noun, but also some which look like the cross-linguistic category adjective.

    4) Due to all of the problems outlined in (3), then (2) cannot be true: “thinking that we learn something about a p-category by classifying it as an instance of a g-category is committing the General Category Fallacy.” And because (2) does not hold, we can infer that neither does (1): language-particular categories are not tokens of language-general types, and calling something is a noun in one language doesn’t allow us to make any valid inferences about what nouns do in that language, or in fact to learn anything else at all about that language.

    Now, although this is an internally consistent argument, it’s clear to me now that my main disagreement is with your second premise, i.e. that if cross-linguistic categories were natural/real, then we would be able to use them to make *deductive inferences*, i.e. that, for example, calling something a noun *logically entails* some essential feature shared by all nouns.

    My point is that this kind of deductive inference is vanishingly rare in science outside of fundamental physics (and some chemistry). As I attempted to outline in my paper, there are no essential features – and no natural kinds of the type you seem to be looking for – in biology, geology, meteorology, and everywhere else. Most scientific categories have the kind of problems we see in (3). Take your example of tuberculosis, which is actually a complex of genetically related bacteria (and the genetics of bacteria are paricularly wild, with rapid mutation and horizontal gene transfer leading to very high levels of variation and intermediate forms).

    However, this kind of complexity and indeterminacy has not made theoretical biologists throw in the towel: the primary mode of scientific inference is not deduction, but *induction*, where we have to work with messy, indeterminate, and inconclusive statistical phenomena, which allow us to draw defeasible, probabilistic, subjective conclusions: all of which is completely in line with linguistic typology and your proposals for comparative linguistics.

    Furthermore, the problems of indeterminacy you describe, for example the arbitrariness of defining a mountain, are only problems if you accept Plato’s line that we should ‘carve nature at its joints’. But that just doesn’t seem to be the way that most complex phenomena in nature work. And the existence of indeterminate boundaries does not imply that there are not real differences to be found. This is the continuum fallacy, or the ‘fallacy of the beard’: Mount Everest is definitely a mountain, and the Eurasian Steppe really isn’t.

    So if we take our cue from the biologists, there is no reason to keep a firewall between cross-linguistic phenomena and cross-linguistic phenomena: that would be like telling epedemiologists to stop tracking tuberculosis, but only particular strains of mycobacterium tuberculosis, and mycrobial pathologists not to look for features of tuberculosis found in other strains. Or telling geologists that they should only work on a case-by-case basis, one mountain at a time.

    At least in principle, then, we should expect an interplay between work on specific languages and work on linguistic comparison. It’s certain that lots of this work will end up being partially mistaken or flat-out wrong, but that won’t be the direct result of messy categories. It will be due to a lack of empirical support and theoretical or predictive success. And in any case, arguably the only way to know what the good categories are is via this kind of theoretical refinement, rather than deciding which categories count as natural or real – or not – in advance. But I say most of this in the paper, so I’ll link to it here again:

    • Many thanks for this detailed comment, Matt! As I said in the post, there’s no disagreement on how to do comparative linguistics, and your claim that deductive inference is not how science normally works only strengthens the main point that I have been making (since my 2007 and 2010 papers): That the identification of a language-particular category with a cross-linguistic type does not give us any insights about particular languages. I have been trying to understand why many other people think that this should work, and my proposal was that if you (maybe unrealistically) assume that linguistics is like (simplified, essentialist) chemistry, then that approach would actually work: The general category presumption would NOT be a fallacy. So I have been trying to give generative grammar the benefit of the doubt, while you are forced to be even more radical in your rejection of that approach (which is becoming more and more prominent in linguistics, despite Chomsky’s loss of interest in a rich grammar blueprint). Yes, indeed, we should “work with messy, indeterminate, and inconclusive statistical phenomena” in linguistics as well, but for our results to be comparable across the discipline, we need precise measurements – universally applicable yardsticks that work in the same way for all languages. This is surely also what biologists or geologists do – they do not define “tuberculosis” or “mountain” differently in different contexts. (I’m still not sure whether a biological concept like “mycobacterium tuberculosis” is indeed identified via a *definition* in the same way as “mountain” is defined, because it seems to me that we identify biological species via their phylogenies, but this is not so important in the present context.)

      • A belated comment, because I wanted to record a recent experience on Facebook, which is quite representative of what I hear from generative linguists all the time. Omer Preminger said about “clitics” (in reply to my question how clitics and affixes are distinguished):
        “There is a terminological confusion here. The term “clitic” is used in at least two ways: one, the more traditional, refers to a phonologically/prosodically-dependent piece of phonological material that is unselective with respect to the identity of its host. The other refers to a particular kind of morphosyntactic element. You could argue that we, as a field, should have (by now) come up with a new term for the second use, which is probably true, but I find arguments about terminology to be the most boring kind of arguments. So let’s just refer to these as P-clitics and MS-clitics, respectively. P-clitics are not necessarily MS-clitics, and MS-clitics are not necessarily P-clitics. (MS-clitics probably never bear primary stress, but some can only ever appear next to a particular host, and that doesn’t make them any less of an MS-clitic.) I’m not enough of an expert in phonology and prosody to know what the state of the art is on the answer to Martin’s question as it concerns P-clitics (can they be identified in an objective way). But as it regards MS-clitics, the answer is very much “yes”: see work by Ruth Kramer, Boris Harizanov, Karlos Arregi & Andrew Nevins, and myself. These works are very theory-laden, i.e., the set of diagnostics for MS-clitic are not theory-neutral. This, to repeat, is a good thing, not a bad thing – at least if one is trying to do science. There is also not a theory-neutral diagnostic for “neutrino” in physics. Two important caveats: 1. The whole Zwicky & Pullum edifice is about P-clitics, and whether it still holds up as a theory of that, it is a nonstarter as a theory of MS-clitics. (

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.