Himmelmann (2019) (finally available on Lingbuzz in January 2021) criticizes my 2010 and 2018 papers on comparison and description, and it might seem that he is advocating a different approach, but (as is easy to see when comparing his work and my work) we actually agree on what is the best approach for descriptive and comparative linguistics – as do other authors such as Dahl, van der Auwera, Dryer, Bickel, Moravcsik and Lazard (whom Himmelmann also mentions). The real disagreement is with generative grammarians, who have long argued that the “shallow” approach to comparison as practiced by the “typologists” is not sufficient and must ultimately be replaced by an approach based on “deeper” analyses (e.g. Coopmans 1984; Baker 2011). Again, this can be easily seen in the practice of generative grammarians, which is totally different from that of non-generative comparativists like Himmelmann and myself.
What explains this different practice? I have long been wondering about this (cf. Haspelmath 2000), and since 2010, I have argued that the difference is the generativists’ reliance on an innate grammar blueprint: a set of universally available building blocks that are assumed to be given in advance (pre-established, Haspelmath 2007), that all grammatical systems ultimately consist of, and that can therefore also serve as the basis for comparison of grammars. Non-generative linguists are taking a more conservative stance and are skeptical about this ambitious approach (which I have recently come to call the “Mendeleyevian vision”, because it envisions linguistics as similar to chemistry, with the ultimate goal of finding a kind of “periodic table of linguistic building blocks”, as in Baker 2001). This conservative stance entails that comparison cannot happen on the basis of pre-established building blocks – it must involve comparative concepts (Bickel’s 2007 “fine-grained variables”).
Himmelmann’s own view is clearly on the conservative side as well, so why does he appear to disagree with me? And why are there other conservative (= non-generative) comparativists who likewise do not find the distinction between comparative concepts and language-particular categories so important? I don’t really know, but it may have to do with the experience that languages are generally very similar (more so than one might have thought), and with the feeling that language-particular categories are often quite well characterized in terms of substantive concepts – for example, English Questions are mostly satisfactorily characterized as a sentence type designed for questions, or Russian Plural forms have an evident close connection with semantic plurality. All this is true, but at the same time, languages may also show strikingly unique structural properties (Haspelmath 2020), and language-particular categories may show unexpected behaviour (e.g. English Questions may be used in requests, even when a request marker like “please” occurs; and the Russian Plural category may include semantically singular nouns like časy ‘clock’). Thus, while as a first rough approximation, we can describe salient aspects of languages in terms of cross-linguistically applicable comparative concepts, we often need language-particular descriptive categories when the description goes deeper. This was the basic insight of the structuralists since Boas (1911) and de Saussure (1916), and its correctness has never been in doubt (Himmelmann does not doubt it).
So the question is not: Are descriptive categories ALWAYS different from cross-linguistically applicable concepts? but rather: Are they EVER different? And if so, this means that one must draw a conceptual distinction (not a “sharp distinction” – simply a distinction). This is what the generativists fail to do (because the blueprint assumption makes it unnecessary), and non-generative comparativists sometimes inadvertently fail to draw the distinction as well (e.g. Dixon 2004 treats “adjective” as a pre-established category, even though in his famous 1977 paper, he had noted striking differences between languages). Himmelmann spends a lot of time arguing that there may be overlaps between cross-linguistic concepts and descriptively useful concepts, something that nobody ever denied (and since Beck 2016, the term “portable terms” has been used for this overlapping area). So there is no disagreement.
It is also clear that describing a new language becomes easier in practice if one already knows about other languages, because languages are often similar, and distinctions that have proved useful for one language may also be useful for another language (e.g. animacy for object marking, or mass/count distinctions for classifier constructions, or habitual action for verbal inflection). So no doubt, comparative knowledge is “useful” and can make descriptions “better”, at the practical level – it can inspire describers, and allow them to come up with more transparent descriptions (Haspelmath 2020). (Perhaps the best account of how comparative knowledge has improved descriptions is Dryer 2006).
But what if there are two competing proposals for describing the same phenomenon? For example, should Russian nominals be described in terms of a “noun phrase” (NP) or in terms of a “determiner phrase” (DP)? Which of the two descriptions is “better”, and by what criteria (cf. Ljutikova 2018)? Generativists usually opt for a description that is motivated by very general, including cross-linguistic, considerations. More conservative grammarians would tend to describe Russian “in its own terms”, i.e. not with a “D” category, because Russian has no articles that occur in the same slot as the demonstratives (this was Bloomfield’s original motivation for proposing the “determiner” concept for English). It is this undisputed situation that I am characterizing by saying that unless one relies on an innate grammar blueprint, one relies on language-particular distributions. “Cross-linguistic properties” of language-particular constructions are useful to know, but they are not the basis for identifying language-particular categories.
Himmelmann is right that the practice of establishing language-particular distributional classes may involve all kinds of factors (especially meaning, when discussing syntactic patterns), but again, there is a principled distinction: Such classes MAY make reference to language-particular entities, so distributional classes are distinct from cross-linguistic concepts. English Questions are defined by a particular English-specific word order (or at least a particular intonation pattern), NOT by their question function. Russian Plural forms are defined by their Russian-specific inflectional pattern (e.g. Nominative/Genitive stol-y/stol-ov ‘tables’, časy/ časov ‘clock’), NOT by their “plural” meaning. This has never been in doubt, so again, there is no disagreement. Himmelmann’s example of Pronominal Articles in Wooi (a language of New Guinea) demonstrates the usefulness of cross-linguistic knowledge for rich descriptions, but the establishment of the category in the first place was based on language-particular distributional information (e.g. occurrence in the “NP-final slot”, after the Wooi-specific “Determiner” slot). Cross-linguistic information is highly useful in practice, but not crucial in principle – and the difference between these two emerges clearly when there is some doubt about the correct description. Cross-linguistic information will never be decisive in discussions of competing descriptions (unless one relies on an innate blueprint).
Himmelmann is also right that “crosslinguistic comparison does not necessarily lead to unduly applying categories and analyses of one language to another” – not necessarily, but often this does happen, unfortunately (especially among authors who rely on an innate grammar blueprint, but not only). Describing Russian as having a DP, or Chamorro as having a verb-adjective distinction (Chung 2012; cf. Haspelmath 2012, and this earlier blogpost) would not occur to authors who do not get distracted by some salient similarities with other (more prestigious) languages. The similarities between languages (and cultures) have of course always been known, but it was only in the 20th century that linguists and anthropologists focused on the uniqueness of each culture and each language. Boas and his followers criticized earlier anthropologists and linguists for seeing other cultures and languages through the Western prism, often as occupying a “more primitive place” on a developmental ladder. They rightly insisted that each culture and language had its own components and interactions, worth studying even without a necessary comparative component. Of course, comparison enriches our understanding of the bigger picture, but it is not strictly necessary for description. Hagman (1977) described Khoekhoe Pronominal Articles without being inspired by Wooi, and Topping (1973) described Chamorro word classes without needing guidance from English. A structural definition of “German Adjective” will NOT make reference to the property-concept meanings of items in this class, but to their inflectional and syntactic distributional options (again, Himmelmann is right that any German grammar will note that German Adjectives generally refer to properties and that properties are typically expressed by Adjectives; but this is not part of the definition). Thus, good grammatical descriptions may in practice come about through knowledge of other languages (and thus be “typologically informed”), but the ultimate criterion is the correctness of the description (in that it comes up with the right distributional classes), not its cross-linguistic naturalness.
And again, Himmelmann is right that “in order for falsification to be possible, there has to be an operationalizable link between descriptive categories and comparative concepts”, and anyone who has engaged in large-scale comparative project knows that this is not easy. For example, when trying to compare something like “plural forms” across languages (e.g. Chen 2019), one encounters variation of the following sort: (i) some languages have dual, trial and paucal forms – should these be included? (ii) some languages have “plural words” rather than morphological plural forms (iii) some languages (like Welsh) seem to have “basic plural forms” from which singulative forms are derived. So how do we operationalize this? There are a number of ways of doing it (and all of them could be called “plural” in a comparative study, so that the term “plural” ends up being linguist-specific, or study-specific), but crucially, none of these can refer to language-specific distributions. (In fact, comparisons need not even refer to language-particular descriptions but can make use of parallel texts or reactions to stimuli, as Himmelmann also agrees.) Comparative concepts are like questions asked by an empirical sociologist: For a cross-linguistic study, each language must answer the question (based on the same understanding of the question in each case), but the answer need not make reference to the language-particular category. It is sufficient if it is clear what the phenomena are (cf. Haspelmath 2018: §10). As Himmelmann says: “The obvious answer is that language-specific constructions have crosslinguistic properties, as argued here and by many other authors on the topic.” But as we saw, these cross-linguistic properties are not always crucial for delimiting the language-particular categories. Thus, one reasonable definition of “plural” for comparative purposes would be “a special form that can denote more than three items”. The English Plural matches this comparative concept, even though it is defined/delimited by the condition “more than one”. But if we want to compare English Plural to similar forms in languages with a dual and a trial, we probably want to exclude those forms, so that the comparative concept cannot be “more than one”.
Himmelmann construes my view as somehow different from Lazard’s (2002), but this is not right. Comparative concepts are “arbitrary” (I took this term from Lazard) only in the sense that they are not a research result, but an instrument for research, like a yardstick or other unit of measurement (I have therefore also referred to the non-generative approach as requiring “measurement uniformity”). They are obviously not arbitrary in the sense that “anything goes” (let alone “alternative truths”). Anyone who knows comparative work in the tradition of Greenberg, Hawkins, Croft and Dryer knows that these authors are using a standard scientific approach. As Lazard notes (in the passage of his 2002 paper quoted by Himmelmann), comparative concepts prove particularly useful if they help us find an “invariant or universal”. But they can also be useful in other ways, e.g. simply for describing a salient type of difference between languages in a clear way.
Thus, Himmelmann and I are using exactly the same approach to describing and comparing languages, which contrasts with the generative approach that relies on a hypothesized innate grammar blueprint. Where Himmelmann sees conceptual differences, there are in fact merely differences of terminology, or perhaps differences of emphasis. If we disregard the Chomskyan school, the scientific field of language description and comparison is not in any way divided by fundamental differences.