General linguistics must be based on universals (or non-conventional aspects of language)

4.1 First solution: non-conventional aspects of human language

The problems that I will highlight in §5 below do not concern those areas of linguistics where non-conventional behaviours are studied. In particular, general inferences from reaction times in psycholinguistic experiments, from electrophysiological experiments in neuroscience, or from slips of the tongue are not problematic because the aspects of behaviour that are of interest here are not socially learned, i.e. are not conventional. Of course, the differences between languages still put limitations on general inferences, and there is now an increasing awareness of the need for cross-linguistic research in psycholinguistics as well (e.g. Bak 2016; Bornkessel-Schlesewsky and Schlesewsky 2009; Jaeger and Norcliffe 2009; Slobin 1985). But even without such awareness, psycholinguistic research can lead us to some general conclusions.

Another non-conventional aspect of language is learnability. It has often been argued that some aspects of grammars cannot be learned on the basis of positive evidence and must therefore be present at birth (this is called the argument from the poverty of the stimulus, e.g. Lasnik and Lidz 2016). For example, it has been claimed that principle C of Chomsky’s (1981) binding theory for English cannot be acquired on the basis of positive evidence (i.e. the rule that a pronoun in subject position cannot have the same reference as a full nominal in other positions such as non-subject position, e.g. *She _i thought much about Maria _i ). Such learnability considerations do not depend on historically accidental conventions of particular languages and are therefore a valid solution to the general linguistics paradox.^[15] However, in practice they do not play an important role in linguistics – there are very few specific proposals about characteristics of languages that must be innate because they cannot be acquired without innate prerequisites.

A lot more could be said about non-conventional aspects of linguistic behaviour, but the focus of this paper is on the problems that arise when one limits one’s attention to language structures, or linguistic conventions (see also §6.1 below on what I mean by conventions). This is why the non-conventional aspects of language are in parentheses in the title of this paper.

4.2 Second solution: general linguistics through worldwide language comparison

An obvious way in which one could learn about Human Language in general is by comparing languages worldwide and by finding general tendencies in representative samples of these languages. This is the well-known Greenbergian approach (Croft 2003; Greenberg 1963; Song 2018), and I do not need to say much more about it here, except perhaps for reminding the reader that Greenbergian comparison does not rely on mental grammars, but can be based entirely on the social rules (cf. (2a) and (3a) above). Thus, there is no need for a “deep” or “true” analysis, but basically any description that gets the facts right (that is observationally adequate) is sufficient as a basis for comparison (cf. Haspelmath 2004, 2014). The comparative concepts on which the comparisons are based are not the same as the descriptive categories used for describing the languages, so that different languages can be approached non-aprioristically, i.e. with no preconceived idea of what the possible categories might be (Haspelmath 2010b; see also §6.3 below).

Explanation in the Greenbergian programme is typically of the functional-adaptive type: Cross-linguistic tendencies are explained with reference to the adaptation of language structures to their users’ needs. For example, special reflexive pronouns tend to be used when coreference is unexpected to help the hearers establish the correct reference (Comrie 1999; Haspelmath 2008), and word order tends to be consistently right-branching or left-branching because this minimizes constituent recognition domains (Hawkins 2014).^[16]

4.3 Third solution: general linguistics based on a natural-kinds programme for the building blocks of languages

Finally, we get to a third solution to the general linguistics paradox: We could learn about Human Language by finding the innate building blocks from which all the mentally represented grammars are constructed. This is the principles and parameters programme of Chomsky (1981), most clearly laid out (for a wider audience) by Baker (2001) (see also Baker 2010; Huang and Roberts 2016; and many others). D’Alessandro (2019: 10) observes that if there are such innate building blocks, then

the constraints that are discovered about one language could be used to describe a different language. This, I think, is the key difference between generative grammar and other linguistic enterprises, such as typology: while typologists assume that, say, the existence of wh-movement in English cannot tell us anything about Chinese, generativists assume that this isn’t the case.

The “innate building blocks” include both architectures (e.g. the distinction between D-structure and S-structure, between syntactic computation and spell-out, between Merge and Agree, between level I and level II, between Gen and Eval in OT) and innate features, categories and constraints (e.g. ±N, ±V, vP, CP, SUBJ, OBJ, +WH, ±coronal, NoCoda). There is a very rich literature with proposals about the kinds of entities that might be innate building blocks of this kind, which are thought to be part of universal grammar. Chomsky (1965) called the architectural building blocks “formal universals”, and the features and categories “substantive universals” (there were no OT constraints in the 1960s, but these would surely also fall under substantive universals).

As noted by Baker (2001), this research programme is not unlike the programme pursued for chemistry in the 19th century that resulted in Mendeleyev’s Periodic Table of Elements. Theoretical chemists found that all chemical compounds are built from about 80 to 100 building blocks, and that there are a limited number of ways in which these can combine to form chemical compounds.

If linguists found that there are just a few dozen (or maybe a few hundred) features or categories (or OT constraints) that recur across languages and from which more complex structures can be constructed in a limited number of ways, then it would indeed be plausible to attribute them to the innate universal grammar, and they might also solve the problem of language acquisition despite the poverty of the stimulus (called “Plato’s Problem” by Chomsky 1986).

I call this the natural-kinds programme (or naturalistic programme) for general linguistics,^[17] because chemical elements are the primary example of what are called natural kinds in philosophy.^[18] Natural kinds are categories or classes of entities that exist in nature independently of any scientific observation. The categories are given in advance of observations, and are thus available a priori. In addition to chemical elements, examples of natural kinds are particles in physics, and (more controversially) species in biology. Clearly, some aspects of human behaviour and cognition are given by nature, e.g. the fact that we can distinguish five basic tastes (sweet, sour, salty, bitter, umami), and maybe that there are six basic emotions (anger, disgust, fear, happiness, sadness, surprise; see Barrett 2006). So it is readily conceivable that our biology might give us four basic parts of speech (e.g. noun, verb, adjective, preposition), or three components of grammar (e.g. phonology, morphology, syntax), or 27 distinctive features for spoken phonology, or 65 semantic primitives (see also Aronoff 2016). There could also be hundreds of different optimality-theoretic constraints, just as there are hundreds of different cell types in the human body. For phonology, linguists are fairly close to a consensus of what the innate features might be, as can be seen in any phonology textbook (though even in phonology, there are dissenting voices of linguists who do not think that the features are universal and innate, e.g. Mielke 2008). In morphology and syntax, there is no such consensus, but it could still be that the search will ultimately be successful.

However, there is a serious problem: In contrast to the Greenbergian programme for cross-linguistic comparison, the natural-kinds programme has no clear methodology for progressing, and no clear criteria for success.^[19] There are a large number of new proposals about the building blocks of the innate grammar blueprint, but there is little (if any) convergence among them. There is no agreement about serial versus parallel architectures, lexicalism, DP versus NP, antisymmetry, phases, cartography, and many other core aspects of grammar. Those new ideas and generalizations that have been widely accepted belong to the level of phenomena (D’Alessandro’s 2019 “mid-level generalizations”), not the explanatory level of innate natural kinds. And of course, the existence of a large number of domain-specific innate elements in just one species is inherently unlikely, given the relative recency of our capacity for language (perhaps only 200,000 years old). Moreover, in the 21st century, the natural-kinds programme has basically been given up by some influential authors, as I will briefly note in §5.1.

This leads me to the last section of this paper – the question whether it is possible to do general linguistics without large-scale cross-linguistic comparison and without a natural-kinds programme. I will end up with a very skeptical answer, resulting in the claim of the title of this paper: General linguistics must be based on universals.

5.1 The 21st century shift

What I described as the natural-kinds programme (§4.3) was developed in the last few decades of the 20th century, and it was widely adopted, also by linguists who did not agree with Chomsky’s particular proposals for syntactic building blocks in the 1970s and 1980s. Syntactic frameworks such as Relational Grammar (Blake 1990) and Lexical-Functional Grammar (Bresnan 2001) saw themselves as making use of universal innate building blocks as well, and the situation in phonology and morphology has been rather similar. A widespread view has been that linguists should try to find a descriptive framework that allows us to describe all languages, but is at the same time restrictive enough to explain language acquisition and the limits on worldwide linguistic diversity (this describes what I called the restrictivist approach in Haspelmath 2014).

But since Hauser et al. (2002) and Chomsky (2005), this programme has been basically abandoned by many linguists as an explicit goal. The 21st century Chomskyan idea is that the innate universal grammar contains only rather minimal building blocks, perhaps only the operation Merge, which explains that languages can have recursion. In Chomsky et al. (2019: 230), the authors even say that “universal grammar” is merely the label for whatever biocognitive differences there are between humans and other animals. One of the motivations for this much less ambitious view of what is innate seems to be that a rich set of innate building blocks has come to be seen as implausible from the perspective of biological evolution (and thus the abandonment of a rich UG would help solve “Darwin’s Problem”; cf. Berwick and Chomsky 2016).

So on this view of Human Language, there is no substantial natural-kinds programme anymore. The general theory is no longer restrictive and cannot explain observed limits on cross-linguistic variation, because with Merge alone, a large number of unattested grammars are possible. Now what does this mean for the practice of the linguist who studies the structures of particular languages? It would seem to me that a necessary consequence is that descriptions of particular languages are no longer carried out in terms of a set of universal building blocks – in other words, description can and should be non-aprioristic, or framework-free (Haspelmath 2010a; see §5.3).^[20] But this is not what has generally happened, as I note in the next subsection.

5.2 The implicit continuation of the natural-kinds programme

Even though the idea of a rich innate blueprint for grammars was given up by influential authors (§5.1), this is not reflected in the practice of mainstream generative grammar. Journals such as Linguistic Inquiry, Syntax and Glossa continue to publish many papers on particular languages that work with a highly technical metalanguage for describing/analysing the morphosyntactic phenomena of languages. It is quite common for research articles to consist of two parts: One part lays out the phenomena in a way that is generally comprehensible to any linguist, and another part (typically called “analysis”) describes the phenomena a second time, using the highly technical metalanguage of current mainstream generative grammar (or more rarely, of some other generative approach, such as Distributed Morphology or Lexical Functional Grammar).

For example, Welch (2016) first describes various conditions for the use of copulas in Dogrib (a Dene language) in a generally accessible way, and then in his §6 (“Analysis”) describes the same facts using technical generative vocabulary such as “merge”, “AspP”, and “φ-agreement”. And Holmberg et al. (2019) describe a generalization about the interaction of question formation and passivization in ditransitives in some Germanic and some Bantu languages, and then in their §3 (“Analysis”) describe the same facts again using technical vocabulary such as “phase”, “specifier” and “ApplP”. Anyone who has a certain amount of experience in this field will confirm that this is very typical: Studies of particular languages make use of highly specific concepts that are thought to be universally applicable.

How does this approach contribute to general linguistics? Clearly, these authors build on the assumption of a rich innate grammar blueprint, because otherwise there would be no reason to have the “analysis” section in addition to the generally comprehensible description of the phenomena in the first part of their papers. The “analysis” adds something, because it redescribes the same facts by means of a universal vocabulary of innate building blocks. This is not redundant because it embodies the claim that these descriptions capture the mental grammars of the speakers, and these mental grammars are instantiations of the innate grammar blueprint. By studying a wide range of mental grammars in this way, we can hope to get closer to the true innate building blocks that underlie all human languages and thus get insight into Human Language. This is formulated by Huang and Roberts (2016: §14.2) as follows:

The P&P model is a very powerful model of both linguistic diversity and language universals. More specifically, it provides a solution to Plato’s Problem, the logical problem of language acquisition, in that the otherwise formidable task of language acquisition is reduced to a matter of parameter-setting. Moreover, it makes predictions about language typology: parameters make predictions about (possible) language types, as we will see in more detail [below]. Furthermore, it sets the agenda for research on language change, in that syntactic change can be seen as parameter change … Finally, it draws research on different languages together as part of a general enterprise of discovering the precise nature of UG: we can discover properties of the English grammatical system (a particular set of parameter values) by investigating Chinese (or any other language), without knowing a word of English at all (and vice versa, of course).

So despite the fact that the natural-kinds programme has generated few generally accepted results (as noted by authors such as Newmeyer 2005 and Boeckx 2014, and as admitted by Baker 2008), these authors continue to pursue this idea, and they are not alone.

Regardless of how promising it may be to look for innate building blocks, this is a coherent position (cf. also Dryer 2016: 314). But many other generative linguists are adopting what appears to be an incoherent stance: They work with the technical vocabulary and rule notation of mainstream generative grammar which is supposed to apply to all languages, but at the same time, they do not endorse the natural-kinds programme of an innate grammar blueprint. This view is articulated, for example, by Koeneman and Zeijlstra in a blogpost (https://dlc.hypotheses.org/1082), where they explain the choices made in their (2017) syntax textbook:

although we build up the theory using technicalities that are adopted from current minimalism, we do not adhere to or to try to persuade students about most of its philosophical or biological underpinnings, such as innateness claims or conjectures about the biological function of language

It seems that they think of their complex formal apparatus (which can account for only a very small part of English grammar) as somehow merely a notation, but they do not discuss alternative notations, of which there are many different ones that are much simpler and would be able to describe a much larger part of English. So one can make sense of what they actually do only if they adopt the traditional (20th century) view of the notation of generative grammar as a theoretical claim about a rich innate grammar blueprint.^[21]

Even though many linguists have nominally abandoned the idea of an innate grammar blueprint, it apparently got entrenched in linguistics to such an extent that we do not think much about it. We propose categories, architectures and formalisms that can be applied to all languages without thinking much about the reason for their possible generality. As a final example, let us consider the following opening paragraph of a paper about Russian Nominalization constructions (Smirnova and Jackendoff 2017: 877):

Despite the fact that the analysis of nominals has for a long time received a certain amount of attention (e.g. Chomsky 1970; Grimshaw 1990), argument realization and case assignment in the nominal domain have been primarily viewed through the lens of verbal syntax. Here we analyze nominals on their own terms, proposing a lexicalist, constraint-based approach to case assignment in Russian nominals, couched within the simpler syntax framework (Culicover and Jackendoff 2005).

Smirnova and Jackendoff only discuss Russian, and they simply presuppose that a discussion of Russian Nominalizations must be relevant to the English phenomena discussed earlier by Chomsky and Grimshaw, which is clearly the case only if “nominalization” is somehow part of the innate grammar blueprint. If it were not, then it could be that Russian is entirely irrelevant to understanding English. Or it could be that it is only historically relevant, because the similarities are due to a shared history (either inheritance or borrowing via French). Since there is no consensus about these matters in the discipline, the authors would have to spell out their assumptions in terms of natural kinds in order to be understood also by linguists who do not share these assumptions. The reviewers of the journal apparently did not deem this necessary, apparently because the traditional (20th century) generative position is still an implicit default position.^[22]

So the answer to the question at the beginning of this section is: General linguistics can be based on a single language if one adopts the natural-kinds programme, but not otherwise. The “Mendeleyevian” or “Bakerian” approach (§4.3) thus remains the only coherent way in which one can do general linguistics if one wants to adopt a universal framework. But let us now see how the framework-free approach solves its own challenges in view of the general linguistics paradox.

5.3 Framework-free descriptive linguistics

While generative linguists who study a single language cannot contribute to general linguistics without subscribing to a natural-kinds programme of some kind, the situation is quite different for linguists who are happy to describe each language in its own terms, in the Boasian tradition. They need not endorse an inherently unlikely proposal about an innate grammar blueprint, but their work does not automatically contribute to general linguistics. Their work is particular-theoretical, not general-theoretical (cf. §2). As noted by D’Alessandro (2019) (in the quotation in §4.3 above), a Boasian/Greenbergian linguist cannot use phenomena from one language for the description of another.

Perhaps surprisingly, this mode of language description has been advocated by a number of prominent typologists (Cristofaro 2009; Croft 2001; Dryer 1997; Haspelmath 2018; Lazard 2005; Levinson and Evans 2010), but how is it compatible with cross-linguistic comparison? How can it be reconciled with the idea that language descriptions should contribute to an understanding of Human Language through the discovery of universals? After all, experience shows that grammatical descriptions are easier to understand and more relevant to general readers if they are written with the needs of comparativists in mind. Evans and Dench (2006: 5) put it like this:

Describing each language entirely on its own terms is a noble and galvanizing task, but unless grammarians orient their findings to what typologists know about the world’s other languages, their grammars can all too easily become obscure, crabbed and solipsistic.

The tension between a kind of description that is faithful to the categories of each language and a description that makes the language appear relevant to general concerns is just a special case of the “general linguistics paradox” that we saw earlier in (5). The solution must consist of two parts.

First, language description is true to the categories of each language, but is inspired by the accumulated knowledge of comparative linguistics (Haspelmath 2020b: §4). So if there is an affix on the verb that is very similar to an applicative or an evidential, it should be given this label, rather than some other idiosyncratic label. But at the same time, it cannot be assumed that what we know about other languages (or about Human Language) determines the language-particular analysis. Language-particular analysis is determined by the facts of each language, not by other languages or by a general theory.

Second, comparison of languages is not expected to be based on the categories of the particular languages, because there is no uniform set of building blocks of which all grammatical systems are composed. Thus, the building blocks themselves cannot be the basis of comparison, but we must use independent yardsticks for comparative “measurement” (cf. Haspelmath 2019: §8 on measurement uniformity vs. building-block uniformity). I will elaborate on this in §6.4.

Thus, grammatical descriptions can contribute to general linguistics in an indirect way: By providing crucial data for large-scale cross-linguistic comparisons, which in turn can tell us about Human Language if we find strong cross-linguistic universals (as noted in §4.2).

6.1 What exactly are conventional aspects of language?

One colleague objected that it is odd to say that comparative grammar in the Greenbergian tradition is based on “social conventions”, and that reaction time experiments involve “non-conventional aspects of language”. And indeed, linguists do not often talk like this, because there is a strong mentalist bias in the field. We talk about acceptability judgements as if they involved “introspection” into our mental grammars, whereas what we actually do is assess the social acceptability of a possible sentence (in terms of linguistic norms). While it is of course true that our knowledge of the social conventions must be mentally represented, the conventions themselves are “social facts”, and when we describe a language, we describe it as a social rule system. This is particularly clear in the case of child multilingualism, where one of the tasks that the child faces is to link sets of linguistic conventions to sets of social situations. And since all languages have different registers, monolinguals are not in a very different situation. Thus, all grammatical regularities and all the words and meanings of a language are its conventional aspects. The two main sources of data for studying conventional aspects of language are corpora and elicitation (including self-elicitation). Until a few decades ago, virtually all of linguistics consisted in the study of linguistic conventions, or language structures.

In striking contrast to this, when we observe a child’s first attempts to talk, or an aphasic’s speech, or a slip of the tongue, we do not enter the results into a “child language dictionary”, set up an “aphasic’s transformational rules”, or posit a “phonemic error system”. Language description of the classical sort (resulting in grammars and dictionaries) is not relevant for non-conventional behaviour. And similarly, when we investigate aspects of ordinary language behaviour that plays no social role (such as reaction times, or event-related brain potentials), this is different from conventional behavioural patterns that derive from social learning and that may be considered socially appropriate or not.

These distinctions should be clear enough, but I elaborate on them here because linguists do not yet have terms for structural linguistics (the study of conventional patterns) and non-structural linguistics (the study of other kinds of linguistic behaviours) that are widely understood. It also needs to be noted that the explanatory factors that are invoked are not necessarily conventional aspects of language. The innate building blocks of traditional generative linguistics are not conventional, but they are claimed to constrain the kinds of conventions that people can acquire. In the non-aprioristic approach, each language has its own building blocks, and these elements are themselves conventional.

6.2 Aren’t there analogous “paradoxes” in other disciplines?

Regarding the “general linguistics paradox” of §4, one colleague objected that they do not see “how this is any more of a paradox than scientists in other fields face”. For example, biologists want to explore and understand the nature of bat echolocation or crab locomotion, and what they can observe directly is only particular bats or crabs.

But how is it that biologists can generalize from a single bat to all members of the species? If species are natural kinds (as I said in Haspelmath 2018: 90, admittedly simplifying matters), then the answer is simple: By their nature, all bats and all crabs have the same essential properties, so by studying one specimen, one learns about the entire species. This view of biological species is the basis of taxonomists’ practice of attaching scientific names to “types”, i.e. specific specimens kept in a research collection.^[23] This approach is possible because the properties of different specimens do not vary by historical accident the way that the observable properties of languages vary. The same goes for other disciplines that work with natural kinds such as chemistry and physics, but there are also natural sciences that study their phenomena using comparative concepts just like linguistics (e.g. the study of clouds in meteorology, or the study of topographic features in geomorphology).

6.3 Why must comparative concepts be different from descriptive categories?

While language structures are often similar, each language is structurally unique (e.g. Haspelmath 2020b). Phoneme systems carve up the phonetic space in different ways, and semantic systems are often different even in closely related languages (cf. German fahren vs. gehen, which work quite differently from English drive vs. go). Syntactic classes are often strikingly different, as can be seen in the very different behaviour of English auxiliaries and German auxiliaries, in the very different behaviour of Polish person clitics and French person clitics, and the very different classifications of Arabic gender (masculine vs. feminine) and Swedish gender (neuter vs. non-neuter). Perhaps most notoriously, what is meant by the syntactic term “subject” differs from language to language in confusing ways. In English, even an expletive like there can behave as a “subject” (cf. I believe there to be two unicorns in the garage), and in Icelandic, a “subject” can be in Dative case, while this is never so with Latin “subjects”. There are usually enough similarities between different languages to make it tempting (and in some sense useful) to reuse the same terms (e.g. “auxiliary” both for English modal auxiliaries and for German tense auxiliaries), but the categories are really defined by their language-particular structural behaviour (e.g. the lack of non-finite forms of English modal auxiliaries), not by instantiating some general (aprioristic) category.

These differences mean that language systems are incommensurable, so that making them comparable requires extra effort. In most cases, we cannot simply translate from one language to the next by substituting different morphs. The most straightforward way of making languages comparable is by creating comparative concepts to which the structural elements of each language are mapped. For the case of fahren/gehen/drive/go, we can start with the comparative concepts ‘go’, ‘go by car’, ‘go by vehicle’, and ‘go by foot’. The simple equation of German fahren with English drive fails because the latter means ‘go by car’, whereas fahren is also used for going by bicycle or by train. And while gehen seems to correspond to go, it can actually only be used for going on foot. So while there are clear similarities here, we need extra concepts to describe how the languages are similar or different. The same goes for phonological categories, where we need phonetics-based comparative concepts, and for syntactic categories, where our comparative concepts must be based on a combination of semantic and phonetic concepts. For example, an ergative case marker is a marker that occurs on a nominal expressing a transitive agent but not on a nominal expressing an intransitive agent (this must be based on a clear comparative definition of “(in)transitive”, see Haspelmath 2011).

So comparative concepts are necessarily based on phonetic and semantic substance that is independent of language-particular structural distinctions, while language-particular categories are based on contrasts, not on substance. It has long been recognized, for example, that phonemes are not defined by their phonetic properties, but by their place in the system of oppositions. But we still want to compare phoneme systems, which must be done in terms of the phonetic features of the inventories.

6.4 How do we decide whether a language-particular category matches a comparative concept?

One colleague asked how – if there are no innate categories – we can decide on what is a particle, or an adposition, or a subject, or a reflexive pronoun. And I agree with them that “this is the number one problem that needs to be addressed”. Indeed, if we want to compare language structures in the world’s languages and find universals, we need to have uniform “yardsticks” for comparison, analogous to measurement in other sciences (cf. Round and Corbett (2020) on the “measurement” metaphor). In Greenberg’s (1963) pioneering work, this issue was mostly left aside, and Greenberg limited himself to saying that he was employing semantic criteria in identifying phenomena like “subject”, “verb” and “genitive construction” (1963: 59).

The issue of cross-linguistic comparability was not widely discussed until Dryer (1997) and Croft (2001) pointed out that language-particular categories are defined by language-particular criteria and thus cannot be compared directly. Since Haspelmath (2010b), the idea of defining comparative terms carefully in such a way that they can be applied to all languages uniformly has been steadily gaining ground. If a category were innate and thus given a priori, it would be reasonable to think that it is identified by different criteria in different languages, but if not, we need to provide definitions that focus on phonetic and semantic substance and make no reference to language-particular structures.

For example, an adpossessive construction (Greenberg’s term was “genitive construction”) is defined as a nominal with a possessed noun and an adpossessor nominal or person index where the possessor referent is in an ownership, kinship, or part-whole relation to the referent of the possessed noun (e.g. Koptjevskaja-Tamm 2001). This fairly complex definition presupposes other terms that are themselves not straightforward (noun, nominal, person index) and that need to be defined carefully in turn. At each stage, it must be ensured that the definitions do not contain concepts that cannot be applied to all languages uniformly. For example, as noted in Haspelmath (2008: 43), the binding conditions of Chomsky (1981) cannot be tested in languages worldwide with a measurement approach because there are no generally applicable definitions of terms like “anaphor” and “pronominal”, and concepts like “c-command” presuppose specific analyses, which cannot be arrived at in an objective manner. For Chomsky’s approach at the time, this was not a problem, because he assumed that his notions were innate building blocks, so they did not have to be defined in such a way that the definition can be applied to all languages uniformly (and it becomes possible to use different criteria for different languages; Haspelmath 2018: §7). But for the Greenbergian approach, the requirements are different, because language comparison is based on uniform measurements of grammatical patterns (cf. 3a), not on true analyses of mental grammars (cf. 2b). The descriptive meta-language is not assumed to be an innate framework, so it cannot simultaneously serve for cross-linguistic comparison and for explanation (Bickel 2015: §2).

6.5 What is a “rich universal grammar”, and why is “richness” relevant?

In the earlier sections I have repeatedly talked about a “rich set of innate building blocks”, or a “rich UG”, but why should this be relevant here? Isn’t it an empirical question how rich our innate knowledge of grammar is?

The reason this is relevant is that de facto, many linguists assume a very rich set of innate building blocks, because each time they use a category that was motivated for one language for another one, this category must be innate. For example, as D’Alessandro (2019) notes (in the quotation of §4.3), wh-movement in English is thought to be informative for Chinese, because a notion such as “question pronoun”, as well as a notion such as “movement”, is thought to be innately given. And the same applies to many other building blocks (CP, ±coronal, anaphor, AUX, specifier, and so on) which are routinely made use of in technical analyses (for example, Welch (2016) uses notions like AspP, Asp’, TP, spell-out, [person], which must be assumed to be innate).

As noted earlier (§5.1), some influential authors have suggested that perhaps much less is innate than has traditionally been thought, and it has specifically been suggested that substantive features are not innate (e.g. Hornstein 2018). This is what I would call “minimal UG”, and the more minimal it is, the more it is compatible with (or even indistinguishable from) the Boasian/Greenbergian approach (as I noted in §2, everyone accepts the existence of biocognitive prerequisites for language). For this reason, it is not sufficient to contrast “UG” versus “non-UG” approaches. It is specifically the rich-UG approach of Chomsky (1981), Baker (2001), and Huang and Roberts (2016) that allows one to combine p-linguistics with g-linguistics, and that can be seen as a competitor of the Greenbergian approach.^[24]

In §5.1, I also said that the 21st century minimalist view of what is innate is “less ambitious” than the traditional generative approach, and I was asked by a colleague whether it wasn’t the other way round: “It is more ambitious to assume fewer innate constraints and still derive the same results.” Now that would certainly be the case, but de facto, leading 21st century generativists have largely given up on the Bakerian parametric programme (cf. Boeckx 2014). They do not derive the same results that authors like Roberts, Cinque and Baker hoped to derive from innate parameters and principles. A lot of 21st century work has been studying particular languages (e.g. Adger et al. 2009; Holmberg et al. 2019; Pesetsky 2013; Welch 2016), but the kinds of explanatorily ambitious proposals that were characteristic of the 1980s and 1990s seem to be largely absent. Thus, by eliminating “richness” of UG, generative grammarians have also tended to reduce the explanatory scope of their analyses. There is no clear contribution to general linguistics in this kind of work.

6.6 Top-down versus bottom-up styles

Several colleagues have mentioned that there is a contrast between a top-down and a bottom-up style of comparative research. One may either generate a highly general hypothesis and then look for confirming or disconfirming evidence, or one may start in a bottom-up way with a survey of the phenomena. In practice, most comparative work represents a mixture of these styles, but it is indeed the case that some linguists spend more energy on top-down proposals, while others spend more energy on bottom-up research. This contrast has also sometimes been called “deductive” versus “inductive”, and some people may want to describe it as “theoretical” versus “empirical” (see note 5). For example, the book series Empirical Approaches to Language Typology (De Gruyter, 65 volumes since 1987) presumably intends “empirical” in the bottom-up sense. This represents a different dimension from the “theoretical-applied” dimension that I focus on in this paper, and crucially, it is not relevant to language-particular analyses (the kind of research that the vast majority of linguists are concerned with). P-linguistics is not more or less “theoretical” or “empirical”; it is only more or less “general” (either focussing on general implications, or leaving these aside).