For a long time, linguists have been more interested in general dependencies between different aspects of language structure than in categories shared by all languages, or, in other words, they have found conditional universals (“if a language has X, it also has Y”) more interesting than unrestricted universals (“all languages have X”) (cf. Table 1 in Evans & Levinson's [E&L's] target article). Although there are obvious similarities between the kinds of categories languages have, and linguists have found it possible to carry over category labels from one language to another ever since Greek grammatical terminology was adopted for Latin, there is something unsatisfactory about claiming, for instance, that “all languages have nouns and verbs.” Clearly, all languages allow their speakers to talk about people, things, and events (nobody disputes this), and all languages have specialized words for meanings such as “mother,” “tree,” “sleep,” and “hit.” But how do we know whether the “tree” word in a language is a noun? The basic problem is that there is no way of defining structural categories such as noun, article, dative case, or future tense in such a way that the definition applies to different languages, other than via meanings that they share – but that these meanings are cross-linguistically comparable is not at issue. Thus, whether the Straits Salish words for “run” and “(be a) man” are verbs and nouns, or whether they are merely subclasses of a more general word-class “major word,” is not decidable (Croft Reference Croft, Vogel and Comrie2000b). There are no agreed upon criteria for nounhood, which makes debates about this rather pointless; and the same applies to all kinds of other structural categories (even word; cf. Haspelmath Reference Haspelmath2009).

Not all comparative linguists share this view, and E&L (sect. 2.2.4) seem to regard it as an empirical question whether the noun-verb distinction will be found to be universal; but in any event the literature has put much more emphasis on dependencies (or correlations) between structural patterns. Since the eighteenth century (cf. Plank Reference Plank and Auroux2001), linguists have often felt that given a certain property X, a language simply had to have property Y as well. Language structures were often seen as coherent systems in which a change in one place would necessarily entail changes elsewhere, perhaps multiple changes throughout the system. This view is particularly associated with the structuralist movement of circa 1920 to 1970, and a famous dictum is A. Meillet's “une langue est un système où tout se tient” (‘a language is a system where everything hangs together’). Although Noam A. Chomsky's original interest in universals came from the claim that languages are unlearnable without rich innate structures, and he dissociated himself from structuralism, the 1980s and 1990s saw a strong movement of generative research into deep dependencies between different structures, due to a few “macroparameters,” whose setting during language acquisition fixes a whole range of patterns (Baker Reference Baker2001). In Pinker's (Reference Pinker1994) words, children “are not acquiring dozens or hundreds of rules; they are just setting a few mental switches” (p. 112).

However, this work on macroparameters has not been a success, and it has largely been abandoned by mainstream generative linguists (cf. Haspelmath Reference Haspelmath and Biberauer2008b). Similarly, a large number of the proposals for deep connections between different parts of grammar that were made outside of the generative paradigm have not been confirmed by subsequent research. What has happened again and again is that a particular proposal for a correlation was based on a few languages (e.g., “languages with vowel harmony tend to have object-verb order,” based on Turkish and Hungarian), but subsequent research on more languages showed that the generalization was too rash. Languages were found to cohere less than the intuitions of linguists indicated.

This does not mean, however, that universal dependencies (“conditional universals”) are not more than a “myth.” Although the structuralist and macroparametric hope that languages have an overall coherence across the entire system has not been confirmed, there are many generalizations with narrower domains that hold true, and for which there is massive empirical support. Consider the examples (1) to (6) (see Haspelmath Reference Haspelmath and Biberauer2008b for more discussion).

1. 1. If a language has an [ø] sound (front rounded mid vowel), it also has an [e] sound (front unrounded mid vowel).

2. 2. If a language has a [θ] sound (interdental fricative), it also has an [s] sound.

3. 3. If a language allows two consonants at the end of a syllable, it also allows one or zero consonants at the end of a syllable.

4. 4. If a language requires accusative marking on a patient object of some type, it also requires marking on patients of all types to the left on the following scale: speaker/hearer>3rd-person pronoun>definite NP>indefinite specific NP>nonspecific NP.

5. 5. If a language can form a relative clause on some clause position, it allows relativization on all other positions to the left on the following scale (Hawkins Reference Hawkins2004): subject>object> oblique>possessor.

6. 6. If a language uses a special reflexive pronoun for an adnominal possessor (e.g., Latin suum), then it also uses a special reflexive pronoun for the patient object (e.g., Latin se) (see Haspelmath Reference Haspelmath2008a).

All these universals, and quite a few more of a similar kind, refer to scales (or “hierarchies”; cf. Croft's [Reference Croft2003, p. 122] notion of “implicational hierarchies”). These scales can generally be interpreted in terms of processing cost. For example, the [θ] sound requires more articulatory effort than the [s] sound; the [ø] sound is less easy to distinguish from other vowels than the [e] sound; accusative marking is less redundant on patient objects that are pronouns or definite than on indefinite and nonspecific NPs; and reflexive use is more remarkable (and hence better coded) in patient object position than in adnominal possessive position (cf. He₁ criticized himself₁ versus He₁ criticized his₁ colleague).

E&L primarily attack the Chomskyan rhetoric of “Universal Grammar” (UG) for understandable reasons, given the widespread acceptance of his approach outside of linguistics. But the cross-linguistic research of the last few decades has put to rest quite a few earlier hunches and claims by linguists about deep dependencies. What remains, however, is no less interesting from a cognitive science perspective than Chomsky's innate UG: Language structures can be shown to follow scalar patterns that in most cases have fairly straightforward explanations in terms of processing cost.