The nature of language universals

There is a widely assumed view of language universals that language universals are

simply properties that are possessed by all human languages. All linguists would agree that there are certain very basic properties that are possessed by all languages. These properties include: utterances made up of discrete meaningful units; conventions of syntax; and duality of patterning (that is, organization of sound structure at a level largely independent of the organization of syntactic structure). These language universals are generally described as design features of language.

The more controversial question is over the existence and nature of universals that

make reference to more specific grammatical entities. Some proposed universals of this type would include the hypothesis that all languages make a distinction between nouns and verbs, or that all languages have consonants and vowels.

Typological research leads to two general conclusions about language universals that

are more specific reference to grammatical categories and structures. The first is that such universals are typically not in the form, “All languages have X”, that is, such universals are not unrestricted universals. Instead, almost all of the language universals that have been discovered are restricted or implicational universals, in the form “If a language has X, then it also has Y”.

Consider for example the relative orders of certain types of modifiers, in particular

adjectives and numerals. In §1, we noted that in some languages, the numeral precedes the noun it modifies, and in others, it follows (and in still others, either order is possible).

The same is true of the order of an adjective and the noun it modifies. When both

modifiers are compared in a single language, the picture changes. In English both

adjectives and numerals precede the noun:

(23) a. red book b. three books

Adj Noun Num Noun

This pattern is found in many languages. In many other languages, both adjectives

and numerals follow the noun:

(24) Kosraean (Autronesian, Caroline Islands)

a. mwet kuh b. mwet luo

men strong men two

Noun Adj Noun Num

A third group of languages has adjectives following the noun while numerals precede:

(25) Jamiltepec Mixtec (Mixtecan, Mexico)

a. ve@he@ lu! hlu b. uvi ve@he@

house little two house(s)

Noun Adj Num Noun

On the other hand, languages with the adjective preceding and numeral following are

virtually unattested (although there are a few; the existence of exceptions will be

discussed below).

The pattern of attested vs. unattested (or at least extremely rare) language types can

be given in the four-cell table below:

Noun-Adjective order Adjective-Noun order

Numeral-Noun order attested (Jamiltepec Mixtec) attested (English)

Noun-Numeral order attested (Kosraean) extremely rare

Table 1. Attested vs unattested adjective and numeral word orders.

The generalization can itself be described in terms of an IMPLICATIONAL UNIVERSAL:

(26) If a language has Adjective-Noun word order, then it (almost always) has

Numeral-Noun word order.

The discovery of implicational universals of word order by Greenberg (1966)

demonstrated that there exist universal properties of language that do not require all

languages to be identical in specific properties. The implicational universal in (26) is not by itself a description of a fact about the grammar of a particular language. In fact, one could not even identify the implicational universal without looking across a set of

languages. The implicational universal captures a contingent relationship between

Adjective-Noun order and Numeral-Noun order. Nevertheless, this contingent

relationship must be a part of individual speakers’ knowledge of language structure and meaning. In particular, a language will not (or is extremely unlikely to) emerge or change to a type that has both noun-numeral order and adjective-noun order.

Hundreds of implicational universals have already been discovered, and more are

discovered every time a typologist investigates a new area of grammar. The existence of so many implicational universals requires a rethinking of the nature of Universal

Grammar, which is usually thought of as a set of unrestricted universals. The part of

Universal Grammar that consists of unrestricted universals specifying ways in which all languages are identical captures only a very small portion of what is universal about language. It misses most of what is universal about language beyond the basic design features. The presence of large numbers of implicational universals requires a model

which allows for a great deal of variation in grammatical structure across languages, but constrains that variation to a significant degree in many different dimensions. This basic observation about language universals has not been addressed by most syntactic theories.

The second general conclusion about language universals that is revealed by

typological research is that the constraints on language variation that are represented by implicational universals are not exceptionless in the way that the design features of

language are. In the example of the relationship between numeral-noun order and

adjective-noun order given above, we noted that the fourth language type, noun-numeral order and adjective-noun order, is not actually unattested: a small number of languages of this type do exist. This is almost always observed for language universals, especially with advent of larger and larger language samples.

This conclusion also changes how language universals must be understood. One

cannot conceive of language universals (beyond the design features) as specifying what constitutes a possible human language. The noun-numeral & adjective-noun language type is not impossible; a few such languages exist. But it is far less frequent than the other three types. Moreover, the other language type with modifiers on the opposite side of the noun, adjective-noun and noun-numeral, while frequent, is not as frequent as the types in which both modifiers precede or both modifiers follow. These differences in the likelihood of language types are significant and must be explained by linguistic theory. In other words, typology shifts the scientific question about language universals from “What is a possible language type, and why? ” to “What is a more probable language type, and why? ”

For this reason, typologists have turned to more sophisticated quantitative methods in

order to identify valid language universals. If all types of languages exist in the area of grammar under investigation, then one must be confident that differences in likelihood are statistically valid. The increase in quantitative sophistication in language sampling was discussed in §2. In the case of implicational universals, Maslova (2003) gives statistical tests for identifying valid implicational universals when one or more types in a table such as Table 1 is of much lower frequency than the other types.

An important type of model for inferring the language universals underlying

generalizations such as the implicational relation in (26) is the competing motivations

model. A competing motivations model posits two or more factors that determine

language structure. However, the motivations typically do not determine a single

grammatical pattern because they are often in conflict. In the case of conflict, there is no single optimal grammatical pattern that satisfies all of the competing motivations, and instead one finds cross-linguistic variation over several suboptimal patterns. In this way, universal properties of language (the motivations) give rise to cross-linguistic diversity.

For example, Greenberg proposed two competing motivations for implicational

universals of word order. The first, dominance, can be thought of as simply a default

preference for one order over another. For example, noun-adjective order (NA) is

dominant, as is numeral-noun order (NumN). The second, harmony, can be thought of as

a dependent relation of one word order upon another. For example, AN order is harmonic with NumN order and NA order is harmonic with NNum order.

Greenberg’s two motivations compete with each other, and the result is described in

the following principle:

(27) A dominant order may occur at any time, but a recessive order occurs only when a harmonic order is also present.

The principle in (27) accounts for the distribution of languages in Table 1. The upper left cell is the language type with both dominant orders (NA and NumN), which are not harmonic with each other. The other two attested types have one recessive order, but the harmonic order is also present. The extremely rare type would have both recessive orders (AN and NNum), neither of which is dominant. That is, the extremely rare type is not motivated by either dominance or harmony, which accounts for its rarity. Finally, one cannot satisfy both motivations at once, since the dominant orders are not harmonic with each other.

A competing motivations model can be found by inspection with a simple case such

as the two interacting word orders in Table 1. But it turns out that comparing the word

orders of several different constructions across languages is much more complex: most possible types are attested, but at highly varying frequencies. Justeson and Stephens (1990) use a statistical technique, log-linear analysis, to examine the relationships between multiple word orders. Log-linear analysis allows one to construct a model of the simultaneous interactions between many word orders; the model that best fits the data provides the best model of which word orders actually might be causally connected. The best fit model for the word orders investigated by Justeson and Stephens identify the dominance and harmonic relations that are actually supported by the data in their 147-language sample. Reassuringly, Justeson and Stephens arrive at the same result that Dryer came to, namely that there is no direct causal relation between adjective-noun order and object-verb order.

Another example of the use of sophisticated statistical techniques in typology is found

with an important recent development in modeling language universals, the semantic map model. The semantic map model allows the typologist to identify language universals without assuming that grammatical categories are the same across languages, as we observed in §4. There, it was pointed out that some languages divide the semantic roles in transitive and intransitive verbs (A, P and S) in a nominative-accusative pattern (A+S vs. P) while others divide them in an ergative-absolutive pattern (A vs. S+P). Still other languages use a tripartite pattern, distinguishing all three of A, S and P, while others have a neutral pattern, not distinguishing any of them. No language (or virtually no language) groups together A and S and distinguishes that group from P.

The range of variation and the limitation of attested types of grammatical relations

can be represented by a conceptual space linking A to S and P to S, as in Figure 1:

Figure 1. Conceptual space for semantic roles.

The attested language types can be mapped as connected regions on this three-point

space:

Figure 2. Map of attested systems of grammatical relations

The unattested (or extremely rare) type of A+P vs. S cannot be mapped on the

conceptual space in Figure 1: A cannot be grouped with P without also including S,

according to the connections in Figure 1. Hence the conceptual space represents what is universal about the relationships among A, S and P, while allowing for the variation of attested language types illustrated by the semantic maps in Figure 2.

The semantic map model is a powerful tool for identifying language universals and

separating language universals from the arbitrary aspects of crosslinguistic variation, and has been widely used. However, it is impractical for identifying the conceptual space when that space consists of many more points than the three exemplified in Figures 1-2, or when the range and frequency distribution of attested language types is complex.

However, these problems can be solved by using multidimensional scaling and related

multivariate techniques. For example, multidimensional scaling was applied to the results of the experimental elicitation for the 71 pictures of spatial relations described in §3, allowing typologists to identify the most significant semantic dimensions of spatial relations in determining attested systems of spatial relations (prepositions, postpositions and case inflection; Levinson et al. 2003; Croft and Poole 2008).

The richness, diversity and complexity of typological evidence are daunting to

analyze. Typology has only recently begun to use more sophisticated quantitative tools to allow researchers to go beyond the easy to identify language universals to the universals that are harder to detect through the noise of arbitrary crosslinguistic variation. The initial successes in using such tools indicates that typologists will be able to find many more restricted universals of language.