How many languages in typological study?

It is impossible to examine all of the approximately six thousand languages of the

world (Comrie, this volume) in order to describe their typological variation and infer

language universals. As with other sciences, a sample of languages must be taken, and

the universals that are inferred are dependent on the quality of the sample. The standard approach to sampling is to ensure that the languages in the sample are drawn from as many different language families and different geographical regions as possible given the

size of the sample (anything from a dozen to several hundred languages). The reason for this is that for statistical purposes, the grammatical constructions in the different

languages in the sample should be historically independent. For example, the fact that the numeral precedes the noun in both English and German is probably due to their descent from a common ancestor language (proto-Germanic), hence including both languages in a typology of word order is to be avoided.

However, a proper sample must be large enough to capture all of the diversity of

human language. With a large sample, the likelihood increases that the typological traits of two or more languages in the sample are historically related, either by descent from a common ancestor or through language contact, recent or ancient. In fact, many linguists believe that all modern languages may be descended from a single common ancestor, even if the time depth and amount of change from that time means we can never construct a complete language family tree with a high degree of confidence. Thus it is possible that some typological traits have been a result of the form of Proto-World, or at least some very ancient protolanguage. For example, the indigenous languages of Papua New Guinea that do not belong to the Austronesian language family are almost entirely subject-object-verb (SOV) in word order. The island of Papua New Guinea was colonized at least 35, 000 years ago if not earlier, and it is likely that the SOV word order of all the languages on Papua New Guinea are historically related through common ancestry or contact for at least 35, 000 years. Some linguists have even suggested that SOV word order may have originated with Proto-World, though the evidence for this claim is fragmentary and circumstantial.

From a traditional statistical point of view, the fact that even distantly related

languages may share typological traits through historical relations is problematic. There is an alternative to the sampling strategy, however, which is precisely to take historical relatedness into consideration when constructing and evaluating language universals.

Dryer’s sampling method is the first in linguistics to do this (Dryer 1989). He does not discard any language from his sample, even if they closely related genetically or in the same geographical instead. Instead, he aggregates the data from closely related languages into genera (a genetic grouping of approximately the same time depth as a major subgroup of Indo-European such as Romance or Germanic), and then aggregates those results into six large linguistic areas in the world (Africa, Eurasia, Southeast Asia & Oceania, Australia & Papua New Guinea, North America and South America). For example, using this method, Dryer demonstrates that the hypothesized relationship between adjective-noun order and object-verb order is a consequence of a high preponderance of this correlation among genera in the Eurasian area, and is not an inherent causal connection between these two word orders (Dryer 1991).

More recently, Maslova and her colleagues have used a sophisticated mathematical

model by taking pairs of closely related languages, inferring from the match or mismatch of typological traits in those pairs the likely rate of change of the typological trait, in order to determine what the stationary distribution of the typological traits are, and hence their inherent dominance or preference (Maslova 2000). For example, in an unpublished paper Maslova and Rakhilina demonstrate that the current distribution of alignment patterns between ergative and accusative (see §4) is not a stationary distribution: it

appears that a stationary distribution would imply a higher proportion of the accusative alignment pattern, and hence a higher dominance of the accusative pattern.

The best solution to the problem of historical relatedness of typological traits in

crosslinguistic research is to use the linguistic family tree directly to model the evolution of typological traits through time and identify any causal relations between those traits.

This method has not been applied to linguistic traits in published studies (but see

unpublished word by Michael Dunn and Russell Gray), but it has been applied to cultural traits using language family trees. For example, Holden and Mace (2003) use the structure of the Bantu language family tree, derived from linguistic data using techniques from molecular phylogeny, to support a causal connection between the adoption of cattle and the loss of matrilineal descent patterns in prehistorical southern Africa. Pagel and Meade (2005) use a similarly constructed tree for Indo-European and argue that proto-Indo-European culture had monogamy and a dowry system, and daughter societies were first to polygyny and then to the absence of a dowry or presence of a bride-price (i.e. a shift to polygyny causes the abandonment of a dowry). The same methods can be applied to, for example, the relationship between object-verb order and genitive-noun order—two orders which are strongly though not completely correlated across languages—in various language families, in order to determine the causal connection between those orders and the directionality of causation.

These new techniques will allow us to use historical relations among languages to

discover language universals rather than treating them as obstacles in inferring language universals. The techniques drawn from molecular phylogeny will have to be modified in order to incorporate historical relations through language contact as well as common ancestry. Finally, the integration of historical relations between languages in inferring language universals requires a shift to analyzing language structure as part of a dynamic evolving linguistic system over time, and language universals as universal constraints on the evolution of linguistic systems. This shift in approach, necessitated by methodological issues in constructing language samples, actually dovetails with the general shift in typological theory from universals of synchronic language systems to universals constraining language change (§6).