Volume 61, Issue 3

The World Atlas of Language Structures is a recently published resource providing data on the geographical distribution of grammatical structures for a large sample of the world′s languages. The articles in this special issue are the result of some first attempts to use the wealth of data available in the Atlas. This introduction discusses some general aspects of the data available in the Atlas and summarizes the articles.

In linguistic typology, implications typically emerge as a result of in-depth study of a few related linguistic parameters. Another way to find implications is to infer them from a typological database with a large number of variables. In this contribution, I discuss the LINFER computer program that was developed precisely for this goal. I have applied the program to the database on which the World Atlas of Language Structures (WALS) is based, arguably one of the largest typological databases available. In this paper, I will report on some first findings. Further, I will show that around one in every 2000 value pairs from the WALS leads to an acceptable and potentially interesting implication. A simulation shows that this is 4 times as many as a purely random dataset of the same size would produce.

The paper is the first attempt to study the quantitative representation of language type in the world′s languages as a random variable, that is, to describe the a priori probability p(x) of a randomly selected language type being represented by x languages. The major goal of this approach is to understand the general properties of cross-linguistic distributions, which might then be used as the standard of comparison for specific typological studies. Using the WALS database of language types, I show that the number of representatives of a language type can be approximated by a Pareto (“power-law”) distribution.

A central methodological issue in language typology is sampling – how to choose a representative set of languages for a typological investigation. Most proposed typological sampling methods are a priori in the sense that they are based on assumed, rather than observed, effects of biasing factors – such as genealogical and areal proximity. The advent of the World Atlas of Language Structures (WALS) creates for the first time a chance to attempt a posteriori sampling. The basic idea is to create a sample by removing from the set of available languages one member of each pair of languages whose typological distance – as defined in terms of the features in WALS – does not reach a predefined threshold. In this way, a sample of 101 languages was chosen from an initial set of the 222 languages that are best represented in WALS. The number of languages from different macroareas in this sample can be taken as an indication of the internal diversity of the area in question. Two issues are discussed in some detail: (i) the high diversity of the indigenous languages of the Americas and the tendency for these to be underrepresented by previous sampling methods; (ii) the extreme areal convergence of Mainland South East Asian languages. It is concluded that areal factors cannot be neglected in typological sampling, and that it must be questioned whether the creation of elaborate sampling algorithms makes sense.

Typological distributions are the combined result of universal structural principles, areal diffusion, and shared descent. The core concern of quantitative typology is to disentangle and to identify these various factors. While areal and structural factors can be tested against each other in standard multivariate designs based on sample stratification, genealogical factors cannot be handled by sample stratification since about one third of all proven families (the strata needed) are isolates, i.e. count only one member. In response, typologists have since long sought to control for genealogical relations during sampling rather than during statistical testing. But available methods suffer from a number of drawbacks. Most importantly, they are not sensitive to the fact that different typological variable have different degrees of stability (genealogical dependence) within families, and that this again varies from family to family. This article proposes a refined method for genealogical control during sampling, which is based on DRYER′s (1989) proposals but is sensitive to actual distributions within genealogical units at each taxonomic level.

This paper presents an algorithm intended to quantify the diachronic stability of linguistic characteristics. It is argued that a linguistic feature whose presence or absence is best predicted by language families is a stable feature. Conversely, a feature that correlates better with geographical areas than with families is one that is sensitive to diffusion. Contrasting the structural heterogeneity within families with that found within geographical areas, it is thus possible to make a statement regarding the varying diachronic stability of specific features. While the main aim of the paper is methodological exploration, and while the method certainly not devoid of problems, I propose that the current approach can be useful in studies of language contact and long-range historical comparison.

In this paper we propose a method to measure the stability of typological features given a data set such as HASPELMATH et al. (2005). We determine the frequency with which one and the same value of a feature occurs in shallow genealogical groupings ('genus′ level). If the languages within such a group tend to show the same value, there is reason to believe that the feature is inherently more stable than if much variation is found within such groups. The proposed metric may be used to evaluate the utility of linguistic features when employing typological data for inferring genealogical relationships among languages.

In this note, we present three methods to discover the most consistent features in the World Atlas of Languages Structures (WALS). These methods measure the fit between each individual WALS feature and the overall dataset of all features combined. Features that show a strong fit to the overall dataset are hypothesised to be more central for the structure of human language than those features that show a weak fit. The three techniques we will use are based on (i) Mantel′s congruence test (MANTEL 1967), (ii) the evaluation of feature coherence relative to the overall dataset, and (iii) the comparisons of ranks. All three methods attempt to identify those features that fit best to the dataset in its entirety, though it turns out that they do not identify exactly the same features. Still, we are able to give some indications of the kind of features that appear to be most promising for future research. Finally, we investigate whether such highly consistent features might be suitable to uncover genealogical relationships between languages.