Beyond binary dependencies in language structure

The study of the regularities in the structures present across languages has always been a quest in close contact with the analysis of data. Traditionally, causal dependencies between pairs of typological variables (like word order patterns or the composition of segment inventories) have been argued for on the basis of language counts, namely how many languages out of a sample exhibit certain pattern in contrast to other. Regularities of this kind have been used in virtually all theoretical camps, and researchers made them part of their discussion on functional pressures on language, cognitive schemes and the architecture of a putative common computational core underlying language, among others. This popularity resides, without doubt, in the strength and the simplicity of the idea: if a set of languages with no recent genealogical history nor traces of areal contact tend to share the same pair of properties again and again, then there seems to be under scrutiny something about the properties of probable languages in general.

While venerable and potentially useful, this procedure is complicated by many factors. First, the nature of a proposed dependency can affect how the pattern of observations translates into support for the dependency. In the first section, we show how different notions of causality and causal strength are appropriate for different types of dependencies involving two variables. Secondly, these dependencies can be distorted not only by historical relations between languages (as usually acknowledged in the literature) but also due to complex causal dependencies involving multiple variables. Addressing these concerns requires appropriate formalisms and statistical techniques. These exist and are widely used for addressing the problem of historical relations (which we cover in the second section), but methods for dealing with relationships between more than two variables are underdeveloped in linguistics. In the final section, we discuss some new approaches to detecting causal dependencies between more than two variables.