How Natural Are Artificial Languages?

Abstract

A full account of the structure and rules of natural languages remains an elusive goal for linguists. One way to gain insights into the mechanics of human language is to create an artificial language. Such languages are often constructed by speakers who are motivated by the desire to design a language with a “perfect” structure that lacks the idiosyncrasies of natural languages (e.g. Lojban). Others are invented by people interested in creating a utopia (as in the case of Esperanto) or a dystopia, as portrayed in certain works of science fiction (e.g. Star Trek and its Klingon language). In all cases, creators of the artificial languages strove to make them functional languages that would or could be accepted and usable by speakers. It seems therefore reasonable to assume that the inventors of these languages drew on their native linguistic knowledge and intuitions. They deliberately designed a lexicon appropriate to their purposes and probably reflected on the morphological and syntactic properties of their languages. By contrast, the statistical properties of natural languages are opaque to everyday speakers, although they have been shown to play a role in language acquisition (Safran et al. 1996) and linguistic behavior (Fine et al. 2013). Just as phonological and syntactic features of natural languages arguably draw from a universal inventory, statistical properties may set natural languages apart from other forms of superficially similar information-encoding systems. Rao et al. (2009) undertook a statistical analysis of several natural languages and non-linguistic systems including human chromosome sequences and the programming language Fortran. Rao et al. showed that the two kinds of data can be clearly differentiated in particular with respect to entropy, which measures the unpredictability of elements (such as words) in a sequence (such as a phrase or sentence). We extend this approach by comparing the statistical properties, including entropy, of two different artificial languages, Klingon and Lojban, to Rao et al.’s data. The results reveal both similarities with, and differences from patterns that characterize natural languages and non-linguistic sequences.