Language Evolution and Computation Research Unit
Department of Linguistics, University of Edinburgh
simon@ling.ed.ac.uk
http://www.ling.ed.ac.uk/~simon

abstract

Language is an adaptive system. In order for languages to persist over time, they must be transmittable from one generation of language users to another. The mechanisms of language transmission – i.e. use and acquisition – influence the relative transmittability of language variants, and therefore the emergent structure of language. An important question, obviously, is whether we can predict the structure of language given a particular hypothesis about these mechanisms. For example, it might be that we can simply read-off the properties of Universal Grammar from the structure of the language acquisition device (a view implicit in much of generative linguistics), or it may be that the link between language use and acquisition on the one hand, and the structure of syntax on the other, is rather more indirect.

In this paper, I explore these issues by examining the types of possible mappings between meaning space and signal space. Human languages typically employ mappings that are topographic. Informally, this means that structure in one space is preserved in the other. The most striking (and uniquely human) example of the topographic nature of language is recursive compositionality. Put simply, the meaning of an utterance tends to be a function of the meanings of parts of that utterance. Compositionality, however, is not a universal property of signals in human language. For example, highly frequent meanings seem to map onto signals in a rather different way from less frequent ones. They tend to be expressed irregularly (i.e. non-compositionally), and with shorter signals.

How are we to interpret these fundamental properties of the morphosyntax of human language? Should they be trivially predicted by our theory of the structure of the language faculty, suggesting that language users are directly constrained by their biology to acquire only languages structured in this way? Or does taking an adaptive systems perspective on languages make a less direct link (and more plausible language faculty) possible? Supporting the latter hypothesis, I will argue that the particular patterns of compositionality found in language are emergent adaptations by language itself to two simple properties of language users: generalisation, and avoidance of effort.

That this is the case can be demonstrated using a simple computational model of the process of language transmission (an extension of the model presented at the London evolution conference). It can be shown that, if the simulated language learners are predisposed to look for generalisations across examples, completely compositional encodings are inevitable "attractors" for evolving languages, even if initially, languages are completely non-compositional. This result, though interesting, is not entirely satisfactory, because it does not account for frequent meanings being short and irregular.

If a least-effort principle for the speakers is modeled, however, the mean length of utterance in the languages of the simulation tends to be much lower. Now, if the distribution of meanings in the simulation is skewed so that some are more frequent than others, the most frequent meanings do appear to behave differently to the others. They are shorter, and are expressed with either partial compositionality or as fully idiosyncratic strings. In other words, the combination of generalisation and least-effort on the part of learners and speakers in the simulation gives rise to meaning-signal patterns that look very natural.

This result can be explained in terms of adaptation. The mechanisms of acquisition and language-production pose adaptive problems for language to solve if it is to be faithfully transmitted from generation to generation. It turns out that the optimal language types for these two "problems" are actually different. In other words, the pressures from generalisation and least-effort are in competition in the evolution of language. Simplifying somewhat, the best language for learning is the most compositional encoding of the meaning space (as evidenced by the results from earlier simulations), whereas the best language for production is the minimal-length encoding of the meaning space. Generally, these cannot be the same.

The solution that language finds to this conflict of interests is interesting because the relative importance of the learning and production pressures changes with the frequency of meanings. Infrequent meanings must be expressed using a compositional syntax otherwise they are unlikely to survive the process of learning from a sub-sample. This is less true for frequent meanings. Conversely, the more frequent a meaning, the more it is subject to erosion from speakers, tending to make it shorter (and eventually unable to be compositional).

These results show us that the relationship between the structure of the human language faculty and the emergent structure of language can be quite indirect. In order to understand the origins of the particular features of human language, we should not simply look at the way humans have adapted to be better at learning language, but also the way language has adapted to being better at being passed on by us.

 Conference site: http://www.infres.enst.fr/confs/evolang/