Language Evolution without Natural Selection: How syntax emerges in a population of induction algorithms

The dominant approach to the origins of syntax in human language has been to examine the selection pressures that affect the evolution of the language faculty in humans. This talk will demonstrate a computational model in which the rudiments of a syntactic system emerge from a qualitatively simpler communication system in a population of induction algorithms. This demonstrates that we may be able to explain the origins of syntax without the need to appeal to functional pressures or biological change. The central claim of this paper is therefore that syntactic organisation is the inevitable outcome of the dynamics of learned communication systems.

The model put forward uses a population of simple induction algorithms with no initial linguistic knowledge (i.e. zero grammars). Members of this population are selected for replacement completely at random (so there is no natural selection pressure). Each agent must produce utterances for meanings randomly selected from a set having a predicate/argument structure. Initially the agents do not produce any utterances except for rare innovation events when an utterance is produced that is a completely random string of phonetic gestures. New agents in the simulation attempt to learn how to produce utterances based on the utterances produced by their neighbours.

The results from this simple model are surprising. The "language" in the population develops from a simple idiosyncratic vocabulary with limited expressive power and little coordination among members of the population, to one with nouns and verbs, word order expressing meaning distinctions, full compositionality, all the meaning space covered and complete coordination.

The hypothesis put forward in this talk is that syntax is the inevitable outcome of the replicator dynamics for generalisations over observed behaviour. In other words, idiosyncratic "rules" relating a single meaning to an arbitrary word are poor replicators through the channel of language use and acquisition. General rules, on the other hand, are far more successful replicators since they are more likely to be expressed in the input to the next generation as they are used for a large number of meanings. The linguistic system that will inevitably emerge in this view will always be the most general and concise that the learning mechanism can support as this system will consist of optimal replicators. It is argued that for any plausible biological observational learning mechanism such a system will be syntactic.