Word order and case in models of simulated language evolution

An attempt to simulate the emergence of case-like behaviour in populations of communicating software agents is presented. An implementation of an Iterated Learning Model is described based on Kirby [44]. The occasional emergence of grammars with distinguishable subject and object noun categories is noted. Changes are then made to this model to enable the use of multiple word orders, in the hope that this will promote such behaviour. These changes do not appear to promote the emergence of such grammars to the degree anticipated, due to the fact that there is no requirement for agents to understand each other. Further changes are made to the model such that learners will reject utterances which they believe to mean something other than what the speaker intended. As a result, there is a rise in the relative number of two-noun category grammars emerging, although the changes to the model also have a destabilising effect, resulting in a decrease in the absolute number. Experiments are also described involving external manipulation of the subset of the meaningspace that agents are permitted to use, also known as the learning bottleneck. The results of this appear to show that in the presence of a very strong bottleneck, regular and fully compositional grammars with a single noun-category are favoured, whilst relaxing the bottleneck to an intermediate value seems to promote the emergence of two-noun category grammars. Finally, a different approach to the emergence of case is described. This involves attempting to achieve proper inflectional case markings in the absence of free word order by treating inflection as another level of compositionality. The emergence of inflectional endings again seems to be favoured by the imposition of a moderately sized learning bottleneck.