Abstract: | A simulation of language evolution as described in John Batali, "Computational simulations of the emergence of grammar", fromJ. R. Hurford, C. Knight, and M. Studdert--Kennedy, "Approaches to the evolution of language: social and cognitive bases", Cambridge, 1998, is replicated, in which a population of communicative agents implemented as simple recurrent networks develops, from no explicitinitial knowledge, effective linguistic representations for a small set of predefined meanings. The author of the original paper claims to have found a language emerging from his simulation which was compositional, in that it reflected regularities in the structure of the meanings. Such results could not be fully reproduced. A detailed analysis of the results of the replication, and additional experiments, suggest that the training regime used seeks to maximise the information content of the linguistic expressions with respect to meaning space, which fails to support compositionality. It is shown that while languages which emerge naturally from the simulations can be transmitted successfully over hundreds of cycles, ideally compositional languages, which were artificially generated, can not be learned properly.
|