Abstract: We describe recent work applying Artificial Life to investigate the evolution of language. First we present work in which the physiological cost of supporting a language ability is represented. It is seen that the benefits of successful communication can lead to evolution of language ability, despite the cost. It was observed, as in a number of other models, that different dialects of signal use emerged. We investigate the evolution of diversity in signal use and compare our findings with other published results.

To help investigate the biological evolution of language ability and the cultural evolution of languages, a variety of models have been developed [1,2,3]. These models have generally either shown the emergence of a common language in populations of non-evolving language capable agents, or the evolution of innate communication schemes. An exception is [4] where evolution acts on the initial weight values of ANN agents to improve their success at learning self-organizing languages. A shortcoming in these works is that the significant physiological evolution required to allow the use of complex languages is not modelled or represented in these models. This is an important feature of language evolution, since not only was physiological adaptation necessary for language use, but many of the adaptations are otherwise costly and detrimental.

Thus, we developed a model in which populations of simple ANNs learn use of an abstract language from the interactions of individual agents. As others have done, we show that it is possible for a self-organising language to emerge from the interactions of language learners.

Within our model, the signalling ability of an agent is determined by the number of active nodes at its signal layer. By making this a hereditary trait it is possible to experiment with the evolution of language ability. Experiments are performed both where there is no cost associated with language ability and where a fitness penalty is applied according to the number of language nodes possessed by an agent. Fitness is gained in our model by successfully understanding signals received from other agents.

Given only the natural constraint that agents interact locally, and most often with those closest to themselves we demonstrate the evolution of a costly language ability. Removing, or weakening this constraint we show that linguistic ability does not evolve where there is a cost associated with it [5]. Throughout the experiment the agents in every generation communicate and learn from one another, and are able to make use of language despite having potentially heterogeneous and/or limited signalling abilities.

It was observed in this work that, where agent interactions were bound by locality, there often existed variation in the signals used by agents over the whole population. This appeared to show the emergence of dialects within the model. As diverse dialects could appear in clusters of agents that had converged to identical ANN, genetic factors were not responsible and the diversity could only be due to the effects of distributed and localised learning. While other works have similarly noted dialect-like phenomena, few Artificial Life works have explicitly investigated the evolution of dialects and linguistic diversity. A review of these papers finds a number of explanations of linguistic diversity, based on adaptive benefits or stochastic noise effects [see 6]. This contradicts the apparent emergence of dialects in our model, so further study of this was performed.

The existing model was altered by fixing the ANN architecture of agents to an arbitrary size, removing the evolution of the language users from the model. Inter-generational communication was added to the model such that agents in the child generation learn from agents in the parent generation. The agent architecture was chosen such that easy visualisation of language evolution would be possible, by making each signal a three-bit binary value and using the three bits to generate red, green and blue components of a pixel. Using this scheme we are able to observe both the signal variation within one generation and the evolution of signal use over many generations (see figure).

We can compare results with observed phenomena in geographical linguistics, and we show that there are similarities. This lends support to our argument that linguistic diversity evolved not due to adaptive pressure, nor required linguistic innovation, but that it is a natural consequence of a distributed learning process. Current work is investigating further the minimal necessary features required for continued diversity in signal use over some population of language learners.

References

[1] Hurford, J. R., M. Studdert-Kennedy, C. Knight, (1998). Approaches to the Evolution of Language, Cambridge University Press.

[2] Hurford, J. R., M. Studdert-Kennedy, C. Knight, (In Press). The Emergence of Language, Cambridge University Press.

[3] Hawkins, J. A. and M. Gell-Mann, Eds. (1992). The Evolution of Human languages. Santa Fe Institute Studies in the Science of Complexity, Addison-Wesley.

[4] Batali, J. (1994). Innate biases and critical periods: Combining evolution and learning in the acquisition of syntax. Proceedings of the fourth artificial life workshop, Cambridge, MA, MIT Press.

[5] Livingstone, D. and C. Fyfe (1998). A Computational Model of Language-Physiology Coevolution. Presented at the 2nd International Conference on the Evolution of Language, London.

[6] Livingstone, D. and C. Fyfe (1999). Dialect in Learned Communication. Presented at AISB 99, Edinburgh.