Language Evolution and Computation Bibliography

Of the known animal communication systems, human languages appear to be unique in their use of recursively characterizable structural relations among sequences of sounds or gestures and the meanings those sequences can be used to express.

The patterns of structural relations that recursion makes possible can serve a range of communicative functions, tremendously extending the expressive resources of the system. Certain structural relations may be used to express specific meanings, or to modify, or extend, or restrict the meanings conveyed by words and other simple constituents.

Despite the unbounded complexity it makes possible, a recursive communicative system can be learned relatively easily because the constituents of a complex construction may themselves be simpler instances of that same kind of construction, and the properties of complex constructions are often predictable from simpler counterparts.

The research described in this paper is an investigation of how recursive communication systems can come to be. In particular, the investigation explores the possibility that such a system could emerge among the members of a population as the result of a process I characterize as ``negotiation,'' because each individual both contributes to, and conforms with, the system as it develops. The members of the population are assumed to possess general cognitive capacities sufficient for communicative behavior, and for learning to modify their behavior based on observations of others. However they are given no external guidance about how their communication system is to work, and their internal cognitive mechanisms impose few specific constraints.

A specific model of the ...

A model of simple agents capable of sending and receiving se- quences of characters and associating them with elements of a set of structured meanings is used to explore the emergence of systematic communication. In computational simulations, each member of a population alternates between learning to interpret the sequences sent by other members, and sending sequences that others learn to interpret. Eventually the agents develop highly coordinated communication systems that incorporate structural regularities reminiscent of those in human languages.

If the members of a population of animals are to enjoy the benefit that might accrue from the exchange of information, their communicative behavior must be coordinated -- most of the time that an animal sends a signal in some type of situation, others respond to the signal in a manner appropriate to the situation that inspired it. We investigate how coordinated communication could emerge among animals capable of producing and responding to simple signals, and how such coordination could be maintained, when new members of a population learn to communicate by observing the other members. We describe a learning procedure that enables an individual to achieve the maximum possible accuracy in communicating with a given population. If all new members of the population use this procedure, or one of the approximations to it we describe, the coordination of the population's communication will steadily increase, ultimately yielding a highly coordinated system. Our results are derived mathematically from a formal model of simple communication systems. We illustrate these results with computational simulations. and discuss their biological plausibility and their relevance to more complex communication systems, including human language.

Recurrent neural networks can be trained to recognize strings generated by context-free grammars, but the ability of the networks to do so depends on their having an appropriate set of initial connection weights. Simulations of evolution were performed on populations of simple recurrent networks where the selection criterion was the ability of the networks to recognize strings generated by grammars. The networks evolved sets of initial weights from which they could reliably learn to recognize the strings. In order to recognize if a string was generated by a given context-free grammar, it is necessary to use a stack or counter to keep track of the depth of embedding in the string. The networks that evolved in our simulations are able to use the values passed along their recurrent connections for this purpose. Furthermore, populations of networks can evolve a bias towards learning the underlying regularities in a class of related languages. These results suggest a new explanation for the ``critical period'' effects observed in the acquisition of language and other cognitive faculties. Instead of being the result of an exogenous maturational process, the degraded acquisition ability may be the result of the values of innately specified initial weights diverging in response to training on spurious input.