Population and Agent Based Models for Language Convergence

Adaptive agents cooperating in rich, open environments will need shared ontologies and linguistic conventions to communicate critical information. In the face of open systems and environmental change, pre-defined ontologies are suboptimal. In open systems, new agents or web services can require new conventions not captured in existing conceptualizations. A more adaptive and more theoretically interesting approach is to have agents negotiate repertoires of categories, meanings, and linguistic forms among themselves. Many related issues arise in the general arena of “emergent semantics”; the case of language serves as an excellent “model organism” for studying the issues, and results generalize nicely to other domains including “collaborative tagging” and bioinformatics. One central problem in emergent, adaptive language is getting a population of agents to converge on shared meanings and forms—this is called the “Language Convergence” problem. The problem can be framed as follows. Suppose we have a population of N agents, each of which is using their own language. The population arrives at a converged state when all the agents are speaking the same language. When this state is achieved, the agents can communicate with each other. There are many different ideas on what a language is, but for our purposes we will use a simple definition. We assume that the lexicon (the set of possible words) is fixed, and that a language is a mapping from meanings (objects in the world to words). This is a common abstraction of language. When the mapping is represented as a matrix, it is often called an association matrix