Quantifier Learning: An Agent-based Coordination Model

We consider the problem of learning the meaning of natural language expressions. In contrast to traditional settings, in which agents infer prescribed meanings from observations, we focus on an algorithm for the coordination of meaning among many agents. We do not assume any external correctness criterion. We propose an agent-based iterative algorithm for coordinating the semantics of upward monotone proportional quantifiers. We describe simple instances of our model in terms of Markov chains. We observe a mathematical connection between the possibility of convergence and specific levels of agents authority and complexity of communication patterns. We discuss the possibility of extending the model to cover the parameter of spatial separation.