Language, as a shared set of conventions for mapping meanings to expressions, can emerge from the self-organization - into a global consensus state - of a population of distributed agents connected through some communication network and playing local collaborative games such as the Naming Game. Concepts and methods involved in this problem are very similar to those applied in statistical physics. In this work we propose a kind of self-organizing Semantic Overlay Networks, inspired by the mechanics of the Ising spin model – and undergoing a variant of distributed simulated annealing – which can converge to a consensus vocabulary through the abrupt transition from disorder to order; the condition which grants the convergence (the mean-field condition — a.k.a. Representative Agent condition – of everyone knowing about the state of everybody else) is approximated here by a sampling, performed through a suitably randomized message exchange mechanism. We outline two possible implementation of such kind of networks: one based on a structured, the other based on an unstructured P2P network.