Different communication frameworks have been developed to implement mutual interactions and opinion dynamics on social networks. Opinion formation is a complex process, and a realistic description should also take into account the feedback that the agents' opinion has on the structure of the network and on the opinion dynamics itself. We propose a model in which different kinds of interconnections and interacting behaviours are associated to the agents depending on their opinion\: extremists tend to self-segregate whilst having a stronger convincing power toward other agents. The system is fully described by a static and a dynamical parameter; in this space of parameters, a curve delimiting two different final-state scenarios can be drawn\: a continuous phase transition appears, separating an ordered consensus phase from a pluralistic situation. Pluralism can only be achieved when extremists are not too much self-segregating and their tolerance threshold is high.