Semiotics is a field where research on Computer Science methodologies has focused, mainly concerning Syntax and Semantics. These methodologies, however, are lacking of some flexibility for the continuously evolving web community, in which the knowledge is classified with tags rather than with ontologies. In this paper we propose a multi-agent system for the recommendation of tagged pictures obtained from mainstream Web applications. The agents in this system execute a hybrid reasoning based on WordNet and Markov chains that is able, driven by user feedback, to iteratively disambiguate the semantics of the picture tags and thus to generate knowledge from the, a priori arbitrary, information available in the Internet.