We present an unsupervised learning methodology derived from compact data encoding and demonstrate how to construct models of polysemy, priming, semantic disambiguation and learning using this theoretical basis. The model is capable of simulating human-like performance on artificial grammar learning.