Our contribution tackles the problem of learning to understand anaphoric references in the context of robotic machine learning; e.g. Get the large screw. Put it in the left hole. Our solution assumes the probabilistic theory of learning spelt out in earlier publications. Associations are formed probabilistically between constituents of the verbal command and constituents of a presupposed internal representation. The theory is extended, as a first step, to anaphora by learning how to distinguish between incorrect surface depth and the correct tree-structure depth of the anaphoric references.