Abstract
The current paper introduces a learning model, which has the generalization capabilities such as recognizing unlearned sentences consisting of words included in learned sentences. Our model is composed of a linguistic and a behavioral module, and both of the modules interact with each other through binding neurons (BN) of hub-like network, three-layer feedforward neural network (FNN). We implemented this model to a humanoid robot and trained the robot to learn sentence set of two different grammatical types with corresponding behavioral patterns. One type is a verb followed by an objectival phrase as like "touch the red block" and the other is a verb followed by an objectival phrase and further followed by an adverbial phrase as like "put the green block on the blue one". Our analysis on the result of learning experiment showed that a compositional (grammatical) structure corresponding to two types is self-organized in the BN space.
