Abstract
An integration of a motion language model and a natural language model was proposed to realize a robot with linguistic processing. The motion language model associates words with a motion symbol stochastically. The natural language model represents sentences (sequences of words). This paper describes semantic abstraction. A language space is constructed by locating a word in a multidimensional space based on probability that the word generates a latent state in the motion language model, where latent state is expected to represent underlying linguistic structure. We validate our approach by constructing the language space on optical captured human behaviors.
