In this presentation, an effective approach for predicting nouns, adjectives and verbs is introduced for more effective communication between a humanoid robot and a human actor. There are three important challenges addressed by our approach: The first one is the accurate prediction of words in language. Most of the existing robotics studies predict words in language using the perceptual information only. However, due to noise and ambiguity in low-level sensory information, prediction using perceptual features is often incorrect. The second challenge is the meaning of the words. The existing studies mostly use discriminative methods to predict words, yet the underlying semantics of what, e.g., a certain noun represents,
is not adequately addressed in the literature. The third challenge is representation of the relations between the different words in language. It is known that humans activate in their brains not only the meaning of the word when that word is uttered but also the related words and their meaning. However, this challenge has not been
addressed in the robotics literature. In this presentation, the words in language are first conceptualized and gradually, a web of concepts is built from the interactions of the robot. The web is built using the co-occurrence information of words, modeled as a Markov Random Field and trained using Loopy Belief Propagation, a widely-used method for such tasks. The method shows on iCub, a humanoid robot, that such a web of concepts addresses to a certain extent all the challenges discussed above: the web improves prediction of word categories; it represents the meaning of words in concepts, and it represents the relations between the words and their meaning. As such, this contribution makes a first important step towards grounded
representation of a semantic network on a humanoid robot, which can be used for several high-level cognitive tasks, such as contextual reasoning, planning, language understanding, etc.