Co-learning nouns and adjectives

Learning and using context on a humanoid robot using latent dirichlet allocation

“…Audio information is collected from a microphone. In total we use a feature vector of length 92 [29].…”

“…Basically a concept web is a convenient representation of what we are momentarily perceiving over what we already know about the semantic relations in the world. The robot connects what it sees, be it entities, properties, or applied actions, to previously deduced concepts c ∈ C using the prototype approach developed in previous work [29], [30]. A concept can be a noun (c ∈ N ), adjective (c ∈ A), or verb (c ∈ V), with C = N ∪ A ∪ V. The concepts that are highlighted by the immediate perceptions are regarded activated (c ∈ C act ), which in turn proceed to activate other related concepts (Algorithm 1).…”

“…1) Mapping Perception to Concepts: We represent concepts with a condensed prototype-based representation [29], [30], and check membership by a simple comparison against the prototype of a concept. Instead of the prototype-approach, one could use any classifier to predict the labels (N , A, V) from observations; however, we use the prototype-based approach since it performs better [29].…”

“…Instead of the prototype-approach, one could use any classifier to predict the labels (N , A, V) from observations; however, we use the prototype-based approach since it performs better [29]. For the sake of concision, the details and example prototypes are provided in Appendix I.…”

“…Each object belongs to 1 noun category, to several adjective categories, and if applicable (any action has been performed), to several verb categories. For details, the reader is referred to [29], [30].…”