ロボットの語意学習のための主観的整合性に基づくマルチモーダルカテゴリゼーション

抄録

This paper proposes a novel method of multimodal categorization for learning word-meaning. It is assumed that the mutlimodal observations do not necessarily capture matched object unlike previous work, but parts of them are still often matched with each other. Subjective consistency is introduced, which measures to what extent probability of object category in one modality is judged to be close to those in different modalities and expected to be utilized for finding the matched parts in multimodal observations. We apply this idea to extend the previous method called Multimodal Latent Diriclet Allocation for coping with the above assumption. Experimental results both with real and artificial data show the efficiency of the proposed method for multimodal categorization using multimodal data involving unmatched observations which are considered to be normal in more realistic situation of learning word-meaning through interaction with humans.