Embodying Pre-Trained Word Embeddings Through Robot Actions

We propose a promising neural network model with which to acquire grounded representation of robot actions and the linguistic descriptions thereof. Properly responding various expressions, including polysemous words, is an important ability for robots that interact people via dialogue. Previous studies have shown can use words are not included in action-description paired datasets by using pre-trained word embeddings. However, embeddings trained under distributional hypothesis grounded, as they derived purely from text corpus. In this letter, we transform embodied ones robot's sensory-motor experiences. extend bidirectional translation incorporating non-linear layers retrofit By training layer alternately, our proposed able adapt dataset. Our results demonstrate synonyms form semantic cluster reflecting experiences (actions environments) robot. These allow properly generate $unseen$ notation="LaTeX">$words$