Embodying Pre-Trained Word Embeddings through Robot Actions

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