Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks