Harnessing the Universal Geometry of Embeddings

Neural Information Processing Systems 

We introduce the first method for translating text embeddings from one vector space to another without any paired data, encoders, or predefined sets of matches.