While modern machine translation has relied on large parallel corpora, a recent line of work has managed to train machine translation systems in an unsupervised way, using monolingual corpora alone. Most existing approaches rely on either cross-lingual word embeddings or deep multilingual pre-training for initialization, and further improve this system through iterative back-translation. In this talk, I will give an overview of this area, focusing on our own work on cross-lingual word embedding mappings and both unsupervised neural and statistical machine translation.
Mikel Artetxe is a Research Scientist at Facebook AI Research working on the area of multilinguality. He has mostly worked on cross-lingual representation learning and machine translation, focusing mostly on unsupervised and weakly supervised approaches. Prior to that, he recently finished his PhD on unsupervised machine translation at the University of the Basque Country under the supervision of Eneko Agirre and Gorka Labaka, and has been an intern at Google, DeepMind and Facebook.