Machine translation is a mature technology with great economic importance, which still has considerable room for improvement when few bilingual texts are available. This project proposes a radically different method of automatic translation: unsupervised translation, i.e. based exclusively on monolingual texts without any bilingual resources. The method is based on deep learning of sequences and the latest advances in cross-lingual word embeddings.

In addition to being a highly innovative proposal, it opens up a new paradigm of automatic translation with ramifications in other disciplines. Since we propose to represent phrases from two languages using the same coder, this has implications for the way current linguistic processors are trained, which can entirely change the way natural multilingual language processing is done and impact the language industries.

The project has the potential to disrupt the translation industry. Current machine translation has problems translating language pairs with little contact (e.g. German and Russian), and specific domains with few bilingual texts (e.g. medical or legal). Since the proposed unsupervised translation system is further enhanced with bilingual resources, this project will improve the quality of automatic translation in such cases, with a real impact on the translation industry.
Webpage:
Organization:  Ministerio de Economía, Industria y Competitividad. (Explora)
Main researcher: Eneko Agirre
Participants: 
Eneko Agirre, Itziar Aldabe, Nora Aranberri, Mikel Artetxe, Xabier Artola , Ander Barrena, Arantza Díaz de Ilarraza, Gorka Labaka, Mikel Lersundi, Oier López de Lacalle , Olatz Perez de Viñaspre