Mikel Artetxe is co-founder of Reka and an honorary researcher at the University of the Basque Country. His research concerns multilingual language models and learning translation systems when parallel training data is scarce.[1][2]
Start with Unsupervised Neural Machine Translation →
Education
Links
Website
X 
GitHub LinkedIn Google Scholar Co-founder and researcher · Reka AI
Honorary researcher, IXA research group · University of the Basque Country
Career
Artetxe completed his doctorate at the University of the Basque Country in 2020, advised by Eneko Agirre and Gorka Labaka. He worked at Meta’s FAIR from 2020 to 2023, then joined Reka as a co-founder while retaining an honorary research role at his university. His earlier placements included DeepMind, FAIR and Google internships.[1]
Contributions
Unsupervised Neural Machine Translation, coauthored with Gorka Labaka, Eneko Agirre and Kyunghyun Cho, trains from separate bodies of text in each language instead of paired translations. It combines denoising, which reconstructs sentences from corrupted input, with backtranslation, which produces synthetic translation pairs. Cross-lingual word representations help connect the languages. The experiments show translation can be learned with much less direct supervision, while also showing that parallel examples remain useful.[2]
Ideas
Artetxe focuses on languages and settings with fewer resources. That perspective asks whether a method depends on expensive aligned data or on assumptions that hold mainly for well-represented languages. His research includes cross-lingual representations and later multilingual language-model work.[1]
Tools, writing & talks
Writing, research & talks
Start here
Denoising and backtranslation train a translator using monolingual text.
Read the training loop and trace how a synthetic translation becomes the input to the opposite direction. Basic encoder-decoder vocabulary helps; compare the unsupervised and partially supervised experiments.
Co-author[2]