Self-Supervised Representation Learning for Low-Resource Languages
We present a novel architecture that achieves state-of-the-art performance on NLP tasks for languages with <10K training tokens, using cross-lingual alignment without parallel corpora.