Low Resource Machine Translation

Overview

Explore the intricacies of low resource machine translation in this comprehensive lecture by Marc'Aurelio Ranzato. Delve into key concepts such as beam search, alignment techniques, and translation with uncertainty. Examine the challenges posed by the long tail of languages, focusing on a case study of Nepali-English translation. Learn about various machine learning approaches, including supervised, self-supervised, and semi-supervised learning methods. Discover the FLoRes evaluation benchmark and process, and gain insights into domain adaptation and unsupervised machine translation. Conclude with an analysis of source-target domain mismatch and final remarks on the future of low resource machine translation.

Syllabus

– Welcome to class
– Machine translation
– Beam search
– How alignment works
– Translation with uncertainty
– Evaluation
– The long tail of languages
– Study case: Nepali ↔ English translation
– Low resource machine translation
– FLoRes evaluation benchmark and process
– ML perspective
– Supervised learning
– Self-supervised learning DAE
– Semi-supervised learning ST
– Semi-supervised learning BT
– Semi-supervised learning ST + BT
– Multi-task/-modal learning
– Domain adaptation
– Unsupervised MT
– FLoRes Ne-En
– Source target domain mismatch
– Final remarks