Researchers tracked CO2 emissions from training German, French, and English machine translation engines, and found this language pair to be most carbon-intensive. Continue reading
Combing through papers released from 2010 to 2020, researchers identify overreliance on metric BLEU to the exclusion of 100+ alternatives — a major weak point. Continue reading
Facebook, Naver, and George Mason University release open-sourced metric based on Covid-19 terms; hope to bridge gap between academia and industry. Continue reading
Scientists propose new system to standardize automatic evaluation of NLP systems by adding humans in the loop; encourage fellow researchers to submit text generation models for evaluation. Continue reading
Though beam search can boost BLEU scores, it can also lead to high rates of misgendered pronouns, even when translating between two gendered languages. Continue reading
The most-cited neural machine translation research papers show how NMT came to dominate the field — and how academia and industry’s interests have evolved. Continue reading
Google study shows human-paraphrased reference translations and new evaluation metric, BLEUP, produce better translations. Findings to be integrated into consumer-facing products. Continue reading
Facebook AI invites developers to improve a speech translation dataset and a open sources a toolkit that evaluates simultaneous speech and text translation. Continue reading
Association for Computational Linguistics names Best Research Papers of 2020 from a field of more than 3,000 submissions; paper that argues for retiring BLEU is runner-up. Continue reading
As machine translation quality improves, precision-based metrics struggle to identify the best systems. Why have newer metrics not ousted longtime standard BLEU? Continue reading
As the language industry transitions to an operating model where the first draft is mostly generated by machines, many are looking for ways to quantify linguists' value-add. Edit-distance can help. Continue reading
Google has admitted machine translated content is a challenge to ranking algorithms. Researchers at the University of Tokyo may have found a solution: back-translation. Continue reading
NVIDIA Senior Deep Learning Engineer Chip Huyen details progress on MT challenges, including data requirements, sentence-length limitations, and imperfect quality evaluation techniques. Continue reading
Facebook has been granted a patent for its machine translation quality evaluation system, which relies on user engagement data to decide which translation users prefer. Continue reading
Your information will never be shared with third parties. No Spam.