After a while, here is a dump of what I found most interesting on arXiv about machine translation and multilinguality, covering May and June of this year. Google Research published a pre-print of their NAACL paper: SCONES (Single-label Contrastive Objective for Non-Exclusive Sequences). The paper is about a simple trick: they replace softmax with binary classifiers with a sigmoid output and use the sum of binary cross-entropies as their loss function. It gets a slightly better BLEU and BLEURT score […]

Here is my monthly summary of what I found worth reading on arXiv in the past month. A preprint from JHU studies zero-shot cross-lingual transfer using pretrained multilingual representation and comes to the conclusion that it is an under-specified optimization problem. In other words, with a multilingual representation model, there are potentially many solutions that are good for the source language, but only some of them are good for the target language. In practice, the solution is probably proper training […]

Highlights from Machine Translation and Multilinguality in August 2022 | Jindřich’s blog

Here are my monthly highlights from paper machine translation and multilinguality. A preprint from the Nara Institute of Science and Technology shows that target-language-specific fully connected layers in the Transformer decoder improve multilingual and zero-shot MT compared to the current practice of using a special token to indicate what the target language is. A very similar idea is also in a preprint from Tianjin University, but in this case, they add language-specific parameters for the other part of the Transformer […]

Machine Translation and Multilinguality in August 2022 | Jindřich’s blog

Here is my monthly summary of what I found worth reading on arXiv in the past month. A preprint from JHU studies zero-shot cross-lingual transfer using pretrained multilingual representation and comes to the conclusion that it is an under-specified optimization problem. In other words, with a multilingual representation model, there are potentially many solutions that are good for the source language, but only some of them are good for the target language. In practice, the solution is probably proper training […]

After a while, here is a dump of what I found most interesting on arXiv about machine translation and multilinguality, covering May and June of this year. Google Research published a pre-print of their NAACL paper: SCONES (Single-label Contrastive Objective for Non-Exclusive Sequences). The paper is about a simple trick: they replace softmax with binary classifiers with a sigmoid output and use the sum of binary cross-entropies as their loss function. It gets a slightly better BLEU and BLEURT score […]

Here are some of my notes and comments on what I had a chance to see at ACL in Dublin last week (my first in-person conference since 2019). ACL D&I 60-60 initiative ACL announced its 60-60 initiative, for the 60th birthday of ACL, all materials that ACL produces should be available in 60 languages. The initiative already machine-translated titles and abstracts of ACL 2022 papers and did an automatic voiceover for plenary talks at the conference. Although this is definitely […]

Another month is over, so here is my overview of what I found most interesting in machine translation and multilinguality. Rotation ciphers as regularizers A paper accepted to ACL 2022 from Simon Fraser University experiments with using rotation ciphers on the source side of MT as a data augmentation technique. They tested it in low data scenarios and it seems to work quite well, which actually seems quite strange to me. It’s just systematic replacing characters with different characters – […]

Here is a monthly summary of what I found most interesting on arXiv this month from machine translation and mutlilinguality. This month was the camera-ready deadline for ACL 2022, so many of the interesting papers are accepted to ACL. Overlapping BPE When training, BPE merges actually do not have to follow the simple objective of merging the most frequent token pair. In massively multilingual models, there is an imbalance between languages, and some of them got segmented almost down to […]