Here is what I found interesting on arXiv in December 2022 and January 2023. At the beginning of January, there a relatively few new pre-prints in general. But now it is catching momentum again, with more papers appearing every day. BLOOM+1: Adding Language Support to BLOOM for Zero-Shot Prompting In this paper, folks from the Big Science Workshop elaborate on how to add language support to the already trained BLOOM model. They tried two approaches: MAD-X (clever stuff with adapters, […]

In this post, I comment on our (i.e., myself, Helmut Schmid and Alex Fraser) year-old paper “Why don’t people use character-level machine translation,” published in Findings of ACL 2022. Here, I will (besides briefly summarizing the paper’s main message) mostly comment on what I learned while working on the one-year-later perspective, focusing more on what I would do differently now. If you are interested in the exact research content, read the paper or watch a 5-minute presentation. Paper TL;DR Doing […]

Last week I was at EMNLP in Abu Dhabi. Besides losing my passport and figuring out what to do on such an occasion (many thanks to the personnel of the Czech embassy in Abu Dhabi), I had plenty of interesting conversations and saw many interesting posters. When I was at my first NLP conference 8 years ago, I was amazed by the papers presented at the conference and returned with a long list of ideas of what I should try […]

Here are my monthly highlights from paper machine translation and multilinguality that appeared on arXiv in November 2022. A preprint with 19 authors from 13 institutions presents something like the T0 model: but instead of starting with the (more or less) monolingual T5 model, they use multilingual BLOOM and mT5 and call the resulting model BLOOMZ and mT0. The main idea is finetuning the underlying model (or the foundation model?) on as many tasks as possible so that the model […]

Here are my monthly highlights from paper machine translation and multilinguality that appeared on arXiv, many of them preprints from the upcoming EMNLP conference. Folks from Amazon published a pre-print that introduces a simple method of how to make pre-trained multilingual representation more robust towards noisy inputs. It is a very straightforward approach: they sample typos based on Wikipedia logs and use those during model training. In addition, they add a contrastive loss that forces the noisy versions of sentences […]

After 100 MT Weekly posts (which took me 130 weeks to write), I realized that weekly blogging is impossible while weekly teaching. So I decided to change the format of the post and write monthly summaries of what I found most interesting in machine translation and multilinguality. This is the first issue that summarizes what interesting happened in February. Exciting news about WMT There will be some exciting changes in WMT competitions. WMT is an annual conference on machine translation […]

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 […]

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 – […]

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 […]