This week, I will follow up the last week’s post and comment on the news from this year’s WMT that was collocated with EMNLP. As every year, there were many shared tasks on various types of translation and evaluation of machine translation. News translation task The news translation task is the oldest task at WMT and sort of a flagship task providing benchmarks for MT research in the long term. Test sets are created by manually translating recent news stories […]

Today, I am going to talk about a topic that is rather unknown to me: the safety and vulnerability of machine translation. I will comment on a paper Targeted Poisoning Attacks on Black-Box Neural Machine Translation by authors from the University of Melbourne and Facebook AI. The main issue making machine-translation users vulnerable is that they typically do not understand the target language and do not have any other choice than trusting the system that target-language output is adequate. Most […]

Last year an EMNLP paper “On NMT Search Errors and Model Errors: Cat Got Your Tongue?” (that I discussed in MT Weekly 20) showed a mindblowing property of neural machine translation models that the most probable target sentence is not necessarily the best target sentence. In NMT, we model the target sentence probably that is factorized using the chain rule into conditional token probabilities. We can imagine the target sentence generation like this: The model estimates the probability of the […]

In this extremely long post, I will not focus on one paper as I usually do, but instead will show my brief, but still infinitely long notes from this year’s ACL. Many people already commented on the virtual format of the conference. I will spare you of that and rather talk about the content of the conference including a list of short summaries of papers. Focus on Evaluation Many papers commented on how we evaluate our models and many of […]

This week, I will comment on a recent pre-print by Facebook AI titled Pre-training via Paraphrasing. The paper introduces a model called MARGE (indeed, they want to say it belongs to the same family as BART by Facebook) that uses a clever way of denoising as a training objective for the representation. Most of the currently used pre-trained models are based on some de-noising. We sample some noise in the input and want the model to get rid of it […]

It is a well-known fact that when you have a hammer, everything looks like a nail. It is a less-known fact that when you have a sequence-to-sequence model, everything looks like machine translation. One example of this thinking is the paper Paraphrase Generation as Zero-Shot Multilingual Translation: Disentangling Semantic Similarity from Lexical and Syntactic Diversity recently uploaded to arXiv by researchers from Johns Hopkins University. The paper approaches the task of paraphrase generation, i.e., for a source sentence, they want […]