Ke Tran

You want to train BERT/RoBERTa for your language in a day or two and you have only 1 GPU. Come to DeepLo workshop, #emnlp2019 to learn more about our work on transferring entire English pretrained models to foreign languages.

1/4 #ACL2024 Excited to share our new paper on the impact of fine-tuning on the qualitative advantages of LLMs in machine translation! 🤖 Our work highlights the importance of preserving LLM capabilities during fine-tuning. arxiv.org/abs/2405.20089

@AriannaBisazza @GroNlp @univgroningen @NWONieuws congrats Arianna 🎉🎉🎉!

Dear colleagues and friends @GroNLP, @univgroningen, and everywhere, I am super happy to announce that I received a @NWONieuws #Vidi grant to improve language modeling for (low-resource) morphologically rich languages, taking inspiration from child language acquisition insights!

Does Vocabulary Selection cause human perceived translation quality degradations not visible in BLEU? Yes! Find out more in our #NAACL2022 paper: arxiv.org/abs/2205.06618 Code: github.com/awslabs/sockey… joint work with @EvaHasler @sonytrenous @ketran @hifelix84 @unattributed

I'm recruiting for 3 Phd + 2 post-doc positions in #NLProc on multilingual neural machine translation at the University of Amsterdam. Apply by Feb 13. PhD positions: vacatures.uva.nl/UvA/job/PhD-Po… Post-doc positions: vacatures.uva.nl/UvA/job/Postdo… For questions DM or email me. @AmsterdamNLP

update on the bathtub :(

I'm looking for a bright and enthusiastic student that will join me and the @GroNlp group, to design more Interpretable Neural MT models! 4-year salaried PhD position in beautiful Groningen (northern Netherlands), deadline 22 March rug.nl/about-ug/work-… #NLProc @univgroningen

Are these enormous English language models or multilingual 🤔?

@Wietsedv @GroNlp @MalvinaNissim nice to see smart initialization also works for GPT-2. We did the same for BERT/RoBERTa and managed to train foreign models on one GPU within a day arxiv.org/abs/2002.07306 Code: github.com/alexa/ramen

New paper + models! It turns out that you can recycle the English GPT-2 model for new languages by only retraining one layer. In our new @GroNLP paper with @MalvinaNissim we show how you can successfully adapt English GPT-2 to Italian and Dutch without retraining from scratch.

@HarrieOos The defense should also be renamed to Us and Them :)

On Friday November 27th at 16:00 (CET), I will defend my PhD thesis titled: Learning from User Interactions with Rankings: A Unification of the Field While you should not judge a book by its cover, it is the only thing that I can share right now. 🙂 (Design done by myself.)

@nlpnoah @echau18 You also can train a BERT on data in your language in a day arxiv.org/abs/2002.07306

I am extremely excited and proud to share with you the ELLIS PhD program. To students around the world: please apply. This will hopefully become one of the most competitive PhD programs in the world. ellis.eu/en/news/ellis-…

Me: In this work, we bake a bread. Reviewer 2a: Reject! Snorting coffee via ears is not novel. Reviewer 2b: Wearing mask is not a good motivation for using ears. Also you can't snort two lines at the same time. Useless! Me: ... 🤦‍♂️

gained two more happy days by deciding to read the reviews on Sunday.

We're looking for a talented PhD student to work on the Responsible Processing of Text Data, at the intersection of #NLP and #privacy, supervised by @turkmenf @brtvrh and myself at @univgroningen. Details here: cs.rug.nl/~bisazza/misc/… Application deadline: April 1st!

@_stefan_munich @mayhewsw I'm working on the code release :)

You want to train BERT/RoBERTa for your language in a day or two and you have only 1 GPU. Come to DeepLo workshop, #emnlp2019 to learn more about our work on transferring entire English pretrained models to foreign languages.

@alex_conneau @GuillaumeLample @Thom_Wolf @PyTorch this is cool! Btw, is there any info about the number of gpus and training time of this new XLM?

Just released our new XLM/mBERT pytorch model in 100 languages. Significantly outperforms the TensorFlow mBERT OSS model while trained on the same Wikipedia data. bit.ly/2KItiC4 @GuillaumeLample @Thom_Wolf @PyTorch

@ybisk @SCSatCMU @LTIatCMU @CMU_Robotics Congrats Yo!

Life update (it's that time of year): I'll be joining @SCSatCMU – @LTIatCMU & @CMU_Robotics – Fall 2020. So share the word with anyone interested in #RoboNLP and grounding. 🙃🙏🤖

@tallinzen @jaaanaru @tyrell_turing we had a paper with a similar idea of separating syntax/semantics in the encoder and shared the attention between syntax and semantic representation. We applied it for machine translation thought aclweb.org/anthology/W18-…

@jaaanaru @tyrell_turing The idea of separating syntax from semantics (e.g. Chomsky's "autonomy of syntax") predates the brain stuff cited in this paper by a few decades. :) 1/2

Generalization by keeping syntactic & semantic information in separate streams.This simple trick from neuroscience boosts the compositional generalization score in SCAN dataset from 12.5% to 91%🤯.Deep learning people,look into the brain for inspiration! arxiv.org/abs/1904.09708