Lena Voita

@lena_voita

PhD student at UvA/EdnNLP; research scientist at Yandex Research (till spring 2020)

Vrijeme pridruživanja: lipanj 2019.

Tweetovi

Blokirali ste korisnika/cu @lena_voita

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @lena_voita

  1. Prikvačeni tweet
    29. sij

    Happy to receive Facebook PhD Fellowship! I feel extremely lucky to be supervised by and , who always support me, and to be part of such great groups as and

    Poništi
  2. proslijedio/la je Tweet
    28. sij

    Great talk by ⁦⁩ at 2020, and more to read on her blog.

    Poništi
  3. 28. sij

    NLP track at Applied Machine Learning Days at EPFL is about to start! Come talk to me if here :)

    Poništi
  4. 27. sij

    I was exited about giving a talk at before, but the snow has definitely made it even more appealing :) Hope it stays till Thursday!

    Poništi
  5. 23. sij

    Today gave a talk at Google Research Berlin about our Transformer analysis work. Met a lot of great people and feel totally happy :) Want to know details? Here are the blog posts: The Story of Heads: , Evolution of Representations:

    Poništi
  6. proslijedio/la je Tweet
    22. sij

    Interesting finding from masked language models (i.e. BERT) are much better at preserving token identity than either machine translation or traditional language models. (Joint work w/ & . More info here: )

    A figure comparing t-SNE representations of is, are, were, was across the layers of machine translation, language model and masked language model (deeper on the right, shallower on the left). Only the masked language model maintains separate representations across all layers: language models collapses them to one blob and machine translation to two(ish).
    Prikaži ovu nit
    Poništi
  7. proslijedio/la je Tweet
    9. pro 2019.

    98: tells us about the relative importance of attention heads in multi-headed attention and evolution of token representations in transformers. and I had fun chatting with Lena for this episode.

    Poništi
  8. proslijedio/la je Tweet
    8. stu 2019.

    You couldn't make it to ? You missed some sessions or are interested what other liked? Our impressions with , and Vassilina are out for you to read over the weekend.

    Poništi
  9. 6. stu 2019.

    3rd day of @emnlp2019: in his great keynote talks about sequence generation with an adaptive order and mentions (among others) our paper by my Yandex student Dima Emelianenko (w/ Pavel Serduykov). By the way, the paper is out!

    Poništi
  10. 5. stu 2019.

    2nd day of : Evolution of Representations in the Transformer! 16:30-18:00, hall 2A, poster P43 (another paper with my research parents and )

    Poništi
  11. proslijedio/la je Tweet
    5. stu 2019.
    Prikaži ovu nit
    Poništi
  12. proslijedio/la je Tweet
    3. stu 2019.

    Mentioned 's NMT paper, 2 other papers presented on Tue: on multi-pass decoding for semantic role labeling (w/Shay Cohen) on training deep and fast Transformers (w/ )

    Prikaži ovu nit
    Poništi
  13. 3. stu 2019.

    If at EMNLP, come talk to me at the poster for this paper at 16.30 on Wednesday, Hall 2A :) (paper with and )

    Poništi
  14. proslijedio/la je Tweet
    30. lis 2019.
    Poništi
  15. proslijedio/la je Tweet
    2. stu 2019.

    Tue at 13.30, Hall 2A: will show how a document-level MT model can be trained *without* any document-level parallel data + look into which phenomena are hard to capture with monolingual data alone (w/)

    Poništi
  16. 30. lis 2019.

    [3/3] Takeaway message: we show how to train models with BPE and make them more robust and up to 3 BLEU better. And yes, I do want to make one of my research parents a bit happy :)

    Prikaži ovu nit
    Poništi
  17. 30. lis 2019.

    [2/3] Usually, models have a pathological behavior of token embeddings: a vast majority of closest neighbors of rare tokens are rare tokens. But not with BPE-dropout! The embeddings are more sensible, and a model is more robust to misspellings (despite not being exposed to any)

    Prikaži ovu nit
    Poništi
  18. 30. lis 2019.

    [1/3] BPE-dropout: our new paper by Ivan Provilkov and Dmitrii Emelianenko! In training, we corrupt segmentation procedure of BPE to produce different segmentations of the same word. In inference, we use standard BPE and outperform BPE and sentencepiece.

    Prikaži ovu nit
    Poništi
  19. 11. lis 2019.
    Poništi
  20. proslijedio/la je Tweet
    11. lis 2019.

    For more info and analyses, check out the excellent blog post by : Paper:

    Prikaži ovu nit
    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·