Yuhao Zhang

@yuhaozhangx

PhD student at Stanford. Working on and its application in the biomedical space.

yuhao.im
Joined February 2012
8 Photos and videos Photos and videos

Tweets

You blocked @yuhaozhangx

Are you sure you want to view these Tweets? Viewing Tweets won't unblock @yuhaozhangx

  1. Pinned Tweet
    Nov 7

    Glad to share our recent work: "Optimizing the Factual Correctness of a Summary: A Study of Summarizing Radiology Reports" Work with my advisors & | , arXiv: Highlights in thread... (1/5)

    1 reply 30 retweets 71 likes
    Show this thread
    Show this thread
    Undo
  2. Retweeted
    Dec 19
    Replying to and

    To add on this, also in rel. to 's point of GCNs being useful in low-res set-ups: Heng Ji's group @ UIUC has a recent paper that shows you can do x-lingual transfer with GCNs for rel. and event extraction by exploiting the shared dep. structures ().

    4 retweets 13 likes
    Undo
  3. Retweeted
    Dec 13

    What ngrams most predict outcome Y controlling for confounds C? Blog post & python package. for text via feature selection & adversarial learning. Predict Y from C then predict Y from Y_hat + text. Pull features from trained model weights.

    3 replies 55 retweets 111 likes
    Show this thread
    Show this thread
    Undo
  4. Dec 6

    An awesome summary of takeaways from !

    Last month, I attended CoNLL+EMNLP. It was a great conference with lots of exciting new ideas. Here are short summaries of some of the ideas I particularly enjoyed, grouped into new methods, datasets and takeaways. Hope you find them interesting too!
    4 likes
    Undo
  5. Nov 7

    Ok seems like arXiv is having a really hard time getting itself back online 🤷‍♂️, but here is another link to the paper!

    Glad to share our recent work: "Optimizing the Factual Correctness of a Summary: A Study of Summarizing Radiology Reports" Work with my advisors & | , arXiv: Highlights in thread... (1/5)
    Show this thread
    1 retweet 8 likes
    Undo
  6. Nov 7

    Together with ’s recent work on critically evaluating neural summarization, we hope our work can draw the community’s attention to this important direction. Also kudos to and for their great CheXpert labeler which enables this study.

    2 likes
    Show this thread
    Show this thread
    Undo
  7. Nov 7

    We apply this training strategy to the summarization of radiology reports and find: 1) optimizing the new reward substantially improves the factual correctness of the summaries 2) surprisingly, optimizing ROUGE alone also leads to notable gains in correctness over baseline. (4/5)

    1 reply
    Show this thread
    Show this thread
    Undo
  8. Nov 7

    Our work is based on a simple yet effective idea: We first run an information extraction (IE) system on a generated summary and fact-check it against its reference. This provides us with a factual correctness reward which we then optimize via RL training... (3/5)

    1 reply 1 like
    Show this thread
    Show this thread
    Undo
  9. Nov 7

    While existing abstractive summarization models can generate summaries which highly overlap with references, they are not optimized to be factually correct. In this work we show the success of an RL-based approach on improving factual correctness of summarization models... (2/5)

    2 replies 1 like
    Show this thread
    Show this thread
    Undo
  10. Nov 4

    Despite being very useful and studied for years, KBP is still a very challenging task for the NLP community, and evaluating KBP systems is especially tricky. They show that strong models such as BERT still has a 20%-30% gap in F1 compared to human performance.

    4 likes
    Show this thread
    Show this thread
    Undo
  11. Nov 4

    Came across this *KnowledgeNet* paper by and the (EMNLP19, ). KnowledgeNet is a continuously expanding, exhaustively annotated dataset for evaluating knowledge base population (KBP) systems. (1/2)

    1 reply 3 retweets 6 likes
    Show this thread
    Show this thread
    Undo
  12. Nov 1

    Truly brave hearts.

    Firefighters in Santa Paula, CA, battle the , the state's newest wildfire, which exploded overnight and forced thousands of evacuations.
    2 likes
    Undo
  13. Oct 29

    Interesting work from 's group on relation extraction: label distributions are shifted between splits in distantly supervised datasets, while fully-supervised datasets such as TACRED do not have this issue. A simple label adaptation method can improve this.

    In our work @emnlp2019, we identified an important yet long-overlooked issue in distantly supervised RE. We also found heuristic threshold and bias adaptation to be helpful in addressing this. arXiv: Code:
    Show this thread
    1 retweet 6 likes
    Undo
  14. Oct 10

    Thank you Rob for the photo! To be fair many credits should go to the community for their great work over many years!

    Demo of the library at by The upcoming language coverage, 83(!), is really impressive!
    9 likes
    Undo
  15. Oct 10

    Happening today! I am speaking about StanfordNLP, our new toolkit at Dev Conference. Come talk to me if you are around and interested!

    . will talk in the early afternoon about our -based stanfordnlp library for analysis of over 50 languages
    3 retweets 10 likes
    Undo
  16. Sep 10

    Pls come and join us if you are close by!

    Very excited to have speak at our next meetup to talk about their new python library for in multiple languages! & Jason Bolton will speak. Please join us if you are free on Sept 26!
    Show this thread
    5 likes
    Undo
  17. Retweeted
    Aug 6

    Up over 1000 followers, but surely there are more people on Twitter than that! One especially good reason to follow this account: from 2020 onwards, we won't be switching from ACL to ACL: will become ACL 2021 once ACL 2020 is done with it!

    4 replies 53 retweets 132 likes
    Show this thread
    Show this thread
    Undo
  18. Retweeted
    Jul 17

    Our new-ish, neural, pure Python stanfordnlp package provides grammatical analyses of sentences in over 50 human languages! Version 0.2.0 brought sensibly small model sizes and an improved lemmatizer. Try it out: pip install stanfordnlp

    24 replies 794 retweets 2,155 likes
    Undo
  19. May 23

    This article from walks through the cool functionalities in our Python library: . Quote: "[cons] the size of the models is too large (English is 1.9 GB)" - we now have models that are an order of magnitude smaller! Try it out!

    2 retweets 2 likes
    Undo
  20. May 17

    Also official website: Kudos to my colleagues Jason Bolton and for all the hard work that went into this!

    2 retweets 5 likes
    Show this thread
    Show this thread
    Undo

@yuhaozhangx hasn't Tweeted yet.

Loading seems to be taking a while.

Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.

    You may also like

    ·