Discover and read the best of Twitter Threads about #ACL2019
Most recents (2)
I just volunteered for the #ACL2019 mentoring program. Thanks to @radamihalcea, Robert Frederking, and @arnaik19 for organizing this opportunity to welcome newcomers. Remembering my first major #NLProc conference ten years ago, I'm glad to help improve the experience. (1/thread)
In 2009, as a student, I attended my first ACL. I registered late and received no information about mentorship. I was the first person from my lab to go to an #NLProc conference, and I had no one to introduce me to people. Nobody had heard of my advisor, who didn’t work in NLP.
It's difficult being an outsider at a conference. Here are some observations I can pass along to students who feel like they’re starting from the same place I was, i.e., attending your first conference in a community where you have few (if any) contacts. #phdchat
When and why does king - man + woman = queen? In my #ACL2019 paper with @DavidDuvenaud and Graeme Hirst, we explain what conditions need to be satisfied by a training corpus for word analogies to hold in a GloVe or skipgram embedding space. 1/4
blog: bit.ly/2X18QEd
blog: bit.ly/2X18QEd
paper: arxiv.org/abs/1810.04882
In turn, our theory provides
1. An information theoretic interpretation of Euclidean distance in skipgram and GloVe embedding spaces.
2. Novel justification for the surprising effectiveness of using addition to compose word vectors. 2/4
In turn, our theory provides
1. An information theoretic interpretation of Euclidean distance in skipgram and GloVe embedding spaces.
2. Novel justification for the surprising effectiveness of using addition to compose word vectors. 2/4
3. A formal proof of the intuitive explanation of word analogies, as proposed by Pennington, @RichardSocher, and @chrmanning in the GloVe paper.
Most importantly, we provide empirical evidence in support of our theory, making it much more tenable than past explanations. 3/4
Most importantly, we provide empirical evidence in support of our theory, making it much more tenable than past explanations. 3/4
