Discover and read the best of Twitter Threads about #emnlp2019
Most recents (5)
Of the many excellent papers at #emnlp2019, this is the one I can't stop thinking about.
On NMT Search Errors and Model Errors: Cat Got Your Tongue?
by Felix Stahlberg & Bill Byrne
aclweb.org/anthology/D19-…
Our current NMT models might be "wrong"🚫
👇thread 👇 1/7
On NMT Search Errors and Model Errors: Cat Got Your Tongue?
by Felix Stahlberg & Bill Byrne
aclweb.org/anthology/D19-…
Our current NMT models might be "wrong"🚫
👇thread 👇 1/7
2/7. One relatively well known open dirty secrets of NMT was that if you increase the beam width at decoding, your BLEU sometime decreases
3/7. This shouldn't happen. With ♾-width, you should get the "best" translation, and (n+1)-width > n-width
Unsupervised Domain Adaptation of Contextualized Embeddings for Sequence Labeling by @XiaochuangHan and @jacobeisenstein
Link: aclweb.org/anthology/D19-…
Code: github.com/xhan77/AdaptaB…
#emnlp2019
Link: aclweb.org/anthology/D19-…
Code: github.com/xhan77/AdaptaB…
#emnlp2019
Since BERT is trained on Wikipedia, Han and Eisenstein are interested in scenarios wheee labeled data exists only for canonical source domain, and target domain is distinct from both pretraining and labeled source domain.
#emnlp2019
#emnlp2019
One of the tasks is part of speech tagging for Early Modern English. Given labeled data in 20th century English, they solve for unlabeled data in 15-17th century English.
#emnlp2019
#emnlp2019
Manning: Lehnert (1977) says that NLU can be measured by asking questions. #conll2019 #emnp2019
Manning: SQuAD leaderboard performance (read: BERT) is impressive, but hard to say these models exhibit true understanding. #conll2019 #emnp2019
Why should we study bias and fairness in NLP? Tutorial session led by Margaret Mitchell, @vinodkpg, @kaiwei_chang & @bluevincent. #emnlp2019
Human biases in data! 1) selection data: a) selection does not reflect a random sample, gender bias in Wikipedia and britannica; b) turkers are mainly from n America; 2) biased data representation: a) some groups are less represented less positively than others;
3) biased labels: a) annotations reflect world views of annotations eg. Western wedding ceremonies vs. other wedding ceremonies.
Most NLP models treat numbers (e.g., “91”) in the same way as other tokens, i.e., they embed them as vectors. Is this a good representation for downstream numerical tasks such as DROP, math questions, etc?
Yes! Pre-trained vectors (BERT, GloVe, ELMo) know numbers.[1 / 6]
Yes! Pre-trained vectors (BERT, GloVe, ELMo) know numbers.[1 / 6]
We begin by testing QA models on questions that evaluate numerical reasoning (e.g., sorting, comparing, or summing numbers), taken from the DROP dataset. Standard models excel on these types of questions! [2 / 6]
To understand how models can capture numerical reasoning, we probe their token embeddings (e.g., BERT, GloVe) using list maximum, number decoding, and addition tasks. [3 / 6]
