@jmhessel
Jack Hessel
@jmhessel
3 years
💐🌼🌻 @XandaSchofield and I have a new ACL Short! 1) We set a new SoTA in most-scooped-result-during-review! :-) Yep: shuffling doesn't hurt BERT on GLUE that much. 2) We explore token reps/attn dists to quantify impact on shuffling 3) We run BERT on diff-privatized BoW docs
Tweet media one
Tweet media two
Tweet media three
5
12
61

Replies

@jmhessel
Jack Hessel
@jmhessel
3 years
While we originally subbed to EACL, arxiv during the ACL review period has hosted (as our AC put it) a "growing bouquet of permutation papers" 💐🌼🌻 we hope ours🌷can contribute :)
2
0
7
@jmhessel
Jack Hessel
@jmhessel
3 years
(And, yes! Permutations have been explored in several works before this bouquet, too! 🌸🌹++ We look forward to highlighting these in the camera ready!)
0
0
3
@thegautamkamath
Gautam Kamath
@thegautamkamath
3 years
@jmhessel @XandaSchofield Sounds very cool! Is there a readable version of your work?
1
0
1
@jmhessel
Jack Hessel
@jmhessel
3 years
@thegautamkamath @XandaSchofield Not just yet, unfortunately --- we're working for the ACL camera ready. The DP setup is very rudimentary (we actually use DP PCA on the word-doc matrix, then re-inflate), but it shows it /can/ be done, even with aggressive noising :-)
1
0
2
@ashimgupta95
Ashim Gupta
@ashimgupta95
3 years
@jmhessel @XandaSchofield Cool work! Looking forward to reading the paper.
0
0
1
@ajaydiv
Ajay Divakaran
@ajaydiv
3 years
@jmhessel @XandaSchofield Nice.
0
0
0
@ChengleiSi
CLS @NAACL🇲🇽
@ChengleiSi
3 years
@jmhessel @XandaSchofield Shameless plug of one of our old papers that also did such shuffling experiment on BERT:
What does BERT Learn from Multiple-Choice Reading Comprehension Datasets?

Multiple-Choice Reading Comprehension (MCRC) requires the model to read the passage and question, and select the correct answer among the given options. Recent state-of-the-art models have...

arxiv.org
0
0
2