MIT NLP Group

MIT Natural Language Processing group.

[MIT NLP Group @mitnlp]

Hello from the MIT Natural Language Processing group! Follow us to get updates on what our lab members have been up to in #NLP

[Anastasios Nikolas Angelopoulos @ml_angelopoulos]

I’m thrilled to announce Conformal Risk Control: a way to bound quantities other than coverage with conformal prediction.

https://t.co/h2fDKcp66E

Check out the worked examples in CV and NLP!

The best part is: it’s exactly the same algorithm as split conformal prediction🤯🧵1/5

[YujiaBao @yujia_bao]

Need to debias your new task? Learn how, from your old one.

Check out our #ICML2022 paper "Learning Stable Classifiers by Transferring Unstable Features" with @CodeTerminator and @BarzilayRegina

Paper -> https://t.co/7M3XcsSd4o
Code -> https://t.co/qOZYZlJ53a

[Tal Schuster @TalSchuster]

New Preprint with @adamjfisch, T.Jaakkola and @BarzilayRegina. We present Consistent Accelerated Inference via 𝐂onfident 𝐀daptive 𝐓ransformers (CATs)

CATs can speed up inference 😺 while guaranteeing consistency 😼. The code is available🙀
🔗https://t.co/rdXKyDKFQB

#NLProc

[Adam Fisch @adamjfisch]

Large pre-trained Transformers are great, but expensive to run. But making them more efficient (e.g., early exits) can give undesirable performance hits.

In our new work, we speed up inference while guaranteeing consistency with the original model up to a specifiable tolerance.

[Adam Fisch @adamjfisch]

New #NAACL2021 paper out on robust fact verification. Sources like Wikipedia are continuously edited with the latest information. In order to keep up, our models need to be sensitive to these changes in evidence when verifying claims.

Work with @TalSchuster and @BarzilayRegina!

[Tal Schuster @TalSchuster]

Is your Fact Verification model robust enough? Consider adding #VitaminC 🍊

Check out our new #NAACL2021 paper: "Get Your Vitamin C! Robust Fact Verification with Contrastive Evidence" with @adamjfisch and @BarzilayRegina

🔗 https://t.co/wYgIcNKiZP

#NLProc #FakeNews📰
🧵1/N

[Tal Schuster @TalSchuster]

In our @IEEESSP paper, led by @RoeiSchuster, we control the embeddings of words by introducing minimal changes to the pretraining data (e.g. #Wiki edits).

This #word_embeddings attack affects many downstream #NLProc tasks!

@cornell_tech @MIT_CSAIL

link: https://t.co/Cbn4eu5ymL

[MIT NLP Group @mitnlp]

Congratulations!

[YujiaBao @yujia_bao]

Accepted to #ICLR2020. Congratulations to Regina, @CodeTerminator, @menghua_wu and myself! Thank the reviewers and AC for their valuable suggestions and comments. Our code and data are already available on GitHub. Camera-ready will be coming soon.

https://t.co/LctiOdw3nH

[Shiyu Chang @CodeTerminator]

#NeurIPS2019 Our work with MIT improves the interpretability of NLP models with an adversarial class-wise rationalization technique, which can find explanations towards any given class. Poster: Tue @ East Exhibition Hall B + C #1. @MITIBMLab @neurobongo @MIT_CSAIL @Bishop_Gorov

[MIT NLP Group @mitnlp]

If you're at @emnlp2019, don't miss our talks:

Towards Debiasing Fact Verification Models
* Wednesday 15:42 (2B) *
@TalSchuster @darshj_shah

Working Hard or Hardly Working: Challenges of Integrating Typology into Neural Dependency Parsers
* Thursday 15:30 (201A) *
@adamjfisch

[Darsh J Shah @darshj_shah]

Check-out our new paper - https://t.co/9xm6YGei1F
Automatic Fact-guided sentence modification.

Method to automatically modify the factual information in a sentence.

Joint work with @str_t5 , Prof. Regina Barzilay.

[Tal Schuster @TalSchuster]

Are we protected from GPT-2, #GROVER style models generating fake content?
What happens if they are also used legitimately as writings assistants?
Check our new report: https://t.co/dLET2BFaW9
with @RoeiSchuster, @Darsh71307636, Regina Barzilay.
#NLProc #emnlp2019 #FakeNews #GPT2

[YujiaBao @yujia_bao]

Few-shot Text Classification with Distributional Signatures. What happens if you take meta-learning for vision and apply it to NLP? Prototypical Networks with lexical features perform worse than nearest neighbors on new classes. How can we do better? ;)
https://t.co/Y2MCy7mnSQ

[Tal Schuster @TalSchuster]

Our #emnlp2019 paper is now on arxiv:https://t.co/UoICIb2xHX

* Extending #FEVER (fact-checking) eval dataset to eliminate bias.
* Regularizing the training to alleviate the bias.

Coauthors: Darsh Shah, @yeodontsay, Daniel Filizzola, @esantus, Regina Barzilay

@emnlp2019 #nlproc

[MRQA Workshop @MRQA_workshop]

Development datasets released! 6 in-domain and 6 out-of-domain including BioASQ, DROP, DuoRC, RACE, RelationExtraction, TextbookQA! Also released BERT baseline results. All the information at https://t.co/WGMmNk4FSS. Check out and let us know if you have questions! #mrqa2019

[MIT NLP Group @mitnlp]

The new shared task in QA at @MRQA2019 Workshop is released

[MRQA Workshop @MRQA_workshop]

Announcing new shared task at #mrqa2019 workshop @emnlp2019 Tests if QA systems can generalize to new test distributions. Details and training data available at https://t.co/OdJOkloUWU, more updates to come! #NLProc

[Yujie Qian @Yujie_Qian]

Our paper "GraphIE: A Graph-Based Framework for Information Extraction" has been accepted to #NAACL2019. We study how to model the graph structure of the data in various IE tasks. Joint work with @esantus @jiangfeng1124 @ZhijingJin and Regina Barzilay. (https://t.co/I7djXcLshm)

[MIT NLP Group @mitnlp]

Congratulations Tal! @str_t5

[Tal Schuster @TalSchuster]

Happy to share that our paper "Cross-Lingual Alignment of Contextual Word Embeddings, with Applications to Zero-shot Dependency Parsing" was accepted to #NAACL2019. The preprint is now available at https://t.co/Mm5h3tA6aL

[David Alvarez Melis @elmelis]

Our paper: "Gromov-Wasserstein Alignment of Word Embedding Spaces" is now available (https://t.co/OaLNt7hdgj). TL;DR: The Gromov-Wasserstein distance provides a simple, principled objective to align (w/o supervision) word embedding spaces, even of different dimensionality!

[MIT NLP Group @mitnlp]

@yujiabao2510 @CodeTerminator @Bishop_Gorov

[MIT NLP Group @mitnlp]

Here's our new #EMNLP2018 paper. By learning and transferring the mapping between human rationales and machine attention, our model yields over 15% average error reduction on benchmark datasets.

Paper & code : https://t.co/lkLadj1V3d, https://t.co/YVqxZxZeht