🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Obtaining Faithful Interpretations from Compositional Neural Networks deepai.org/publication/ob… by Sanjay Subramanian et al. including nitish gupta, Sameer Singh, Jonathan Berant, Matt Gardner #Computation #Language

ACL paper w/ Qiang Ning, Daniel Khashabi 🕊️ and Dan Roth: arxiv.org/abs/2005.04304. We train LMs to better understand time (events' duration, frequency, etc.) and to give quality predictions. Our model also does better on related extrinsic tasks (e.g., sub-event RE) via fine-tuning.

Reading arxiv.org/pdf/2005.00724… Sanjay Subramanian et al. Evidence of "non-faithfulness" in NMNs: the disentangled neural modules we get from e.g. CLEVR don't arise in models trained on natural data. But what does it mean for an NMN to be faithful? 1/

Happy to host our first speaker Nitish Gupta for the summer zoom seminar series at UU Data Science. He will be speaking on neural module networks for complex reasoning over text, more details here : datascience.utah.edu/club/summer-se…. Stay tuned for upcoming exciting talks. @UtahSoC University of Utah

1/ We have a very poor understanding of privilege in India. This has far-reaching effects on the products we design and build. I am writing this in English. You are able to read it. Congrats - we are both privileged.

It was great to meet fellow QA researchers at the #acl2020nlp Birds-of-Feather QA Session. Great job organizing Daniel Khashabi 🕊️ and nice to see Jonathan Berant Hanna Hajishirzi Nitish Gupta Sewon Min . Seems we should take the discussing to a future workshop? 😊

Come join us in the Q&A session tomorrow on Obtaining Faithful Interpretations from NMNs. We'll be happy to answer questions, discuss, or just chat! (Ben Bogin Sanjay Subramanian Sameer Singh Jonathan Berant Matt Gardner) virtual.acl2020.org/paper_main.495…

In case you haven't seen enough papers on arXiv recently a new EMNLP findings paper #emnlp2020 led by Inbar Oren: arxiv.org/abs/2010.05647 We know already seq2seq models generalize badly to new compositions in semantic parsing... 1/3

It wasn't easy to review #nlproc research for an extremely unusual year, but here's an attempt. Listen to the episode or see below for the main themes I covered, and links to some of the representative papers for each (mostly in order I talked about them). Warning: Long thread...

Check out SacreROUGE, from my (former) lab mate Dan Deutsch I am a big fan of Sacre* metrics libraries and would like to see more. github.com/danieldeutsch/…

A reminder that the AKBC deadline (June 17th) is 4 weeks away! Please help RT and submit your work! The submission site is open and all the details can be found at: akbc.ws/2021/cfp/ #AKBC2021

Thrilled to receive the NSF CAREER award! Looking forward to exciting research & collaborations. Thanks to my mentors, colleagues, & esp students. Also thanks to Shashank Srivastava for his support & to this gentleman for waiting until the proposal submission ddl to arrive into the world.

Really excited about this work w/ Sameer Singh and Matt Gardner! We show that enforcing consistency between semantic parses of related NL utterances leads to improved learning in weakly supervised setting! Also, it got accepted to ACL 2021! Paper - arxiv.org/abs/2107.05833

Some life updates: I defended my PhD in April! Special thanks to my advisor Dan Roth and Sameer Singh & Matt Gardner for being there during the journey! Very excited to have joined Google AI as a research scientist this week! Looking forward to continuing research in NLU 😀

#NLPaperAlert: QA Dataset Explosion!🔥 A survey of 200+ QA/RC datasets proposing a taxonomy of formats & reasoning skills. Also in the bag: modalities, conversational QA, domains & beyond-English data. Honored to work on this with Matt Gardner & Isabelle Augenstein arxiv.org/abs/2107.12708

Super excited to be part of this journey with Partha Talukdar and other amazing colleagues at Google Research! #GoogleForIndia

Excited to release IndicGenBench: A suite of evaluation datasets for multiple tasks in 29 Indic languages! Evaluations over many LLMs reveal huge room for improvement. IndicGenBench is multi-way parallel opening doors for interesting research! Great work by Harman Singh!

IndicGenBench (arxiv.org/abs/2404.16816) is now available on Huggingface!

Excellent and very timely work on mitigating Reward Hacking for Reward Models!