Some personal update: I will join
@UWCheritonCS
as Assistant Professor and
@VectorInst
as Faculty Member in 2024! I am *very* excited to be back in
@Canada
to help grow the Canadian AI ecosystem! Please apply if you are interested in a PhD at the intersection of NLP and ML!
Dear 2020 PhD students: what you're doing is nothing short of amazing. It is hard to start a PhD/collaborate with new people virtually/move to a new city/participate in a new community virtually during a pandemic. It will get better! We're all proud of you! Wish you all the best!
I wrote a guide on how to get started in NLP/ML research if you do not have a background in AI. If you are an undergraduate student looking to get into research, I hope this helps you!
I am hiring strong PhD students in ML and NLP at the University of Waterloo to start in 2024. This is an excellent opportunity to be a part of a vibrant new NLP group w/ 5 professors. Please see more details here: . Deadline is Dec 1!
Our
#emnlp2020
reviews were largely constructive and detailed - I barely squeezed in my response at 900 words after responding to each question! If you are starting out in research and are horrified by academic Twitter please rest assured that there are also great reviewers!
Super excited to share our
@iclr_conf
2020 work RTFM: Generalising to Novel Environment Dynamics via Reading ()! Joint work w/
@_rockt
and
@egrefen
at
@facebookai
research London on RL policies that generalise to new envs via reading! 1/3 👇
The Dockerized
@PyTorch
implementation of our
@acl2018
paper Global-Locally Self-Attentive Dialogue State Tracker () is now available on Github ().
#nlproc
I am on the job market this year for academic and research positions - please share broadly!
My research is on reading to learn: instead of learning specific problems, how can we learn to interpret language to generalize to new problems?
More at
#nlproc
Our latest reading to learn paper Language Dynamics Distillation will appear at
#NeurIPS2022
! In LDD, we pretrain the agent to read to model env dynamics. LDD improves generalization on 5 distinct language grounding envs over naive RL, VAE, inverse RL. 🧵
Many prior works in language grounding study single environments. How do we build unified techniques that apply across multiple environments? Our
#NeurIPS2021
paper proposes the multi-environment Symbolic Interactive Language Grounding benchmark (SILG).
Our
#emnlp2020
paper Grounded Adaptation for Zero-shot Semantic Parsing proposes GAZP, a framework for zero-shot language-to-SQL parsing by synthesizing data in new DBs after reading their schema. 💡
Paper 📰
Thread 👇1/9
#NLProc
#MachineLearning
Our
#acl2019nlp
paper Entailment-driven Extracting and Editing for Conversational Machine Reading (E3) is out! E3 studies conversational machine reading (CMR), a task oriented dialogue problem where reasoning rules aren't fixed but implied by procedural text. Paper/code👇1/10
One thing that I really appreciate is how so many junior faculty paid for students w/o presentations to come to
#NAACL2022
. Meeting & getting to know my peers has been one of the most rewarding aspects of academia. COVID has deprived a generation of students of this opportunity.
Fantastic work by Victor Zhong (
@hllo_wrld
) during his internship at
@FacebookAI
Research London. Procedurally generating environment dynamics and their textual descriptions forces the agent to perform multi-hop reading comprehension.
Paper: w/
@egrefen
Reading reviews from co-reviewers and I just want to say: not every component of the model must be new. Not every single model must be novel. This is one of the frustrations we must deal with with neural nets because there are infinite ways of combining differentiable functions.
Thanks
@Apple
! I am honoured and grateful for this fellowship. I would not have received this fellowship without support from my awesome advisor
@LukeZettlemoyer
and the excellent research institution that is
@uwcse
.
I will be at
#NAACL2022
in Seattle and giving my first in person talk in 2 years! Message me if you want to meet up! The talk is titled “Reading to Learn”, happening Friday July 15 2pm at the Multimodal Workshop. I’m also going on the job market this year! Let’s chat!
#nlproc
I am on the job market this year for academic and research positions - please share broadly!
My research is on reading to learn: instead of learning specific problems, how can we learn to interpret language to generalize to new problems?
More at
#nlproc
So I finally tried
@raydistributed
- it is **fast**! Built a simple lib to preprocess data. Here's an example of parallel parsing with
@stanfordnlp
Stanza. I'm seeing 2x speed on laptop w/ toy text + 3 procs. Different large job on server is 100x faster!
Our work on Reading to Learn, along with terrific work from
@AnimaAnandkumar
and
@karthik_r_n
was recent featured in
@QuantaMagazine
!
By far the most professional interactions I've had with a news org - a lot of work put into fact checking+editing.
OMG I finally got my first primary contributor patent approved!! On the one hand this is pretty mundane but on the other hand it feels like one of those immigrant story moments. 😅 Cheers to the good times
@CaimingXiong
@RichardSocher
@SFResearch
!
Fantastic to have
@Thom_Wolf
speak at
@uwnlp
today! Many exciting things happening at
@huggingface
- BigScience is such a massive and important undertaking and I cannot wait to see what will happen! Also it has been so exciting to see the growth of
@huggingface
over the years!!
You know what feels better than getting good convergence? Getting good annotations. Building a dataset right now and boy do I enjoy seeing these good annotations come in.
I,
@sewon__min
, and
@RichardSocher
will be presenting two posters tomorrow at
@acl2018
on dialogue state tracking ( ) and efficient question answering (). Come check out our work at 12:30!
#nlproc
The paper and videos for Language in Reinforcement Learning
#ICML2020
workshop is now online! Check it out at . Also please join us this Saturday for exciting invited talks and posters!
If you work in
#nlproc
I highly recommend that you design some sort of project that involves raw Wikipedia or wikidata. You really come face to face with this vast accumulation of human knowledge and it’s honestly very humbling despite all its flaws.
There are no shortcuts to learning. Inspirational YouTube videos do not replace 10 years of practice. The ML “education” hype reminds of the predatory iOS and mobile game bootcamps from the 2000’s, except now global instead of relatively confined to the Bay Area.
So in
@sirajraval
's livestream yesterday he mentioned his 'recent neural qubit paper'. I've found that huge chunks of it are plagiarised from a paper by Nathan Killoran, Seth Lloyd, and co-authors. E.g., in the attached images, red is Siraj, green is original
I’m happy to chat at
#emnlp2020
about my research on reading to generalize as well as
@uwnlp
! Feel free to reach out on rocket chat or email. I find these chats especially helpful at virtual confs. I’m currently in two productive collaborations that resulted from these meetings.
Had super exciting discussions about reading to learn + language grounding at
#NeurIPS2022
! I’m at
#emnlp2022
Dec 7-12 for works led by
@machelreid
,
@TianbaoX
,
@ChenHenryWu
. Please DM or email to chat! I am on the job market and would love to discuss opportunities as well!
I will be presenting Coarse-grain Fine-grain Coattention Network for Multi-evidence Question Answering
@iclr2019
on May 9th 11am-1pm. Hope to see you there! The 6-month-old CFC is still the top public model!
paper:
leaderboard:
@stanfordnlp
@StanfordAILab
I am actually so excited about this. This has been discussed and in the works for soooooo long. People waaay under-appreciate the effort required to maintain large software tools for so many years in a graduate research lab with students who rotate in and out of the program.
#NAACL2022
stragglers: some personal favourites for spending the weekend in Seattle, a thread. Places: Ballard, Ballard Locks, Discovery Park, Golden Gardens, Myrtle Edwards, Gas Works, Green Lake, Arboretum, Carkeek
The TACRED dataset—person/organization relations for training relation extraction systems—is available from LDC! (At last—sorry for the delays, etc.; $25 for non-members.) See more in our papers: and
While in London this summer, I’ve been using a British keyboard and have finally changed my ze’s to se’s. No longer will reviewer 2 have to point out that this confused Canadian keeps mixing American spelling with British spelling.
I will be presenting our work E3: Entailment-driven Extracting and Editing for Conversational Machine Reading today at 4pm at poster
#7
. Please come check it out!
#ACL2019
#nlproc
@ACL2019_Italy
Our
#acl2019nlp
paper Entailment-driven Extracting and Editing for Conversational Machine Reading (E3) is out! E3 studies conversational machine reading (CMR), a task oriented dialogue problem where reasoning rules aren't fixed but implied by procedural text. Paper/code👇1/10
I finally got access to DALL-E. Late to the party - I know. First impression is that it generates amazing images, but it doesn't really understand compositional language. Here are "a woman proposing to a man" and "a car on top of a man". Great progress, but lots more to do.
Machel
@machelreid
is one of the most productive people I have collaborated with. He should definitely be in your top 18 under 18 list 😉 PhD programs: you may want to start your recruiting process now!
I'm happy to announce that "LEWIS: Levenshtein Editing for Unsupervised Text Style Transfer" has been accepted to
#ACL2021NLP
(Findings)! This would have not been possible without my amazing co-author
@hllo_wrld
! Preprint will be out soon!
Our
@iclr_conf
poster sessions will be this Thursday 5-7am GMT / 10pm-12am PDT during Session 1 and 5-7pm GMT / 10pm-12am PDT during Session 4. More details here:
Hope to see you there!
Super excited to share our
@iclr_conf
2020 work RTFM: Generalising to Novel Environment Dynamics via Reading ()! Joint work w/
@_rockt
and
@egrefen
at
@facebookai
research London on RL policies that generalise to new envs via reading! 1/3 👇
Happy holidays!! We’ve successfully escaped to sunny LA. Our
@AlaskaAir
flight got cancelled and then fortunately uncancelled (didn’t know this was possible) 😅 Super glad to be out of the
#seattleicestorm
I find it difficult to believe that
@Tim_Dettmers
still manages to improve this already-terrific post for choosing a PhD program. I think every answer I've ever given to "what advice do you have for choosing a PhD program" is found here.
An important but elusive quality to learn in a PhD is research style. It is valuable to be aware of this before you start a PhD. Among other updates, I added an extensive discussion on research style to my "choosing a grad school" blog post. Enjoy!
Our
#emnlp2020
paper Grounded Adaptation for Zero-shot Semantic Parsing proposes GAZP, a framework for zero-shot language-to-SQL parsing by synthesizing data in new DBs after reading their schema. 💡
Paper 📰
Thread 👇1/9
#NLProc
#MachineLearning
I have updated
#WikiSQL
to include a new leaderboard for weakly supervised models! Currently the top model is MAPO by Liang,
@Mo_Norouzi
,
@JonathanBerant
,
@quocleix
& Lao. If you know of any other weakly supervised work, please let me know!
#NLProc
The worst flight ever: delayed at
@Gatwick_Airport
for 10 hours by 30 minute increments, when
@vueling
knew that the flight would have been late by at least 5 hours.
Our
#emnlp2020
paper Grounded Adaptation for Zero-shot Semantic Parsing proposes GAZP, a framework for zero-shot language-to-SQL parsing by synthesizing data in new DBs after reading their schema. 💡
Paper 📰
Thread 👇1/9
#NLProc
#MachineLearning
Before I start at
@UWCheritonCS
, I will be in NYC with
@MSFTResearch
as a postdoc, where I will work on RL + NLP. If you are in NYC or Toronto and want to get in touch, please reach out!
Great work by
@jayelmnop
!!! Language descriptions of state space seems like a promising way to remove noise and reduce spurious correlations. Super excited about more work in this area!
@hmkyale
@YaleSEAS
@dragomir_radev
@Yale
@YINSedge
@YaleCompsci
Very sad & sorry to hear this… I had corresponded with Drago as recently as 10 days ago regarding using figures from my work for his book. We then caught up and he wished me good luck for my job search. He has always been so kind and gracious, especially to junior researchers.
I gotta say the
#emnlp2020
auto caption is pretty good! however it seems to never get the word "parser". There are also very very rare & funny problems like this one. For the record, no LSD was purchased as a part of our work on GAZP .
#NLProc
2. Multi-hop Reading Comprehension through Question Decomposition and Rescoring w/ the wonderful first author
@sewon__min
,
@LukeZettlemoyer
, and
@HannaHajishirzi
, where we propose a new technique for multi-hop QA by decomposing natural language questions.
We are excited to host
@hllo_wrld
from
@uwnlp
for an in-person talk titled "M2RQA: A benchmark for multi-evidence, multi-answer, robust QA". Join us on Wednesday Sep 14 at 12pm @ ICICS X836!
@UBCLangScis
@CAIDA_UBC
@UBC_CS
@BMarcusMcCann
@chrmanning
I hope I don't misquote but: empirically, some of the most influential research are produced by universities; it does not seem the case that good research necessarily requires better tech+more resources, otherwise universities would have ceased to exist a long time ago.
I have had a wonder 5 years at
@uwcse
. I am tremendously grateful to my advisor
@LukeZettlemoyer
for his help in my journey - I could not have imagined a better advisor. I also want to thank
@uwnlp
and my office mates in CSE318. I hope our paths cross often in the future!
1. E3: Entailment-driven Extracting and Editing for Conversational Machine Reading w/
@LukeZettlemoyer
, where we achieve SOTA on
@uclmr
's conversational machine reading task () by modeling latent rules in text & entailment through dialogue.
Talk to your databases -- Photon 🌌 v1.1 is here!
What's new?
- A dual input mode that accepts both NL and SQL queries
- Enhanced GUI with more test databases
- New blog post
Blog post:
Live demo:
ArXiv:
How do we turn natural language instructions💬 into programs executable by machines 🤖? Join us at the first workshop on interactive and executable semantic parsing (IntEx-SemPar) Nov 19 at
@emnlp2020
! Call for paper at ! Deadline Aug 14🗓️
People usually get information from others in a multi-turn conversation. To approach this, we’ve released CoQA 🍃—A Conversational Question Answering Challenge by
@sivareddyg
•
@danqi_chen
•
@chrmanning
. 127K Qs— free-form answers—with evidence—multi-domain.
After living here for a decade I still don't understand America. Why is the state to be able to regulate someone going to the hospital to do something to THEIR OWN BODY and not regulate weapons in public? Where is the boundary b/w fed and state and individuals?