My landlord and I are exchanging ChatGPT-generated emails negotiating the rent increase. I wish we could just let ChatGPT negotiate with itself and report the decision to us.
I just published Tips for Writing NLP Papers
I wrote it for my students so I don't have to sound like a broken record (and edit papers for the same issues over and over again 😃). But some of you might find it useful too.
Last 2 days on Twitter: work 90hr weeks! See on who published the highest number of first-authored papers! Dear junior NLPers, please ignore this nonsense. Choose quality over quantity and have something other than work in your lives.
The slides for my talk "Everything you wanted to know about ChatGPT (Except for what OpenAI doesn't tell us)" are available here: . I used it for both a guest lecture in an applied ML class and for a talk for non-CS audience.
I finally prepared a comprehensive slide deck about LLMs for people outside the field, so I kindly request everyone to stop releasing tools and papers so that it stays up to date. No GPT-5 please 😁
Hinton's talk at
#ACL2023NLP
has many good points, but it bothers me that he criticizes researchers that say LLMs don't understand language as being non scientific, while he himself sees every anecdote as a proof that LLMs understand language. Science needs more nuance.
You know what's worse than packing? Adversarial packing. I'm packing and my husband takes things out of the garbage and recycling bins and asks "why are you throwing it away?".
I finally prepared a comprehensive slide deck about LLMs for people outside the field, so I kindly request everyone to stop releasing tools and papers so that it stays up to date. No GPT-5 please 😁
After 3 years of blogging about NLP, with the goal of writing in simple words that non-experts can understand, I finally dared to write about deep learning... it ended up a very long post :)
New pre-print targeting commonsense QA tasks. We propose an unsupervised framework with one LM as an answer scorer and another as a knowledge source. This is what happens when you don't panic and shut down your models when they talk with each other 😉
Implicit bias in the wild: an immigration officer upon seeing foreign names on the work permit applications incorrectly assumed my husband was the professor and I the spouse. When my husband mentioned losing his US driver's license, he asked "so who drove here?".
How well do contextualized word embeddings address lexical composition? They are good in recognizing meaning shift ("give in" is different from "give") but much worse with revealing implicit meaning ("hot tea" is about temperature, "hot debate" isn't).
Are GANs better than LMs in text generation?
@GuyTvt
and
@HabibGavriel
provide a simple way to approximate GANs to LMs which allows evaluating them using standard metrics like perplexity & BPC. It turns out GANs perform worse! (work with
@JonathanBerant
)
lazyweb: I'm looking for examples of using current NLP tools in real-world applications successfully (e.g. NMT, or Google using BERT for search) or, even more interesting, unsuccessfully (either "bag of words worked better" or "the model was deployed and caused damage" stories).
Who coined the term "causal language model"? I've seen it used to describe left-to-right (as opposed to masked) language models, but I think it's misleading given that these models were not trained for (and do not excel at) modeling causes & effects.
I gave a lot of talks in the last 3 weeks, but today was the most challenging one. I gave two 30 min talks about intro to NLP to two groups of 7th graders. I'm hope to meet some of them again at ACL 2028 😀
I'm happy to update that this paper was accepted to TACL. Final version is available here: (the main change is an additional task, light verb construction classification).
How well do contextualized word embeddings address lexical composition? They are good in recognizing meaning shift ("give in" is different from "give") but much worse with revealing implicit meaning ("hot tea" is about temperature, "hot debate" isn't).
I wrote a blog post about text generation:
If you're new followers (thanks
@kdnuggets
!): my blog is about NLP topics for non-experts, with high-level & not too technical explanations.
Me, writing a proposal: LMs confuse semantically-similar mutually-exclusive terms.
Also me, absentmindedly holding a bag of lemons: should we keep the tomatoes in the fridge?
New pre-print: pre-trained LMs treat first name mentions as references to prominent named entities ("Donald" always refers to Trump). We turn this "common knowledge" into strong empirical evidence!
We came up with the title before everyone was grounded 😬
I hate having to break a student's heart by recommending to reject their paper, just because (clearly!) their advisor was too lazy to review the submission, not to mention teach them best practices such as doing literature review. As a result, other people's time was wasted.
@yoavgo
It's adorable :) Since my parents learned what I'm working on, my dad occasionally sends me basic NLP tutorials he finds online, and I don't have the heart to tell him I can teach these topics.
This thread is sad. As a junior PhD student I used to think "I wish I was religious, it would forbid me to work on Saturday". At some point I realized I'm not less productive and I'm much happier taking weekends off. Now I'm religious about work-life balance!
I've written a python script that automates generating bib files by searching for bib entries in several sources:
It passes sanity checks, but I take no responsibility for anyone's bib files.
Are GANs better than LMs in text generation?
@GuyTvt
and
@HabibGavriel
provide a simple way to approximate GANs to LMs which allows evaluating them using standard metrics like perplexity & BPC. It turns out GANs perform worse! (work with
@JonathanBerant
)
Every talk at
#emnlp2019
:
Chair: The next speaker is X from Y university who will speak about Z.
Speaker: Hi, I'm X from Y university, I will be talking about Z.
Many thanks to CIFAR
@CIFAR_News
and Vector Institute
@VectorInst
for this recognition and opportunity, and congratulations to all the other new Canada CIFAR AI Chairs!
Finally,
@VeredShwartz
@VectorInst
@UBC_CS
is building models capable of human-level understanding of natural language, particularly focused on the implicit meaning ("reading between the lines") in human speech.
Attend our 1ST OFFICIAL IN-PERSON TALK of 2022! Join us on Thurs. 02/10/22 from 12pm-1pm @ the Eng Science Building w/ Vered Shwartz (University of British Columbia) on her talk titled, "Incorporating Commonsense Reasoning into NLP Models." Abstract/bio in our FB/IG!
I stopped myself from tweeting something snarky right after reading my ACL reviews. An hour and a slice of cheesecake later, I'm ready to express some more concrete observations :)
Idea: a
@kaggle
competition in which you get several images of Starbucks coffee cups ordered by the same person, and you need to predict the real name of that person.
Me before going on vacation:
- booking flights ✅
- packing ✅
- writing and running a bash script to utilize the lab servers for the entire time I'm away ✅
Reading an author response where each comment is an argument against my question/comment. It's exhausting and makes me regret the time I've spent helping them improve the quality of the paper. Dear fellow researchers, peer review is not a war and author response is not a battle.
Me, doing a literature review: everyone is doing the same thing.
Husband: there are more people working in your field than ideas, so according to the pigeonhole principle multiple people will be working on the same idea.
Check out our new Findings paper
#acl2023nlp
! We test LLMs on how well they can interpret noun compounds, e.g. a "chocolate bunny" is a chocolate shaped like a bunny.
1/5
A few years ago, my research areas within NLP were pretty diverse. In 2023, most of them (lexical semantics, discourse, and multiword expressions) fit into the same oral session 😲
#eacl2023
I'm in Dublin for
#acl2022nlp
. I look forward to talking research in person! I will be presenting a Findings paper and
@TuhinChakr
will be presenting our TACL paper. See details below.
I'm preparing my slides for
#emnlp2020
and I'm replacing previous "US president" examples in different examples in hopes that the prior examples will be outdated by the time of the conference in mid November🤞🤞🤞
Wizzair agent: you didn't pay to select a seat
Me: I know, assign me one
W: you didn't pay. You don't have a seat
M: I didn't pay for a *specific* seat, I'll take *any* seat
W: but you didn't pay
M: I paid for *a* seat. What else did I pay for??
W: a flight ticket
#catch22
I was curious about it too so I ran a very small-scale MTurk study on this recently. Here are a few highlights. I'll share more details later if people are interested 😀
In Polish, 3 am is night, but 4 am is morning. How is it for other languages? Is nighttime a blur or does day start between 3 and 4?🤔
#linguistics
#nosleep
2) A short paper should *not* be judged as a 4 pages long paper, but rather as a small, focused contribution. Reviewers often expect authors to solve NLP in 4 pages. And they reject the paper based on lack of experiments which are clearly outside the scope of the paper.
I'm in Canada, where snow geese fill the serene blue sky (top image), reading the horrendous news from Israel, where Hamas is filling the sky with missiles (bottom image), while slaughtering innocent civilians and kidnapping women, children, and elderly people from their homes,1/
4) It's completely OK to point out a problem without offering a solution. (I guess this goes back to the thread from a few days ago about other contributions apart from new models).
The difference in my motivation, and the pace and quality of my grant proposal writing before and after eating that chocolate chip cookie is so significant that I'm almost considering listing it as a co-applicant.
Every time I forget something I knew, search for it and find something I wrote about it (lecture notes, Quora answer...), I feel a combination of pride and shame. But I think a more comforting way to look at it is that I store most of my knowledge on the cloud instead of locally.
I'm increasingly seeing left bracket written without a preceding space. I'm suspecting it's a generational difference (I'm reading mostly text written by younger people these days...). Any ideas why this is changing? I'm assuming there's a better explanation than efficiency.
3) I don't think it's fair to ask authors to perform experiments that require more resources than most institutes can afford. (And certainly not when these experiments are in any case beyond the scope of the paper).
Someone I hardly know and don't work with sent me an email on Friday at 11:30 PM, another one on Monday at 4am ("I think you missed my previous email"), and their colleague sent an email at 5 am. Have you heard of weekends, folks. I'll get to it when I'm done with my actual work.
Git is so passive-aggressive when you accidentally commit a large file: doesn't warn you before you do, lets you go on with commits without telling you something is wrong, eventually bursts out when you push, and there's nothing you can do to fix it.
@tallinzen
@nlpnoah
The SustaiNLP workshop at EMNLP will focus on this exactly! (Except for the invite-only part, we welcome everyone! I hope the lack of exclusivity doesn't make it less appealing ;) ).
There is a special place in hell to the Vered whose email address differs in one letter from mine, who keeps mistyping her email address and signing me up to websites and mailing lists.
This video is an excellent demonstration of the state of text generation with LMs: hallucinations, inconsistencies, memorization of the training corpus, and nonsense... But the AI-generated cake "does kinda taste alright"
HR: you are all encouraged to take days off and rest!
My brain: Take a day off now and spend it staring at the ceiling? Here's a better idea. Why don't I work until I break and won't take a vacation until I can spend it sipping cocktails on the beach in an exotic destination?
I will be hosting a "Discourse and Pragmatics" birds of a feather session on Monday Nov 16 19:00-20:00 UTC / 11:00-12:00 PST. Join us in () and rocket chat ()
#emnlp2020
Are you concerned about the cost, environmental impact, and usability of large NLP models? Are you convinced that abundant data and compute are imperative for true language understanding? Help us collect questions for the SustaiNLP panel at
#emnlp2020
!
We are collecting questions for the SustaiNLP panel at
#emnlp2020
- do you have any questions to our panelists regarding their view on efficiency of NLP models?
I'm super excited about our new dataset and I look forward to see models developed to solve it! This project was fun to work on and
@_eunjeong_hwang
did amazing job😃
Check out our new work! We present the task of meme captioning and release a new dataset, MemeCap, containing 6.3K memes along with the title of the post containing the meme, the meme captions, the literal image caption, and visual metaphors. 1/3
We are organizing an AKBC workshop on commonsense and KBs (with
@ABosselut
,
@xiang_lorraine
, and
@billyuchenlin
). Consider submitting new or previously-published papers. We will also have awesome invited speakers and panelists! More details here: .
@emilymbender
Emily, your silence since this tweet is troubling. I think your intentions are good and I don't think you condone the Hamas, but I find the lack of condemnation troubling.
You're in a position of power as soon-to-be ACL president. 1/
TFW you read a paper that cites an old paper of you that does X, yet still claims "To the best of our knowledge, we are the first to propose X" (even I wasn't the first! 😄).
What annoying habits have your neighbors developed during the pandemic? My neighbor is learning to play the saxophone and she has been playing the same 7 note sequence for hours a day since March.
This seems consistent with my experience lately. We had to manually verify text written by annotators and filtered out quite a bit of text that looked LM-generated. In addition, the annotations for verification/ranking tasks were so noisy that we decided to not do them in mturk.
One of our key sources of human data is no longer fully “human"!
We estimate that 33-46% of crowd workers on MTurk used large language models (LLMs) in a text production task - which may increase as ChatGPT and the like become more popular and powerful.
Join us tomorrow at
@csrr_workshop
to learn about new research in commonsense! We have an exciting program with excellent invited speakers and lightning talks. We will be in person at the Liffey A and on zoom.
#acl2022nlp
Long paper with Ido Dagan: a model for noun compound paraphrasing (apple cake -> cake made of apples) that better generalizes for unseen noun compounds ("pear tart" has the same relation as "apple cake") and for similar paraphrases ("w2 made of w1" and "w2 containing w1").
Proud of
@ShanyBarhom
for her first paper! We jointly predict entity and event coreferences based on their predicate-argument relationship. Our model achieved state-of-the-art performance on the ECB+ dataset. We will upload the camera-ready version soon.
I'm more than happy to announce that our long paper "Revisiting Joint Modeling of Cross-document Entity and Event Coreference Resolution" got accepted to
#ACL2019nlp
@ACL2019_Italy
! 😃 with the wonderful
@VeredShwartz
and
@AlonEirew
.