Share this page!

Enter URL or ID to Unroll

You can paste full URL like: https://x.com/threadreaderapp/status/1644127596119195649
or just the ID like: 1644127596119195649

How to get URL link on X (Twitter) App

  1. On the Twitter thread, click on or icon on the bottom
  2. Click again on or Share Via icon
  3. Click on Copy Link to Tweet
  4. Paste it above and click "Unroll Thread"!
  5. More info at Twitter Help
(((ู„()(ู„() 'yoav))))๐Ÿ‘พ Profile picture
(((ู„()(ู„() 'yoav))))๐Ÿ‘พ
@yoavgo
Jun 29, 2023 โ€ข 5 tweets โ€ข 1 min read
yes, the neural LMs learn a model of the world as projected through documents (and soon also images) found on the internet. the big question is to what extent this projection---even in the best case---provides an accurate or complete representation of the real world. here the @emilymbender camp says "nothing whatsoever", and i and others say "to some extent, and certainly enough to be useful", but i don't think there is any doubt that its very far from being complete or accurate.
Dec 23, 2021 โ€ข 6 tweets โ€ข 2 min read
for me there are two, both from my CS undergrad, both were the first assignment in a course, both involve programming, and both beautifully captured the essence of the course in a single, simple to explain and somewhat open-ended assignment, that you had to figure out on your own the first is in the Compilation course by Prof. Mayer Goldberg (no family relation): we had to write an interpreter for a simple-but-turing-complete language (easy), and then we had to write a compiler from high level code to this language.
Dec 9, 2021 โ€ข 6 tweets โ€ข 2 min read
ื—ืœืงื›ื ืื•ืœื™ ื”ื‘ื—ื ืชื ืœืคื ื™ ื›ื—ื•ื“ืฉ ืฉ ynet ื”ื•ืกื™ืคื• ืื•ืคืฆื™ื” ืœื”ืงืจืื” ืฉืœ ื›ืชื‘ื•ืช ื‘ืขื‘ืจื™ืช, ื•ื–ื” ืืคื™ืœื• ืขื‘ื“ ืžืžืฉ ืœื ืจืข.

ื”ื”ืงืจืื” ื”ืชื‘ืกืกื” ื‘ืžื™ื“ื” ืจื‘ื” ืขืœ ื˜ื›ื ื•ืœื•ื’ื™ื™ืช ื”ื ืงื“ืŸ ืฉืคื•ืชื—ื” ื‘ื“ื™ืงื˜ื”, ื‘ืคืจื•ื™ื™ืงื˜ ืžืจืฉื™ื ืžืื“ ื‘ื”ื•ื‘ืœืช ืื‘ื™ ืฉืžื™ื“ืžืŸ ื•ืคื™ืชื•ื— ืขื™ืงืจื™ ืขืœ ื™ื“ื™ ืื‘ื™ ื•ืฉืืœืชื™ืืœ ืฉืžื™ื“ืžืŸ (ื•ืžืขื•ืจื‘ื•ืช ืžืกื•ื™ื™ืžืช ืฉืœื™, ื•ืžืขื•ืจื‘ื•ืช ืฉืœ ืžืฉื” ืงื•ืคืœ) ืื‘ืœ ืœืžื” ืื ื™ ืžืกืคืจ ืœื›ื ืขืœ ื–ื”? ื”ืื ื›ื“ื™ ืœืกืคืจ ืขืœ ื”ืืชื’ืจื™ื ื‘ื˜ืงืกื˜ื™ื ื—ืกืจื™ ื ื™ืงื•ื“, ื•ืœื”ืชื’ืื•ืช ื‘ื”ื™ืฉื’ ื”ื™ืคื” ืฉืœื ื•? ื’ื, ืื‘ืœ ืœื ื”ืขื™ืงืจ.

ื”ืขื™ืงืจ ื”ื•ื ื”ืกื™ืคื•ืจ ื”ื™ืคื” ื”ื–ื”:
Sep 27, 2021 โ€ข 16 tweets โ€ข 6 min read
This paper is now finally on arxiv!

We (@yanaiela @rtsarfaty Vika Basmov) define a new core NLP task we call "Text-based NP Enrichment" (TNE), which we believe is both:

(a) very useful if performed with high accuracy,

(b) better benchmark for reading comprehension than QA is. The task definition is very simple: for every pair of base-NPs in the text (in our dataset a text is ~3 paragraphs long), decide if they can be related by a preposition, and if so, which.

Why is this task interesting? We argue that its a core component of reading comprehension.
Sep 25, 2021 โ€ข 4 tweets โ€ข 1 min read
so, what do i think?
lets start w/ a disclaimer: i'm not active in summarization research, and only skimmed the paper, focusing on the eval part.

i think not using ROUGE but rather a human eval (for the main metric) is a nice step forward from a visible player like OpenAI, but, but, it is also the really bare minimum of an eval. and it is far from being a good one (for starters, hardly any details are given re eval guidelines, what were the evaluators were instructed to eval). this is sort-of excusable here, since models are so far from human level,
Aug 29, 2021 โ€ข 10 tweets โ€ข 2 min read
a bit more on this: "oh the new large DL models in NLP are so soul-less, they only consider form and don't truly understand meaning, they are black-boxes, they expose and amplify sociatel biases in the data, etc etc etc": well, all true, but at least they work. like, previous-gen models *also* didn't understand meaning, and *also* considered only form. they were just much worse at this. so much worse that no one could ever imagine that they capture any kind of meaning whatsoever. they didn't work.
Aug 27, 2021 โ€ข 23 tweets โ€ข 4 min read
my two cents on why NLP as a field is focusing on the ML-ish / algorithmic / leaderboard-ish aspects (incl., now, LLMs) and not on the underlying language phenomena: it is just so much easier, on so many levels. i'm not talking about "the ease of getting through the review process", but about the ease of acquiring knowledge, the ease of formulating research questions, the ease of conducting research, the ease of measuring your own progress, the ease of "getting results".
May 26, 2021 โ€ข 15 tweets โ€ข 3 min read
i keep getting RTs of this NYT opinion piece (like, i saw it ~30+ times in my TL already?), and while i only skimmed the content itself (which is for the most part correct), i want to say a few words about the misleading accompanying "maps" illustration,

nytimes.com/2021/05/25/opiโ€ฆ i can see how this illustration fits well with the palestinian narrative. and all the maps are correct. but their arrangement in a sequence, and their context, is misleading. and i really hate misleading infographics. so, here is some context:
Feb 19, 2021 โ€ข 12 tweets โ€ข 6 min read
i love this! and spent the past 15 minutes finding a bunch more.
Dec 4, 2020 โ€ข 5 tweets โ€ข 1 min read
the paper is well written, but also pretty benign. its an "opinion piece" kind of paper, and it doesn't say anything we don't know already. it doesn't reveal any google secrets, and could have just the same been written outside of google. and google wouldn't have cared. at the same time, i can def see why lawyers at google don't want to see it published *with google affiliated authors*. it can do serious harm to the company. when these things are being said by outsiders, they are easy to brush off. but when it is signed by your own scientists...
Sep 8, 2020 โ€ข 12 tweets โ€ข 3 min read
We've just had a virtual local conference (ISCOL2020, the Israeli NLP event), over the gather.town interface. The event was small-ish compared to xACL events (150 participants logged in at once), but I really enjoyed the conference experience it provided. I didn't take any screenshots because I was too busy interacting with people, but here is a screenshot by @zehavoc who visited, and which was taken pretty much after everyone left :) (15 minutes after the last session)
Jul 18, 2020 โ€ข 39 tweets โ€ข 12 min read
GPT-3 update: I got contacted by @gdb who sent me an invite! excited to try it out. thanks Greg! initial attempts: very impressive QA results (check out the coref in the gates questions!) but also has some glitches. Image
Jul 13, 2020 โ€ข 5 tweets โ€ข 1 min read
would be interesting to compare view counts to citation counts few years down the line. i suspect only weak correlation? idk. for example i suspect the two winning theme papers will be generically-cited *a lot*. but the next three highly watched theme papers (Henderson, Trott et al, Dunietz et al) probably not so much, despite being great papers. they are just less citable, each for its own reason.
Jul 12, 2020 โ€ข 6 tweets โ€ข 2 min read
Here goes.
The answer is 66 (according to data collected few hrs ago). The vast majority of responses under-estimated.

@ethayarajh suggested to break it down by tracks, and presented results for Ethics, Summarization and Theme, where the median is 104. He chose good tracks. Overall, there were 74662 views of 907 videos. This is very nice! but we can see that the views are not evenly distributed: most works received far fewer views than the 82.2 avg.
There is large variability between tracks, and within tracks, with some "hit" papers in each track.
Jul 10, 2020 โ€ข 4 tweets โ€ข 1 min read
oh wow, watching percy liang's talk live, without the ability to turn of the live captions for some reason (i think i am seeing the slideslive presenter's slides..) --- and wow are they distracting. (and bad. very bad.) so intelligent 16

this Kun B right.
Jul 10, 2020 โ€ข 4 tweets โ€ข 2 min read
Interesting one.
- @HengJi1 is a friend and colleague, who has been working on IE for a long time. I also work on IE.
- IE is a classic case of dual-use tech.
- ACL also gives best-papers awards to MT, which is also clearly military tech.
- It is OK for IE and MT to be awarded. Having said that, I was also surprised about this award, given the army affiliation and explicit use case, and given the current events and climate. The committee was either very brave or very oblivious on this one, not clear to me which.

the work itself is strong and worthy.
Jul 7, 2020 โ€ข 4 tweets โ€ข 3 min read
And now, what will we be presenting tomorrow at #acl2020nlp? three papers (thread) Work by former lab member @roeeaharoni (with very little involvement by me, i must say ;) ) on emergent domain clusters in pre-trained LMs and how we can use them in NMT:
virtual.acl2020.org/paper_main.692โ€ฆ
Jul 6, 2020 โ€ข 4 tweets โ€ข 3 min read
Today at #acl2020nlp is not over, but here are our sessions for tomorrow: Alon's thoughtful comments on what it means to do interpretation "right", and what it means for an interpretation to be faithful.
Jul 5, 2020 โ€ข 5 tweets โ€ข 4 min read
A bunch of works from my group(s) coming up at #acl2020nlp tomorrow. Watch the videos and come visit us in the Q&A sessions! In work with @lambdaviking @gail_w @royschwartz02 @nlpnoah and @yahave we provide *theoretical* results (yes, with proofs) of things that can and cannot be represented by various kinds of RNNs, and under what conditions.
virtual.acl2020.org/paper_main.43.โ€ฆ

+ blog:
lambdaviking.com/post/rr-hierarโ€ฆ
Jul 4, 2020 โ€ข 4 tweets โ€ข 1 min read
ื”ืกื•ืคืดืฉ ืœืžื“ืชื™ ืขืœ ื•ืฉื™ื—ืงืชื™ ืขื sonic pi, ืฉื–ื• ืฉืคืช ืชื›ื ื•ืช ื™ืขื•ื“ื™ืช ืœืชื›ื ื•ืช ืžื•ื–ื™ืงื” (ืœืžืขืฉื” dsl ืžืขืœ ืจื•ื‘ื™) ื•ื–ื” ืžื’ื ื™ื‘ ืœืืœืœื”. ื™ื—ื“ ืขื processing, ืฉื”ื™ื ืกื‘ื™ื‘ืช ืคื™ืชื•ื— ื™ืขื•ื“ื™ืช ืœืฆื™ื•ืจื™ื ื’ืจืคื™ื™ื ืื™ื ื˜ืจืืงื˜ื™ื‘ื™ื™ื, ืืœื• ืฉื ื™ ื›ืœื™ื ืขื ืคื•ื˜ื ืฆื™ืืœ ืื“ื™ืจ ืœืœื™ืžื•ื“ ืชื›ื ื•ืช ืœืžืชื—ื™ืœื™ื / ื ืขืจื™ื / ื™ืœื“ื™ื. (ื•ื’ื ื›ื™ื™ืฃ ืœืžื™ ืฉืžืชื›ื ืชื™ื ื›ื‘ืจ) ื–ื” ื“ื™ ืžื“ื”ื™ื ืžื” ืืงืจืื™ื•ืช ืžื‘ื•ืงืจืช ื™ื›ื•ืœื” ืœื™ืฆื•ืจ, ื’ื ื‘ืจืžื” ื”ื•ื•ื™ื–ื•ืืœื™ืช, ื•ื”ืจื‘ื” ื™ื•ืชืจ ืžื–ื” ื‘ืจืžื” ื”ืžื•ื–ื™ืงืœื™ืช. ื•ืื™ืš ืฉื™ื ื•ื™ ืฉืœ ืคืจืžื˜ืจื™ื ืคืฉื•ื˜ื™ื ื™ื—ืกื™ืช ืžืฉืคื™ืขื™ื ืขืœ ื”ืื•ืคื™ ืฉืœ ืžื” ืฉืžืชืงื‘ืœ, ื•ื›ืžื” ืงืœ ืœื’ืœื•ืช ืืช ื–ื” ื›ืฉื™ืฉ ืฉืคื”/ืกืคืจื™ื”/ืกื‘ื™ื‘ื” ืฉืชื•ื›ื ื ื” ืœืฉื™ื ืืช ื”ื“ื•ืžื™ื™ืŸ ื‘ืคืจื•ื ื˜.
Jul 3, 2020 โ€ข 4 tweets โ€ข 2 min read
This visualization is fascinating. Highlighted are paper I co-authored. Look how spread out they are! I am all over the NLP landscape (well, in the center of it at least)!

...but then Iooked at the tracks. --> Image here is machine translation and question answering. ImageImage