I’ll be at
#NVIDIAGTC
this week! Looking forward to talking with innovative startups, especially at seed/early-stage seeking investors. Also excited to discuss VC investment in AI. If you’re applying AI in the healthcare/medical domain too, I’d love to chat and exchange ideas.
My first neural model for signs classification as part of the programming assignment of
@AndrewYNg
's Specialization 😀 %77 accuracy and it classifies my own hand sign correctly. Good for now!
#DeepLearning
#NeuralNetworks
@stanfordnlp
Stanza's POS tagger is indeed more accurate than CoreNLP. There are cases where VERBs cannot be correctly tagged, which itself results in inaccurate dependency parsing. One example...
#NLProc
🚀 Presenting our work at
@csrr_workshop
today:
💫 Knowledge-Augmented Language Models for Cause-Effect Relation Classification.
Paper:
Code:
#ACL2022
#NLProc
Here in Seattle for NAACL 2022! Looking forward to having great conversations and making new friends ☺️
Also, if you’re working on Biomedical and Clinical NLP, I’d love to talk.
#NAACL2022
PS. ~2k in-person participants!
Also thanks to the organizers of Narrative Understanding for an amazing workshop this year. And, to Yejin Choi (
@YejinChoinka
) for the great talk and for bringing some joy to us with her cheeseburger stabbing example!
#NAACL2022
لغتنامه لکسی پرس رو به صورت عمومی روی گیت هاب قرار دادیم. لکسی یک لغتنامه فارسی شامل بیش از ۴۰۰۰ مجموعه ترادف از مجموعه فارس نت ۲ هست که به صورت دستی نشانه گذاری (مثبت، منفی و خنثی) شدند. این مجموعه قبلا فقط با ارسال ایمیل در دسترس بود.
✨Summary of some findings from the "Capabilities of
#GPT4
on Medical Challenge Problems".
Encouraging results, and there's still a lot to think about before successfully using GPT-4 in real-world clinical applications.
Preprint:
(1/9) Thread 🥁
It was very nice talking to Christopher Manning and having his invaluable feedback on my research at The Science of Deep Learning Colloquium here in Washington, DC. It was also fun listening to Rodney Brooks's lecture on the Steps Toward Super Intelligence.
#AI
#NLP
[1/3] There are many good but scattered resources for causal relation extraction/classification in text. I am trying to unify these resources using a simple machine-readable format, CREST🚀, to make them easier to use.
#NLProc
👇👇
📄GitHub repo:
We are thrilled to announce a brilliant addition to the Lavita Advisory Council -
Dr. Saeed Hassanpour, Professor of Biomedical Data Science, Computer Science, and Epidemiology, Founding Director of Dartmouth Center for Precision Health & Artificial Intelligence (CPHAI) at
Delighted to share ‘ParsiNLU’ which is the result of a great collaborative work appearing in TACL'21! ParsiNLU is a suite of Language Understanding Challenges for Persian that we hope will be a step towards making progress in multilingual/cross-lingual NLP.
#NLProc
Excited that our big collaborative effort, "ParsiNLU: A Suite of Language Understanding Challenges for Persian" will appear in TACL'21!
If you're working on multilingual/cross-lingual NLP, give it a look!
Paper:
✨ Quick insights related to health/medical evaluation from the Claude 3 report:
There are two medical/health-related benchmarks in the evaluation: PubMedQA, a biomedical QA dataset, and MMLU, a commonsense reasoning benchmark that includes a task in the health/medicine domain.
This is a very good and clear example of why it's not correct to assume that credible news sources always publish true news! We also need to figure out what makes such false news headlines and articles compelling.
#FakeNews
#misinformation
#disinformation
#AI
[1/2] “Extension is not Explanation”: This is a short story on what happened to some ACL submissions yesterday that does not sound fair. I’d be happy if
@aclmeeting
can provide an explanation of the reasoning behind their decision.
#NLProc
#ACL2021NLP
A great talk about the history of the internet by one of its fathers, Vint Cerf, at GWU. He mentioned misinformation and disinformation as very hard problems to deal with and part of the unfinished business of internet.
#misinformation
#disinformation
#internet
@deliprao
Got a mentoring request from a brilliant **sophomore** at a high school three years ago, and they published their work as the first author at NLP4IF! 🙂
"To Build Truly Intelligent Machines, Teach Them Cause and Effect." Judea Pearl. It is inspiring, encouraging, and insightful specially when your research is on extracting causal relations from natural text.
@yudapearl
#AI
#NLP
So sad to hear that the beloved Patrick Winston passed away today. Two years ago, he kindly gave me the opportunity to meet him at his office at MIT. That meeting for me was not just about AI and his great work, but all about inspiration and encouragement. RIP dear Patrick.
به ن��ر میاد چَت جی پی تی هنوز یه خرده کار داره برای شعر فارسی گفتن. شعر به کنار، دیوید خوارزمشاه؟! 😄 ولی معلومه آهنگهای فارسی رو شاید خوب بتونه دنبال کنه:
«برای آرزوهای شاد... و برای خواب شادی که تو بودی»
#ChatGPT
#Persian
Great effort by
@Malikeh5
👏 Any thoughts or feedback on how we can make this list better? Is any dataset missing? Feel free to comment and reach out!
#MedTwitter
🚀 Exciting news! Our team
@LavitaAI
has curated a treasure trove of medical question-answering datasets on
@huggingface
, now available for the entire community. 💊🏥🩺
Link:
Preprint of our accepted paper to NLP4IF is on arXiv now. We investigated semantic/linguistic features in fake and satirical articles to better understand the differences between these two content types. We shared some preliminary results of our ongoing research in this paper.
11/17/2018 marks the day that I pitched my idea for running a business for the first time in a competition. It was great meeting people as passionate as me about entrepreneurship. Lots of lessons to learn. Thanks GWU for such an amazing opportunity.
Extending the joy of EMNLP '22 into 2023, I will be delivering my Keynote for the conference on Monday, Jan 9, at 7am PST. Title: Toward Responsible NLP: Walking the Walk. Tune in and spread the word.
نتایج اولیه تحلیل واکنش کاربران فارسی توییتر، عمدتا ایرانیان، به #کرونا با استفاده از پردازش زبان طبیعی بروی بیش از ۵۳۰۰۰۰ توییت فارسی:
نسخه اولیه مقاله:
کدها و گزارشات بروی گیت هاب (به همراه شناسه توییت ها):
#COVID19
#Iran
Topic analysis of 50,000 Tweets in Persian/Farsi on
#COVID
ー19 over past few days [More data, graphs and full report coming soon] (1/4)
#COVID19
#Iran
#كرونا #کرونا_ویروس
It was fun listening to a panel of entrepreneurs talking about their life before and after the Shark Tank. Good to know what exactly is going on behind the scenes of the show! 😀
#Entrepreneurship
#GWU
#SharkTank
It was great to have an opportunity to present our paper at
#sbpbrims
. In our paper, "Does Causal Coherence Predict Online
Spread of Social Media?," we tested the hypothesis that causally coherent news documents, including
#FakeNews
, will be shared more by individuals online.
@mark_riedl
GPT3 not only doesn't lecture me on semantic understanding but also answers me in a none of your business kind of way today!
Q: John mowed his lawn. Why did John do the mowing?
A: John mowed his lawn because he wanted to.
🚀 Excited to represent
@LavitaAI
as a sponsor at this year's
#ACL2023NLP
Clinical NLP Workshop!
Let's connect if you're in
#AI
or
#NLP
in the clinical/biomedical space 🙂
We are excited to announce that Lavita AI will be sponsoring the Clinical Natural Language Processing Workshop at the 61st Annual Meeting of the Association for Computational Linguistics (ACL) in July 2023 ().
ACL is the premier international scientific
[1/2]
#Iran
was among the countries that were hit hard in the first wave of the
#COVID19
spread. As the situation was unfolding, we started analyzing Persian/Farsi tweets to gauge the reaction of users (Iranians mainly) to this unprecedented event using NLP...
I present my ongoing work about “Spread of Coherent Stories on Social Media” today at INSNA 2018. In my work, I test the hypothesis that more coherent news articles, fake or real, are more likely to be shared by people on social media.
#FakeNews
#ArtificialIntelligence
#INSNA2018
Great and insightful panel discussion yesterday at
#NAACL2022
about "The Place of Linguistics and Symbolic Structures".
We had TeamChris vs. TeamSmith at the Oscars, and it was TeamChris vs. TeamEmily at NAACL 😃
A pretty interesting work by a mathematician friend on representation learning via a category-theoretic approach called "Categorical Representation Learning (CRL)":
"Morphism is All You Need"!
#AI
#NLProc
رشته ۱/۴: همزمان با ظهور ویروس کرونا در دنیا و از جمله ایران، شبکه های اجتماعی هم مملو از اطلاعات مربوط به اون هست. تصمیم گرفتیم که تحقیقاتی رو بروی توییت های فارسی انجام بدیم تا درک بهتری از فضای موجود در میان مباحث پیرامون کرونا پیدا کنیم. #کرونا #کرونا_ویروس #كروناويروس
I'm not
@emnlp2019
but my advisor Dr. Mona Diab presented our work on identifying satire vs. fake news. Reach out to her and chat if you want to learn more about what we do. And, please check out our paper at:
#emnlp2019
#NLProc
The number of
#COVID19
related tweets posted by Farsi/Persian speaking users (mainly
#Iranians
) till Nov. 15. Overall, there's a decrease in the number of tweets with some temporary jumps in early May, July, and October. We're running an analysis to find the discussed topics...
One take away from today's tutorial about bot-like behavior on Twitter: We need to pay a closer attention to messages and content not just the medium. Fighting against
#FakeNews
and dis/misinformation needs both computational and social insights.
#sbpbrims
#GWU
Great talk and insights by Mark Russinovich
@markrussinovich
at
@StanfordHAI
, from
@Azure
AI Infrastructure to AI research and machine "unlearning"!
Mark also shared some internal secrets about their AI supercomputer's size compared to previous versions: "Even bigger!"
I'll be at the
#JPM24
in San Francisco! Excited to connect and learn, share insights, and dive into the latest in healthcare innovation. If you're around, let's catch up and chat!
#JPM2024
We have also just released a new set of over 1,400,000 Persian/Farsi tweet IDs related to
#COVID19
. Please check out our paper and reach out with any feedback/questions (will be presented at
#emnlp2020
):
Paper:
GitHub repo:
[2/2] We are sharing the preliminary results of our ongoing study and analysis of
#COVID19
related Persian/Farsi tweets:
Preprint:
GitHub repo:
Get in touch if you have questions or comments about any specific aspect of the work.
I am wondering if we can call it a contribution to a field when someone says that they created a new dataset using blah blah blah methods and explicitly mentioning that it is their contribution but not publishing the dataset?!
#contribution
#confused
Presenting my work in an interdisciplinary research competition and I'm the only male presenter and the only male in the room with judges and moderators included.
Kudos to all of these amazing female researchers for their great work, I just feel a bit lonely! 😃😄
@GWtweets
Long overdue 😅, but here we go: we just made LexiPers, a sentiment lexicon for Persian, publicly available (it was previously available upon request). You can find it at:
#Lexicon
#Persian
#NLProc
Interesting read on how Google is fighting misinformation. At first glance, it seems that there is not much about features of "content" and there is still a lot of room for improvement.
#FakeNews
#misinformation
@google
Glad I had the opportunity to be at
@nlpnoah
's talk today at Georgetown, thanks to
@complingy
. It was interesting to know Noah's perspective on domains and microdomains in text classification. I'm also curious about the Domain Adaptive and Task Adaptive pretraining in my tasks!
پیکره احساس سنتی پرس رو هم به صورت عمومی روی گیت هاب در دسترس قرار دادیم 🙂. نسخه فعلی شامل بیش از ۱۵۰۰۰ جمله منحصر به فرد برچسب گذاری شده هست. امیدواریم هرچند جزیی، بتونه کمکی برای علاقمندان به تحلیل احساس در زبان فارسی باشه.
Course
#2
, done! Thanks to
@AndrewYNg
. Learned a lot on optimization algorithms, regularization, hyper-parameter tuning, and normalization. Excited about the CNN course and Sequence Models!
#DeepLearning
#NeuralNetworks
"...we run different random seeds for each model and report the mean of the *non-degenerate runs* for each model, which has higher than 80% of accuracy in the training set..."
Is this a standard practice that I didn't know about?! 🤔
#NLProc
Thanks to all who joined us, for great questions and discussions. GisPy, is an open-source tool for measuring gist inference score in text. Please check our paper and GitHub repository for more details.
Paper:
GitHub:
#NAACL2022
@pnazemi
جالب هست که از مدتها قبل حتی بحث پیرامون کرونا میان کاربران فارسی زبان در توییتر هم رو به کاهش بود در حالیکه که اوضاع میدانی و حتی آمارهای رسمی ارایه شده لزوما رو به کاهش نبودند و حتی ایران به پیک هم نرسیده بود. یک سری مشاهدات رو ما اینجا گزارش کردیم:
رشته ۴/۴: در تمامی گروههای معنایی بدست آمده از تحلیل و نشانه گذاری توییت ها، دو موضوع غالب در هر گروه از میان سه دسته موضوعی خبری، طنز و کنایه، و اظهار نظر هستند. توییت های سازنده و مربوط به ارایه راهکار در تمامی خوشه ها موضوع آخر به شمار میاد. #کرونا #کرونا_ویروس #كروناويروس
Good to have a new president at GWU with Computer Science background. President LeBlanc mentioned "divide and conquer" algorithm in his interview as a solution to many real world problems! So true! 😀👍
@GWtweets
@ylecun
GPT-4: "By mentioning a well-known expert's name in the prompt, the AI might associate the question with higher credibility and strive to generate a more accurate and reliable response. However, this approach doesn't inherently change the limitations of the AI model itself."
I know no one is probably watching the Golden Globes or could care less, but this was one of the most boring opening monologues in the history of Golden Globes! It was all downhill after Ricky Gervais (
@rickygervais
) stopped hosting the show!
#GoldenGlobes2023
Yet another long overdue!😃 We just made SentiPers, a sentiment analysis corpus for Persian, publicly available (it used to be available upon request). It includes more than 15,500 manually annotated sentences in Persian.
#NLProc
@aclmeeting
You need to also let authors who withdrew their papers before the deadline [assuming they can't make it] know that they *can* submit their papers again.
Who is being targeted by political ads? And what attributes political advertisers tap into when targeting users? In this publication, we took a first stab at answering these questions by exploring
@propublica
, a political ads dataset.
A great event on misinformation. I am glad I had the opportunity to attend
@misinfocon
. It was interesting and helpful to listen to people who are studying
#FakeNews
from a different perspective.
#misinfocon
#DC
(P.S. behind the scenes of getting ready for the group photo!)
نمودار مربوط به تعداد توییت های ارسالی در رابطه با #كرونا توسط کاربران فارسی زبان در بازه زمانی فروردین تا آبان ماه ۱۳۹۹. در مجموع طبق انتظار تعداد توییت ها در گذر زمان رو به کاهش بوده ولی در مقاطعی جهش هایی رو به همراه داشته که در حال بررسی اونها هستیم...
The test sentence at the end of training Emojify model as part of
@AndrewYNg
's Specialization:
"not feeling happy 😞"
Nothing wrong, however this is the sentence I want to test after completing the assignment:
"I feel awesome 😄," and it works pretty well 😀
#DeepLearning
Interesting talk by
@madhumitasushil
on exploring clinical domain knowledge of BERT models.
One insight: adding domain-specific medical knowledge to BERT (e.g. using further MLM training) did not significantly help BioBERT on MedNLI
#NAACL2021
Paper:
🧬 Bio+NLP: I just added the Benchmark Corpus for Adverse Drug Effects (ADE) to CREST.
#NLProc
A collection of 5,671 [DRUG <--> Adverse-Effect] and [DRUG <--> DOSE] relations from PubMed articles.
Take a look if you're interested in Biomedical NLP:
🥁🥳 Many congrats to our Mona Diab, the VP-ELECT of the ACL Special Interest Group on Linguistic Data and Corpus-Based Approaches to NLP (SIGDAT), very well-deserved. Thanks to all who voted for Mona.
#NLProc
"We need to be able to extend it [AI] to do things like reasoning, learning causality, and exploring the world in order to learn and acquire information." from one of the godfathers of modern AI, Yoshua Bengio on importance of creating causal models.
“‘Inclusivity’ has to be more than a word we say to look good,” says one of the fathers of modern AI, about a lack of diversity in the field of artificial intelligence.
I will be presenting my ongoing work on “Fake News on Social Media” next week at SBP-BRiMS 2018’s Doctoral Consortium in DC. Fake News problem is real and AI-based only methods are not the answer to the problem. Check out the abstr…
I sent this message on our lab's slack channel the day I read
@MSFTResearch
's blog post about Turing-NLG and
@Google
's REALM paper, feeling that I need some time to catch up! If I knew the universe is listening to me, I would have asked for something else! 😅😃
#COVID19
#NLProc
“People fail to get along because they fear each other; they fear each other because they don't know each other; they don't know each other because they have not communicated with each other.” Martin Luther King Jr.
#MLKDay
@sarahwiegreffe
@anmarasovic
Great work and interesting insights on discrepancies in
data collection. I'm also trying to unify some datasets with a focus on causality (cause-effect) relations (might be a sub-type of datasets you're working on):
🚀 New CRESTed dataset:
*BioCause*: a corpus of annotated causal relations among biomedical events.
Any dataset with "causal relations" missing? let me know!
Github:
#NLProc
Someone was asking why people always want to create their own dataset and not just focus on improving their methods using existing data resources. I personally would love not to spend time on creating my own data, but when the desired data is not available you have to create it!
President Trump is really helping me with my research! “Beautifully represented” needs to be further investigated as a feature for fake news!
#FakeNews
#AI
The White House Correspondents’ Dinner is DEAD as we know it. This was a total disaster and an embarrassment to our great Country and all that it stands for. FAKE NEWS is alive and well and beautifully represented on Saturday night!
🚀[1/2] Training trick of the day with huggingface!🙂 If you’re experiencing slow training on TPU even using pytorch XLA, you may want to check whether your inputs are dynamically padded. Dynamic padding makes TPU recompile at each step and slows down the training. Solution? ⏭️
[2/2] Summary:
@aclmeeting
did not allow authors who withdrew their papers last minute [assuming they can't make it to the deadline] not knowing about the 1 full-day extension to have an option to re-submit their papers.
#NLProc
#ACL2021NLP
@Tanmoy_Chak
@ReviewAcl
@aclmeeting
[2/2] I see a lot of Ph.D. (graduate) students who are willing to contribute and are incentivized to learn, and usually even spend more time compared to the average on carefully reviewing a *CL-level paper.