As AC/author, I see an increasing trend of reviews criticizing results/approaches are not surprised. A scientific paper is not Sci-Fi. We value novel but reasonable approaches; rigor and comprehensive experiments to support the claims. Why do we expect a surprise in a paper?
If you're seeking an advisor in NLP, consider exploring options with recently appointed faculty. While established senior folks are well-known, junior folks shine as rising stars. In fact, many successful PhDs are the first few students of their advisor. Don't miss the chance!
Honored to be awarded Sloan Fellowship for our work on fairness, robustness, inclusion in Human Language Technology at
#UCLANLP
. The award belongs to my students and collaborators. Thank the NLP/ML/AI community and
@UCLAengineering
to support our work
ACL/ICML highlights threats of LMs like ChatGPT generating paper content. However, I'm more concerned about reviews. Example: ChatGPT generates confident but nonsensical reviews better than any R2. Now, it's AC's job to spot real reviews, but I lack confidence in doing so
The problem is much more beyond data bias. ML systems can be poorly calibrated () and bias in the data might be amplified (, ) i.e., simply diversifying the data source might not solve the problem.
ML systems are biased when data is biased.
This face upsampling system makes everyone look white because the network was pretrained on FlickFaceHQ, which mainly contains white people pics.
Train the *exact* same system on a dataset from Senegal, and everyone will look African.
📢Please join me in congratulating Harold Li
@LiLiunian
for passing his Ph.D. defense today🎉 Harold has accomplished remarkable work and has been a pioneer in vision-language foundation models. 🧵 1/
PM me to grab☕️ at
#NeurIPS2023
. I am seeking PhD/postdocs to join
@uclanlp
in Fall24 in areas:
- Trustworthy NLP👼: Align LLM with human values and ensure they safe uses
- Multimodal🙈: Align vision and language
- LLM Reasoning🤔: LLMs in commonsense, logical, math reasoning
I am honored to be nominated by SIGDAT (the org that oversees EMNLP) to run for VP-elect with other awesome candidates who share the goal of improving our community. Please check your email to vote by 3/24.🗳️ See details:
ACL
#SIGDAT
members, look in your inbox for VP-elect and Secretary/Treasurer elections. The candidates are really awesome, so you may have a hard time picking just one each.
@emnlpmeeting
@IAugenstein
@MonaDiab77
I’m at
#ACL2023
this week.
@uclanlp
members 🐻 and my collaborators at
@AmazonScience
will present the following papers at the conferences on the topics around trustworthy NLP, vision-language, and language+reasoning. See details at 🧵👇
Dear friends,
@hhexiy
@robinomial
@sameer_
and I will present a tutorial on "Robustness and Adversarial Examples in NLP" virtually at
#EMNLP2021
9am AST tomorrow (Nov 11). See you there!
#NLProc
Presentation link:
Slides/Website:
Indeed, it's impressive to see Gemini team act fast on new benchmarks. Our MathVista, released a month ago also been tested there. It's happy to see how the spirit of open science and platforms like
@huggingface
have simplified model reproduction/testing.
P.S. MathVista was led
Hi
@emilymbender
, I'm one of the lead authors of MMMU. I can certify that 1) Google didn't fund this work, and 2) Google didn't have early access. They really like the benchmark after our release and worked very hard to get the results. It doesn't take that long to eval on a
Congrats to
@LiLiunian
for winning Google PhD Fellowship! 🎉🥳🎊 Harold led pioneering efforts in vision-language research, including developing notable models such as VisualBERT, CLIP, and recently introduced Desco. He will be on the market this year!
@uclanlp
@UCLAengineering
In 2009, Google created the PhD Fellowship Program to recognize and support outstanding graduate students pursuing exceptional research in computer science and related fields. Today, we congratulate the recipients of the 2023 Google PhD Fellowship!
In the coming days, I'll tweet in English and Chinese for papers in Ethics & NLP and Theme areas
@aclmeeting
. Kudos to
@emilymbender
for the initiative of multi-lingual live-microblogging 我將會在tweet上即時轉撥ACL會議
One of my favorite papers!🚀 Instruction fine-tuning datasets are typically generated by mimicking human-annotated seed sets. But, why don't we construct instruction fine-tuning data from real datasets in repositories like in 🤗. Check out our
#EMNLP23
paper!
🦖 Dynosaur was accepted at EMNLP 2023!
#EMNLP2023
In the new version, we further show Dynosaur’s great perf. on Longform!
We’ll present Dynosaur in Poster Session 2 14:00-15:30 SGT
Preprint:
Code:
Congrats 🎉 to the newly titled Dr. Lu
@lupantech
on defending his thesis about mathematical reasoning with language models"! 🧮 Pan has published a series of works on quantifying and improving math and scientific reasoning ability in LLMs. Some highlights:
Hello World! 🙌🏼
#SoCalNLP2023
is back this year, to be held
@UCLA
! 🏝️☀️
Keep an eye here for more announcements, but for now, our CfP is open! More details ⬇️
🔗
📆Nov. 17, 2023
@zacharylipton
This is the answer from GPT-2😅:
We will have real, mass-deployed, fully autonomous self-driving cars in the next O(n) years, where n = 10 000. The theoretical limit of this scenario is not n = 10 000, but it is still, a lot of n's, to do n = 10 000 …
DM me if you're interested in a postdoc opportunity at UCLA NLP to work on exciting topics including fairness/robustness in NLP, machine common sense, cross-lingual transfer, language generation, and more!
@VioletNPeng
Congrats Dr. Ahmad (
@ahmadwasi
, ) 🥂🎈🎉 for passing his thesis defense! Wasi has done many impactful works in topics ranging from cross-lingual transfer, information retrieval, to NLP for programming languages.
#proudadvisor
#NLProc
Happy to share our work on Object Detection in the Wild through Grounded Language Image Pre-training (GLIP) (Oral at
#CVPR2022
)!
Superior zero-shot and few-shot transfer learning performance on 13 object detection downstream tasks!
Absolutely second this. A proposal I spent several months on was declined on the same day I received the Sloan award letter. Ping me if you want to hear more failure stories. Sometimes lucky, sometimes not, just believe what you love to do. I'll cheer for you.
Congratulations to the 51 award recipients of the 2019 Amazon Research Awards, who represent 39 universities in 10 countries. View the full list and find out how to be added to the 2020 Call For Proposal distribution list here:
#AmazonResearchAwards
Congratulate Jieyu Zhao
@jieyuzhao11
on winning the prestigious Microsoft Research Ph.D. Fellowship! Jieyu's research on examining and mitigating social prejudices in NLP models has made a great broader impact. The award is well deserved.
Congratulations to the 2020 Ada Lovelace and PhD Fellowship recipients! Their work, ranging from applications in mental health & neuroscience to quantum computing & model bias, leads a new vanguard in technology’s evolution. Learn more about 2020 fellows:
The subtitles of ACL recordings generated by ASR are hilarious..🤣 obvious, the language model is off-the-path and should be pre-trained on ACL anthology.. or the ASR should leverage the presenting paper as input when decoding...
It's application season! We might consider getting help from ChatGPT to draft SOP, reference letters, etc😏. But, hold on!✋ Yixin found that LLMs often perpetuate traditional stereotypes, portraying men as agentic (like natural leaders) and women as communal (such as well-liked
Might be slightly last minute, but I’m thrilled to announce that I will also be giving a talk on our paper in the
@gem_workshop
this afternoon! 🤩 The session will start at 2pm, in room Leo 2, but it will also be co-hosted through zoom. Drop by and say hi! 😁
Check out our new
#EMNLP2021
paper on non-binary gender in NLP. I learned a lot from this diverse team, which brings in many new perspectives. Thank for the team and survey respondents, who make this happens.
Congrats Tao Meng (
@TaoMeng10
) for passing his defense! 🎉 Tao's research centers on integrating constraints into LLM inference and learning, a challenging yet crucial aspect for effectively controlling large LLMs.
I almost feel shy to share personal news during this pandemic time: I'll join UCLA CS as an assist. prof. with my PlusLab this summer. I'm excited about the upcoming collaborations with new colleagues and continuing to contribute to SoCal-NLP.
@CS_UCLA
#SoCalNLP
1/2
Can we generate syntactically controlled paraphrase without parallel data? 🧐 Check out our (w/
@kuanhao_
) recent paper at
#EACL2021
. We also apply this to improve the robustness of models against syntactically adversarial attacks.🛡️
Check out our
#EACL2021
paper on controlled paraphrase generation! We show that it's possible to control the syntax of paraphrases and generate syntactically various paraphrases without using annotated paraphrase pairs. (, w/
@kaiwei_chang
) [1/2]
#UCLANLP
Thank you to everyone who submitted ICML/ACL-ARR reviews on time or is still working on them during the spring break 🏖️& COLM 🦙 deadline. If you still owe reviews, please kindly respond to your AC/SAC so that they know whether they need emergency reviewers 🙏
I am honored to be nominated by SIGDAT (the org that oversees EMNLP) to run for VP-elect with other awesome candidates who share the goal of improving our community. Please check your email to vote by 3/24.🗳️ See details:
Outstanding Papers at
#ACL2023NLP
. Congrats to all authors!!
📚 Backpack Language Models
🔎 CAME: Confidence-guided Adaptive Memory Efficient Optimization
🌍 Causes and Cures for Interference in Multilingual Translation [1/n]
How to transfer a relation/event extraction model across different languages by leveraging parse structures? 🧐 Come to our Graph Attention Transformer Encoder (GATE) paper () at
#AAAI2021
@ahmadwasi
@VioletNPeng
#UCLANLP
(1/3)
Do you have a paper on privacy, robustness, fairness, faithfulness, ethics, or safety in NLP? The 3rd edition of the Trustworthy NLP workshop will be co-located with
#ACL2023
. The submission deadline is 4/24. Please find CFP here:
📢 Attention, trustworthy NLP researchers! 📢
If you have papers on fairness, robustness, ethics, transparency, safety, misinformation, hallucination, etc, in NLP, we invite you to submit your papers to NAACL
@trustnlp
workshop by 3/27!
@guyvdb
Even my acceptance recommendation at the AC level gets overturned 😒. I felt bad for the paper. I spent several hours carefully reading the reviews, discussions, and some of broaderline rating papers like this one. I believe SACs did the same.
📢New paper📢 We propose 𝑷𝒂𝒓𝒂𝑨𝑴𝑹, a large-scale syntactically diverse paraphrase dataset constructed by AMR back-translation, and show its 🔥potential🔥 on several NLP tasks!
Work w/ Varun,
@IHung_Hsu
, Anoop,
@kaiwei_chang
,
@aram_galstyan
[1/n]
I am thrilled to share that I will join the Department of Computer Science and Engineering at
@TAMU
as an Assistant Professor in Fall 2024. Many thanks to my advisors, colleagues, and friends for their support and help. I'm really excited about the new journey at College Station!
Abstract submission deadline is Jan 25 so hurry up!
- the abstract is mandatory
- withdraw allowed by Feb 1
- in case of multiple submissions of the same paper, keep only one
- check your inbox for the automatic confirmation email after submission
#ACL2021NLP
#NLProc
How grammatical errors affect language encoders?🧐Our
#acl2020nlp
: "On the Robustness of Language Encoders against Grammatical Errors"( w/Fan Yin, Quanyu Long, Tao Meng) diagnoses encoders by simulating grammatical errors with adversarial attacks.
#NLProc
How bias are today's open-ended language generation models? 🧐Check out our
#facct2021
paper and a new benchmark on systematically measuring fairness/bias in NLG
@AmazonScience
#NLProc
I'll host a BoF Session on Ethics, Bias, and Fairness in NLP, Wed (Jun 9) 15:00 Seattle / 18:00 New York / 23:00 London / Thu(Jun 10) 3:30 Delhi / 6:00 Beijing / 8:00 Sydney. If you have topics you would like to discuss, DM me or post here
When performing transfer learning with multiple sources, one key question is how much info one can leverage from each source. In
#NAACL2021
paper, Rizwan Parvez
@uclanlp
developed SEAL-SHAP, an efficient source valuation framework for quantifying the usefulness of the sources 1/n
Designing an inclusive LLM 👩👩👧👧 requires attention to every detail. The new paper by
@ovalle_elia
highlights the impact of technical design decisions like tokenization on the model's inclusivity. Congrats on getting it to NAACL-Finding!
@AmazonScience
🚨LLM problems - esp. those that result in harmful biases - can indeed arise from tokenization! Turns out, BPE tokenization hinders LLM abilities to correctly use gender-neutral pronouns, thereby exacerbating LLM misgendering. Thread below ~
📝
🧵 [1/n]
This tweet is about an
#acl2020nlp
paper "Give Me Convenience and Give Her Death: Who Should Decide What Uses of NLP are Appropriate, and on What Basis?" by Kobi Leins, Jey Han Lau, Timothy Baldwin. 1/n
I have to skip EMNLP (but I'll be at Neurips 23). Please talk to our students from UCLA about our work in fairness in NLG, limitations of VIsion-Language models, effective instruction tuning, and Embodied Instruction Following
Ben’s group
@JHU
has a paper with similar title. Similar idea but with a different focus. It’s nice that they put the posters side by side, showing the gender bias in coref is verified independently by two groups.
Thrilled to be a part of the ECOLE program and work with this incredible team. We will develop human-machine collaborative approaches to analyze image, video, and multimedia data through vision and language in the era of generative AI🚀.
@uclanlp
We are very excited to kick off our DARPA ECOLE project, and extremely proud of this super young and diverse group of PIs for our MIRACLE team that I have put up together with
@cvondrick
(1/5) UCLANLP is associated with the following presentations at
#emnlp2019
/
#CoNLL
.
Interested in fairness in NLP? Check out 1) our tutorial (Mon); 2) talk on bias in NLG ; 3) poster on bias in language with grammatical gender
MathVerse
Does Your Multi-modal LLM Truly See the Diagrams in Visual Math Problems?
The remarkable progress of Multi-modal Large Language Models (MLLMs) has garnered unparalleled attention, due to their superior performance in visual contexts. However, their capabilities in
I gave a tutorial on Transferable Representation Learning in NLP at
#ODSCWest
. The talk summarizes our recent research on cross-lingual transfer, Vision&Langauge pre-training (VisualBERT), Programming&Natural Languages pre-training (PLBART) See slides:
Hire him! Kuan-Hao completed PhD with me and now is a postdoc at UIUC. He had a series of significant works in NLP, particularly in (1) generative event extraction and (2) enhancing robustness against syntactic perturbations. I'm always amazed by his ideas🤯 He is one of a kind.
Heading to Singapore for
#EMNLP2023
soon🚀🚀 Let's discuss about language understanding/information extraction/ knowledge generalization!
I am on the faculty job market this year. Feel free to schedule a chat with me!
First, on 7/9, I’ll share insights on leveraging indirect supervision signals for learning vision-language representation at the 'Indirectly Supervised NLP' tutorial.
Interested in privacy, fairness, explainability, accountability, security, ethics in NLP systems? 🤖Consider submitting papers or participating in TrustNLP workshop at NAACL 2021. Details:
Excited to announce our first workshop on Trustworthy Natural Language Processing, held as part of
@NAACL
2021 in Mexico City this June. Call for papers and details on submission can be found at . See you there!
How to bridge typological differences across languages in cross-lingual transfer?🧐 Check out
@ahmadwasi
's
#ACL2021
paper ()! We showed that explicitly encode dep parse tree with GAT into M-BERT improves cross-lingual transfer. Happy to talk about it at ACL
hot off the press -- VisualBert: A simple and performant baseline for vision and language. Language + image region proposals -> stack of Transformers + pretrain on captions = SOTA or near on 4 V&L problems.
@LiLiunian
+Cho-Jui Hsieh +Da Yin
@kaiwei_chang
Thanks,
@emilymbender
's awesome live tweeting. The full talk “Gender Bias in Contextualized Word Embeddings (NAACL 2019)” is on YouTube now. Great job
@jieyuzhao11
!
UCLA CS is hiring this year! 🥳 We have multiple open rank tenure-track faculty positions. We especially encourage candidates who commit to mentor students from underserved populations to apply.
Check out Xiao's paper on argument sufficiency assessment👩🏫. She proposed a framework to determine whether the conclusion is well-supported by premises using probability of sufficiency. Congrats on getting it to NAACL!
Kareem, co-advised by
@guyvdb
and me is on the faculty job market🚀 Grab ☕️ w/ him at
#NeurIPS2023
if your dep. is hiring. He has done amazing work on neuro-symbolic AI 🤖 This specific paper enables us to inject logical constraints in autoregressive models like language models.
Neuro-Symbolic (NeSy) methods inject constraints into NNs, but do not support autoregressive models e.g. transformers. We propose Pseudo-Semantic loss a NeSy framework for injecting arbitrary logical constraints into autoregressive models. At
#NeurIPS2023
!
Check out the new analysis of Visual Math Reasoning with Bard, GPT-4V, Gemini by
@lupantech
.
Interestingly, Gemini and GPT-4V have slightly different skill sets (see the radar graph).
Thanks
@lupantech
for the quick analysis during the confernece,
@JeffDean
to connect us to
💥💥Update Alert! Radar graphs & leaderboard on
#MathVista
now feature detailed scores for the
#Gemini
family models. 🚀
🔍 Insight: Gemini Ultra leads the pack, outperforming GPT-4V by 3.1%! Yet, each model shines uniquely in various math reasoning & visual contexts.
🙏 Big
Congrats to
@UCLA
Asst. Prof.
@adityagrover_
and incoming Asst. Prof. Saadia Gabriel
@GabrielSaadia
of
@CS_UCLA
on being named to
@Forbes
' 30 Under 30 list in science. Grover and Gabriel were each recognized for their work using artificial intelligence.
Appendix A in our NIPS paper explicitly discusses the same issue of 3COSADD as in the "Sensational" paper. This is the reason why we need to design a different experiment.
"Sensational" paper criticizes our paper on bias in word embeddings for using the parallelogram analogy alg. (min_x |x-y+a-b|), but we don't use that alg. at all (see img) and we UPPER-bound |x-y| not lower-bound it. Don't trust everything on arxiv. :-)
Check out our
#ACL2021
paper() on NLP for privacy policy docs. We collaborate w/ security researchers and law school to annotate PolicyIE, a corpus consisting of 5,250 intent & 11,788 slot annotations over 31 policies of websites and mobile applications. 1/
Our
#NAACL2021
paper demonstrates the challenges when applying NLP in analyzing narratives from USA’s CDC National Violent Death Reporting System. We showed that Coref suffers from poor transferability due to domain gaps, especially in narratives involved LGBT individuals 1/n
Can a model learn problem structure/constraints from data? E.g., given pairs of adjacent matrix and corresponding minimal spanning tree, can a model learn to solve MST? Check out our
#ICML2021
paper on constraint mining with ILP w/
@kaiwei_chang
1/4
At
@FAccTConference
, Amazon researchers presented a new dataset and methodology for evaluating bias in language models like BERT and GPT-2. Initial experiments did find evidence of bias in texts generated by popular language models.
#facct2021
Please consider submitting your work to the Responsible AI
#KDD2021
workshop organized by
@sunipa17
, Mehrnoosh Sameki,
@jwaladhamala
, and Cho-Jui Hsieh
Call for papers for Workshop on Measures and Best Practices for Responsible AI
@kdd_news
Paper Submission Deadline: May 10th, 2021
Additional details in the link below.
#KDD2021
@yoavgo
I made a cartoon for a NeuRIPS paper in 2016 as a 3-min highlight. After trying several times and realizing that I cannot speak naturally, I just use a robotic narrative and it's actually not bad🤣...Check it out...
DesCo🪩: Can we teach a vision-language model to recognize objects by language descriptions as a kid does? 🔥 Our new approach leverages LLM for training vision models based on rich descriptions
@uclanlp
Excited to share our new work DesCo () -- an instructing object detector that takes complex language descriptions (e.g., attributes & relations).
DesCo improves zero-shot detection (+9.1 APr on LVIS) and ranks 1st at the
#OmniLabel
Challenge of CVPR2023!
As a reminder, we're tweeting about the KG-BIAS workshop (co-located with
@akbc_conf
), specifically focusing on bias in the context of knowledge graphs. We're keeping live notes here: , and you can find more info at
#KGBIAS
We've added a new chapter to the guide, thanks to
@arjunsubgraph
, all about fairness metrics, bias metrics, and bias mitigation tools for your models. Learn how you can apply bias mitigation to large, contextual language models directly from a config file!