Venelin Kovatchev

The 15th edition of the Lisbon Machine Learnings School (LxMLS 2025) is looking for its monitor team. As always alumni are especially welcome. Apply before the month ends! bgmartins.github.io/lxmls-website-…

The #LxMLS2024 monitors team!

Funded PhD opportunity at the University of Birmingham. I am recruiting a student to work on data-centric NLP. Reach out if you are interested in e.g.: active learning, curriculum learning, adversarial data collection, and evaluation for NLP and LLMs. findaphd.com/phds/project/d…

How much does data impact the evaluation of NLP models? How can we measure data distribution in an efficient and multi-dimensional way? How to predict OOD generalizability? Check our paper with @mattlease - "Benchmark Transparency", accepted at NAACL (arxiv.org/abs/2404.00748)

My heart is full ❤️ So grateful for the community support. I’ll do my best!!

@jessyjli Congratulations! Well deserved!

The School of Computer science at UoB is inviting national and international candidates to apply for funded PhDs. The candidates can choose the area, topic, and supervisor. More info: tiny.cc/z06dvz

@RamonAstudill12 @yoavgo Update, received multiple emails in my spam folder ~2 hours ago.

@sintelion @yoavgo ☝️ I am not the only one

Does everyone get the ACL anonimity survey email?

@elneurozorro When thousands of arxiv papers are getting out, most of them get little attention. Large labs have many (and popular) members that share links to papers (regardless of quality), effectively filling the "reading bandwidth" of neutral readers.

@sintelion But honestly that's an interesting take. Low bar to publication = more noise, which we already have too much of. But how it would favor large labs? The peer review process is infamously biased, and large labs have more resources to toil through it.

Preprints should be: 1) accepted papers or 2) papers that will never be peer reviewed. 1) and 2) should be clearly marked. If you are scared someone will steal your research - implement registration (e.g. OSF). If your research will be irrelevant in 3 months - rethink it.

@elneurozorro It is a related issue and imo all stems from lack of patience and proper scientific approach. Conceptually, pre-prints are a great idea. In practice, they just fuel publishing frenzy and favor large labs that can attract more attention in the crazy article spam.

@sintelion Well that's a different issue entirely which I agree with you on, which is that we should publish way less! And I could see how if preprints lower the bar to "publish" it could make that problem worse. Not sure what the solution to that is but it's more related to incentives

@elneurozorro I guess it depends on the scale. When it gets to hundreds of papers published every day, it becomes an unbearable spam of scientific clickbaits and "fast food" research that is obsolete in a month. It tolerates a culture of quantity over quality, which is unhealthy.

@sintelion What's wrong with pre-peer review preprints? It's not just to mark your work, but to get feedback, and speed up scientific communication.

Reading a rebuttal where the authors have relied heavily on LLMs to "polish their writing". 2 page rebuttal with 0 meaningful content. Should I lower the score as they wasted 30 minutes of my time or give them an extra point for making me laugh on a Friday?

"To overcome the naivety of Naïve Bayes, I suggest that we use a logistic regression!"

At Dynabench, we're gearing up for the AI race, and we embrace the rapid pace of change! We're excited to announce some big updates that make our new-and-improved platform faster, better, and easier to use, for leaderboard users, dataset creators, and benchmark owners! 1/7

Embeddings t-SNE vizualizations of paraphrase datasets show: There are big differences in semantic balance. Left: ETPC (human) by @sintelion Right: MPC (machine) by @jpwahle Research still lacks evenly distributed paraphrase datasets by machines! arxiv.org/pdf/2303.13989…

@jessyjli @UT_Linguistics Congratulations! Well deserved!

Tenure! I cannot express how grateful I am for the support I’ve got from family, colleagues, students, and friends & collaborators 🥰 I’m incredibly lucky and honored to belong in @UT_Linguistics, and the larger UT NLP group! liberalarts.utexas.edu/linguistics/ne…

New lit review: @d_anubrata, @sintelion, & Houjiang Liu call for human-centered NLP for fact-checking: "The state of human-centered NLP technology for fact-checking." Information Processing & Management, 60(2), 2023 utexas.box.com/shared/static/… @UTiSchool @EngagingNews @UTGoodSystems

@KhalilMrini The wording of the prompt matters! If I ask it for "wiki article on ...", I'm also a football player. If I ask it to "write an article on ..." I am suddenly a "little known Bulgarian poet".

#Galactica has revealed my true identity to the world. There is no more hiding from my sports career in 🇹🇳 Tunisia 🙈 ... (there is no Tunisian footballer who has my name 😂) #NLProc #NLP #Meta