🔥Telegram İfşa

New #NLProc paper on ArXiv! More and more papers in NLP claim to evaluate on ‘typologically diverse’ languages. But what does this even mean? In our new paper (with Wessel Poelman, Miryam de Lhoneux/ @mdlhx.bsky.social and Johannes Bjerva), we systematically such claims. arxiv.org/abs/2402.04222 1/🧵

It has come to our attention that many researchers and developers are confused by the existence of both zh-hk (Hong Kong Chinese) and yue (Cantonese) on Common Voice, and don't know which one to use. Our short answer is: use yue, and do NOT use zh-hk. Reasons below.

Very exciting news! I'm hiring a postdoc (Assistant/Associate Project Scientist) to work on AI and whales Project CETI UC Berkeley 🐳 We're a very interdisciplinary lab, as shown by this requirement: - PhD in linguistics, cognitive science, psychology, CS, marine biology, or

Under what conditions do we use placeholders like Engl. "whatchamacallit", Chin. "na ge shenme", Span. "cosa/cacharro/chisme", Germ. "Dingens" etc.? T. Seraku proposes an implicational hierarchy based on data from 56 language. degruyter.com/document/doi/1…

Crazy for pasta, crazy for cultural evolution of pasta ripiena from Italy and Eurasia 🥟🧡 2 phylogenetic branches setting the origin in Northern Italy, while Sardinian Culurgiones are in a separate clade. Which one is your favorite? link.springer.com/article/10.100…

#ThrowbackThursday #NLProc "SocioProbe: What, When, and Where Language Models Learn about Sociodemographics" by Anne Lauscher et. al investigates PLMs' knowledge of sociodemographics via probing and finds it a major challenge in NLP. aclanthology.org/2022.emnlp-mai…

When we hear someone speak a dialect, we can often tell where they're from. Can LMs do the same? Our #TACL paper addresses this question and shows how to boost LMs' geolinguistic skills. 🌍 This paper has been in the making for almost three years, so glad it's finally out! 🧵

📢New paper in PNAS! How rich was Vienna at the time of Mozart or Tuscany at the time of Michelangelo? Historical GDPs per capita are scarce, leaving this unanswered. Here, we provide new estimates using machine learning to augment historical GDPs per capita. /1 🧵

not claiming this as an original insight, but one reason to not think in terms of 'gaps in the literature' is because very, very often good academic work is about reconceptualizing space that has already been 'filled'. if you're looking for 'gaps', you'll never see that

Languages tend to rely on consonant lengthening to mark the onset of words nature.com/articles/s4156…

Traditionally in ML, building models is the central activity and evaluation is a bit of an afterthought. But the story of ML over the last decade is that models are more general-purpose and more capable. General purpose means you build once but have to evaluate everywhere.

Unironically wanting to do computational cow sociolinguistics now

🤖 NEW PAPER 🤖 Chain-of-thought reasoning (CoT) can dramatically improve LLM performance Q: But what *type* of reasoning do LLMs use when performing CoT? Is it genuine reasoning, or is it driven by shallow heuristics like memorization? A: Both! 🔗 arxiv.org/abs/2407.01687 1/n

Somewhere along the way “correlation is not causation” morphed into “until the causal link is proven beyond a reasonable doubt, ideally with a large scale RCT with global external validity, than the correlation likely reflects the impact of some unobserved factor I can’t name.”

If you make a drawing in the weight matrices of your neural network at initialization, it will likely still be visible at the end of training arxiv.org/abs/2012.02550

🚀 Excited to announce that DeSTA2, a spoken LLM. What's amazing? It's trained on just 150 hours of speech data, beating models with tens of thousands of hours! Shoutout to NTU PhD student Ke-Han Lu & NVIDIA team for leading this! 👉 Learn more: kehanlu.github.io/DeSTA2/

It’s crazy how talking with a science buddy is simultaneously what’s most likely to move the project forward and the thing that’s least prioritized in our schedule.

Inequality isn't just about income and wealth. It's also related to transport and accessibility. Here's how poorer people are more impacted by inaccessibility to public services:

Look at all those dialects in the data ✨

What's actually different between CLIP and DINOv2? CLIP knows what "Brazil" looks like: Rio's skyline, sidewalk patterns, and soccer jerseys. We mapped 24,576 visual features in vision models using sparse autoencoders, revealing surprising differences in what they understand.