wonderful paper exploring the impact of anthropomorphic language (esp. in light of our AnthroScore work, which measures prevalence of this language!)

our new paper on Use vs. Mention examines the consequences of LMs misclassifying the mere **mention** of contentious topics as harmful. very relevant to discussions of online harms, moderation, censorship, #WOAH2024 etc. ↓↓↓

TIL about the #catsofeacl hashtag.. have been running into this guy everyday here :D

Presenting AnthroScore at the 4pm poster session today!! and at the Computational Social Science session at 11am tomorrow!!! come say hi :D
#EACL2024

Attending #EACL2024 and passionate about computational social science (CSS) or social applications of NLP?
Join us tomorrow (Wednesday at 2 pm in the Carlson room) for an informal CSS Birds of a Feather gathering!
Let's meet, chat, and share insights!

In Malta 🇲🇹 for EACL this week!! Let me know if you’re around and wanna talk about nlp applications esp education, or just hang with me Myra Cheng Kristina Gligorić … or run with me along the coast!! 😎🏃🏻‍♀️

view from the conference hotel at #EACL2024 👀 let me know if you're here and want to talk about anthropomorphism, implicit language, social biases in LLMs, or just hang (cc Kristina Gligorić Rose !!)

There are many concerns around LLMs being politically biased. But how, if at all, can we meaningfully evaluate values/opinions in LLMs?

In our new paper, we show that current *constrained* evals (e.g. surveys) likely tell us little about LLM values/opinions in the real world

🧵

“It’s a double-bind that the field is caught in, where the users and creators of LMs have to anthropomorphize [when prompting, giving instructions, etc.], but that leads to more and more misleading ideas about what these models can do.”

New article about AnthroScore!

Do LLMs learn impossible languages (that humans wouldn’t be able to acquire) just as well as they learn possible human languages?

We find evidence that they don’t! Check out our new paper…

💥 Mission: Impossible Language Models 💥
ArXiv: arxiv.org/abs/2401.06416

🧵

AboutMe: Using Self-Descriptions in Webpages to Document the Effects of English Pretraining Data Filters

Shows that some quality classifiers act like topical domain filters, and langID can overlook English content from some regions of the world

repo: github.com/lucy3/whos_fil……

This paper is really important, presenting empirical evidence of the imbrication bet. AI & the surveillance biz model. This is notable particularly given that most production surveillance tech is proprietary, its existence and use hidden from the public.

arxiv.org/abs/2309.15084…

By 'surveillance biz model' I'm referring both to the data-centric platform biz model that emerged c. 1990s, & persists as the tech industry's economic engine, & the biz of provisioning surveillance systems in particular, which generally (as here) build on top of this paradigm.