MPower Professor; Professor, Linguistics and Institute for Advanced Computer Studies, University of Maryland; Tech Advisor, FiscalNote, Coleridge Initiative
I complain when a company's behaved poorly, so only fair to praise when they do well. I had a short-notice emergency that led me to not take a @SouthwestAir flight, and they waived their no-show policy to refund. Much appreciated.
@ChrisGPotts Even leaving aside the linguistic and philosophical debate, though, there are important real-world reasons the radical version of distributionalism is a problem — see https://t.co/1tfes7cKKT.
Scientific progress requires investigating a diverse set of ideas. Do we still have that in natural language processing and AI? In this short (3min) video snippet, I raise this question during a panel at the top NLP conference this past summer. https://t.co/fQxcXnqKqH
@ChrisGPotts There’s evidence that he’s more than just hedging (see the rest of the paragraph I quoted) but I don’t doubt you’dhave been inclined to see how far you could take it! :) And quite possibly he’d have been up for that. I dearly wish Lila Gleitman were still around for us to ask.
@ChrisGPotts Harris (1954) speaks of distributional relations “which *correlate with* some aspect of meaning” (p. 156, my emphasis). He never says there is no notion of meaning other than distribution, and in fact he clearly indicates that meaning goes beyond just distributions.
Should we think of LLMs as cognitive models? This talk goes beyond the "argument from amazingness" to a more careful assessment of what it means to model human language processing, and why thinking of LLMs as cognitive models might or might not make sense.
https://t.co/Abra43koq7
To be presented at ACL 2025: Large Language Models Are Biased Because They Are Large Language Models.
Article: https://t.co/0TzlqUVmXs
Short (8min) video: https://t.co/E43jQ4Yw4A
#ACL2025NLP#NLProc#LLMs
Also: "results from the human evaluation indicate that a classical model (LDA, Blei et al., 2003) performs at least as well, if not better, than its modern equivalents." MALLET, with suitable preprocessing (stoplist, phrasal tokens) still has serious legs!
In earlier work, we showed that neural topic model evaluation was broken, and those models didn't improve over classical methods the way people thought. This new paper provides a replacement paradigm that's grounded in the real-world requirements of qualitative content analysis.
(Repost due to mistaken deletion😢):
Evaluating topic models (& doc clustering methods) is hard. In fact, since our paper critiquing standard eval practices 4 years ago, there hasn't been a good replacement metric
That ends today! Our ACL paper introduces a new evaluation🧵
Also joint work with @psresnik and @boydgraber. This work concludes a "trilogy" of topic model evaluation papers
paper 1: https://t.co/V5GWJ4NqOO
thread 1: https://t.co/xnmWyivl7a
paper 2: https://t.co/RVDItyBot3
thread 2: https://t.co/zaGeaYCkRN
@Itay_itzhak_ Thanks for this! Possibly of interest: my new paper on (normativity-related) bias aligns with yours, and offers a root-cause answer to your Section 3 question, "What Causes Bias in LLMs?". https://t.co/E43jQ4Yw4A, https://t.co/0TzlqUVmXs #ACL2025NLP#NLProc
Researchers @UofMaryland are examining how and why some misleading narratives proliferate via strategic use of certain mainstream news articles. The paper—coauthored by Pranav Goel, @_Jon_Green, @davidlazer & @psresnik—was published @NatureHumBehav.
https://t.co/D7QnVHavDH
Looking for guidance on raising successful kids in a complicated world? Rebecca Resnik's new podcast is (objectively, even if I'm personally biased!) a great place to go. https://t.co/2kSKaIqoTp
@UHC If an MD prescribed a medication, they’re saying it’s medically justified. If the generic is unavailable and you insist on extra pre-authorization to approve the brand name, there’s ONLY one reason: to avoid paying out benefits to which someone is entitled. Shame on you.
Just out on arXiv: my paper arguing that harmful biases are an inherent consequence of the underlying assumptions in any large language model, as LLMs are currently formulated. To the extent this is true, those assumptions badly need to be revisited. https://t.co/buCnm02vY8
I am beyond thrilled to report that my first PhD grad, Mona Diab, has been named a new Fellow of the Association for Computational Linguistics, along with four other terrific scholars. Congratulations to @MonaDiab77!
Honored my paper with @psresnik was accepted to Findings of #EMNLP2023! Many psycholinguistics studies use LLMs to estimate the probability of words in context. But LLMs process statistically derived subword tokens, while human processing does not. Does the disconnect matter? 🧵