@LessCrime@BrendanNyhan@ChenhaoTan Sure in many areas the signal is weak, I’m not making a sweeping claim, but it’s strong in many contexts and we’ve relied on it in those
@BrendanNyhan@ChenhaoTan And in the interim, so long as our systems are still premised on receiving that signal, it makes sense to try measuring how much time/work authors put into the paper.
@BrendanNyhan@ChenhaoTan It’s not just about whether claims are true, but whether they are important and interesting. An author committing time to write offers some signal that the idea is worth reading about. We will have to find ways of recovering that signal as the cost to producing papers goes down.
@ifaposto Cool idea, never heard of that. So your family of q-distributions is then conditional on p. I feel like this might break the usual convergence guarantees, but I’m not sure, and maybe that doesnt matter. Would be cool if it works…
@ifaposto Ah I see, that’s super interesting... so you do not update these two coordinate-wise, but you use the coordinate-wise update to define the q-distribution?
@ifaposto@charlesm993 Ah I remember an example. @keyonV's paper on "text-based ideal points". They do the optimal CAVI updates for some latents but then do fixed-form variational updates for others with reparameterization gradients: https://t.co/tLj7cbbP7h. This is full variational Bayes tho, not EM.
Delighted to announce the next step in my career!
After my postdoc, I will begin a joint appointment at TU Wien and the Complexity Science Hub Vienna as an Assistant Professor in NLP! My deepest thanks to all who helped me along the way
I'm hiring—details on PhD positions below
@yoavgo I’m not so sure! I could imagine humans also being able to recognize the false claim when given the text, but months later recalling the claim but forgetting the context
@yoavgo Sure but does this paper provide further evidence of that? I think it’s a genuinely interesting question whether humans, presented with this same experiment, would later recall false facts
@yoavgo I am genuinely asking whether it is known that humans don’t actually read and learn in this way (eg I can imagine giving human subjects the same document and seeing 3 months later if they recall false facts)
Attention @arxiv authors: Our Code of Conduct states that by signing your name as an author of a paper, each author takes full responsibility for all its contents, irrespective of how the contents were generated. 1/
@alemelk My sense is we should be more worried about our fields becoming full of slop than about our fields becoming communities of people who carefully read their own papers