Though I also wonder if all of this generalizes to the inverse case of meaning-seeking minimal pair stimuli

bsky.app/profile/kani...

I am 100% on board with this framing, but wondering if it’ll hold for minimal pair stimuli that are both grammatical but convey very different messages, an example from COMPS:

A robin can fly.
A penguin can fly.

(Similarly for datasets like EWoK)

Chris Potts on Twitter, circa 2023: language models are outstanding (lmao)

Congratulations!!!

Congrats!!!

Keeping it ambiguous as always 🥸😎

Of course!! Thanks so much for the kind words — it’s just a bunch of modifications I made over the old Hugo academic theme!

Woah 😳!

And here’s an alternate link: semanticsarchive.net/Archive/zg1Z... my bad!

Thanks for explaining! Sure, calibration of common ground can potentially change expectations about naturalness. Another way this can happen is from specific constructions (e.g., “hey, wait a min”), which has been shown to have an effect in humans w/o any other context (S&K). See our exp 2 for this!

Check out Syrett and Koev’s work (academic.oup.com/jos/article/32…) among others as evidence for human sensitivity to one being a more surprising response than the other

Also check out our second exp — we saw weakened sensitivity when there were digression triggers (like “hey wait a minute”)

Yes — we test this in exp 2 where we use digression triggers to see if the preference for the at issue content is weakened (it is; check out our exp 2 for details)