My co-lead @KaufCarina and I present: an in-depth investigation of event plausibility judgments in language models.
A 🧵 1/
arxiv.org/abs/2212.01488
A 🧵 1/
arxiv.org/abs/2212.01488
Knowledge of event schemas is a vital component of world knowledge. How much of it can be acquired from text corpora via the word-in-context prediction objective?
2/
2/
We test this Q using a controlled minimal pairs paradigm, with simple event descriptions. We manipulate plausibility by swapping the agent and the patient (The teacher bought the laptop / The laptop bought the teacher) or changing the patient (The actor won the award/battle). 3/
LLMs are almost perfect when assigning likelihood to possible vs. impossible events (The teacher bought the laptop / The laptop bought the teacher) but aren’t as good when it comes to likely vs. unlikely events (The nanny tutored the boy / The boy tutored the nanny).
4/
4/
In follow-up tests, we show that
- LLM scores depend both on plausibility and surface-level factors like word frequency (meaning that distributions for plausible and implausible sentences are highly overlapping)
5/
- LLM scores depend both on plausibility and surface-level factors like word frequency (meaning that distributions for plausible and implausible sentences are highly overlapping)
5/
- LLMs generalize very well between active and passive versions of the same sentence BUT not as well as humans for synonymous sentences (The teacher bought the laptop / The instructor purchased the computer).
6/
6/
- explicit plausibility information emerges in mid LLM layers and then stays high
- as in beh tests, animate-inanimate (AI; impossible) events are easier than animate-animate non-reversible (AAN; unlikely) events
7/
- as in beh tests, animate-inanimate (AI; impossible) events are easier than animate-animate non-reversible (AAN; unlikely) events
7/
Minor but fun:
- a probe trained on both active and passive voice sentences is as successful as a within-voice probe (but a probe trained on only one voice type fails to generalize, especially in early layers)
8/
- a probe trained on both active and passive voice sentences is as successful as a within-voice probe (but a probe trained on only one voice type fails to generalize, especially in early layers)
8/
Check out the paper for an interpretation of these results, including a discussion of selectional restrictions, reporter bias, and more!
9/
9/
This is an international collaboration brought together by @ev_fedorenko and @AlexLenci1966, with vital contributions from @g_rambelli and @EmmanueleChers1 (and our undergrads Selena She & Zawad Chowdhury).
10/
10/
It's been a crazy run. We started in early 2020, had zoom (or skype?) calls during the 2020 covid lockdowns & have coordinated many meetings between Boston, Italy, Hong Kong, and sometimes Germany and Russia. Glad the project has finally come to fruition!
11/end
11/end
Bonus: a preliminary exploration of ChatGPT responses shows that it might also have an impossible-implausible gap (although a more detailed investigation is of course needed).
12/11
12/11
• • •
Missing some Tweet in this thread? You can try to
force a refresh
