A dataset of naturalistic in-context implicatures from Internet corpora
More details in the blog post here.
The dataset covers 3 types of implicatures:
Scalar implicatures with some.
Manner implicatures with double negation
Manner implicatures through wordy paraphrases
Examples are designed in a way that it is possible to measure the effect of the implicature on a language model without any prompting, just by measuring changes in token probability distributions across interventions. More details in the paper!