Inference towards the antonym -- surprisals

Surprisal measures for minimal pairs testing whether the negation of an adjective is likely to imply its antonym.

3 different "paradigms":

Inference forced by the additive particle too, inserted to the left of the antonym ("ante");

Inference forced by the additive particle too, inserted to the right of the antonym ("post");

Inference forced at the "meta" level, via the predicate "means" ("meta").

2 measures of surprisals:

Mean;

Total.

For BERT, two methods to compute MLM surprisals:

Salazar et al., 2020 (original measure)

Kauf and Ivanova, 2023 (labeled "withinwordl2r")

All computed with minicons.