Language Models Require Supporting Context to Predict Appropriately in Negative Contexts
James Michaelov ⋅ Benjamin Bergen
Abstract
Both humans and language models have been reported to show complex patterns surrounding the processing of words in negative contexts - that is, following a negation such as 'not' or a quantifier such as 'few'. In this study, we show that contemporary language models struggle to predict atypical but semantically plausible sentence continuations in negative contexts unless there is a substantial contextual support for the atypical continuation.
Chat is not available.
Successful Page Load