Negated and Misprimed Probes for Pretrained Language Models: Birds Can Talk, But Cannot Fly

TL;DR

This paper introduces two probing tasks, negation and mispriming, to evaluate how well pretrained language models understand factual knowledge, revealing their limitations in handling negation and susceptibility to distractions.

Contribution

The paper proposes novel probing tasks for assessing factual knowledge in PLMs, highlighting their inability to handle negation and their vulnerability to mispriming effects.

Findings

PLMs do not distinguish negated from non-negated questions

PLMs are easily distracted by misprimes

PLMs show limited understanding of human-like factual reasoning

Abstract

Building on Petroni et al. (2019), we propose two new probing tasks analyzing factual knowledge stored in Pretrained Language Models (PLMs). (1) Negation. We find that PLMs do not distinguish between negated ("Birds cannot [MASK]") and non-negated ("Birds can [MASK]") cloze questions. (2) Mispriming. Inspired by priming methods in human psychology, we add "misprimes" to cloze questions ("Talk? Birds can [MASK]"). We find that PLMs are easily distracted by misprimes. These results suggest that PLMs still have a long way to go to adequately learn human-like factual knowledge.