[Paper Review] Negated LAMA: Birds cannot fly

Nora Kassner, Hinrich Schütze|arXiv (Cornell University)|Nov 8, 2019

Topic ModelingComputer Science17 citations

TL;DR

This paper introduces two probing tasks—negation and mispriming—to evaluate factual knowledge retention in Pretrained Language Models (PLMs). It shows that PLMs fail to distinguish between negated and non-negated cloze questions and are easily misled by irrelevant misprimes, indicating significant limitations in learning human-like factual knowledge.

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.

Motivation & Objective

• To investigate whether Pretrained Language Models (PLMs) can accurately represent and distinguish between negated and non-negated factual statements.
• To examine whether PLMs are susceptible to distraction by irrelevant or misleading contextual cues in cloze-style questions.
• To evaluate the extent to which PLMs internalize factual knowledge in a way that mirrors human-like reasoning and robustness.
• To develop probing tasks that expose fundamental weaknesses in PLMs' factual knowledge representation.

Proposed method

• Design a negation probing task where cloze questions are presented in both negated (e.g., 'Birds cannot [MASK]') and non-negated (e.g., 'Birds can [MASK]') forms.
• Use the same factual knowledge (e.g., 'fly' as the correct answer) in both forms to isolate the model's ability to process negation.
• Introduce misprimes—irrelevant or misleading words (e.g., 'Talk?')—before cloze questions to test model robustness.
• Measure model predictions on masked tokens to assess whether negation or misprimes alter the model’s output in a way that reflects understanding.
• Compare prediction accuracy and confidence between standard, negated, and misprimed cloze questions.
• Base the probing framework on the LAMA benchmark (Petroni et al., 2019), extending it with negation and mispriming variants.

Experimental results

Research questions

• RQ1 Can PLMs correctly identify the correct answer in cloze questions when the sentence is negated?
• RQ2 Does the presence of a misprime significantly alter the model’s prediction in a cloze question?
• RQ3 How does the model’s performance on negated and misprimed cloze questions compare to its performance on standard factual cloze questions?
• RQ4 To what extent do PLMs reflect human-like understanding of factual knowledge under syntactic and contextual perturbations?

Key findings

• PLMs do not distinguish between negated and non-negated cloze questions, often assigning similar probabilities to the correct answer in both cases.
• The models show no significant drop in confidence or accuracy when the sentence is negated, indicating poor handling of negation.
• Misprimes such as 'Talk?' significantly influence model predictions, leading to incorrect or inconsistent outputs.
• The model’s predictions are easily distracted by semantically irrelevant context, suggesting weak reasoning and attention control.
• These results indicate that PLMs store factual knowledge in a fragile, context-sensitive manner rather than as robust, structured facts.
• The findings highlight a critical gap in PLMs' ability to learn and apply factual knowledge in ways that resemble human cognition.