The Negation Bias in Large Language Models: Investigating bias reflected in linguistic markers

COLM 2025

TL;DR

This paper investigates "negation bias," a human tendency to use negative phrasing for stereotype-challenging situations, in large language models (LLMs). They found that autoregressive LLMs exhibit this bias, showing higher "surprisal" (perplexity) for affirmative descriptions of non-stereotypical scenarios, while encoder-based models do not. This highlights the importance of analyzing subtle linguistic markers, not just explicit content, for understanding LLM biases.

What Problem Does This Paper Address?

While it's well-established that LLMs can perpetuate social biases and stereotypes, much of the research has focused on explicit associations (e.g., "doctor" and "man"). However, human communication involves subtler linguistic mechanisms that reinforce stereotypes, such as the "negation bias." This bias, observed in social psychology, describes how people tend to use negative constructions (e.g., "not tidy") rather than affirmative ones (e.g., "messy") when describing behaviors that violate common stereotypes. This paper aims to bridge the gap between psycholinguistic insights and NLP bias research by investigating whether LLMs also exhibit this negation bias, which could be a significant, yet overlooked, source of bias in their outputs.

Methods

The authors operationalized negation bias computationally by measuring LLM "surprisal" using perplexity. They created a dataset based on human experiments, pairing stereotypical and non-stereotypical contexts with both affirmative and negated descriptions. For each LLM, they fitted a linear mixed-effects model to analyze the interaction between context (stereotypical vs. non-stereotypical) and linguistic form (affirmative vs. negated) on perplexity. This allowed them to assess if models showed a "crossover pattern" where negation's effect on perplexity reversed depending on the context's stereotypicality. They evaluated both autoregressive decoder-only models and masked language models (MLMs).

Datasets

The core dataset used was derived from Beukeboom et al. (2020), which investigated negation bias in human communication. The authors expanded this dataset algorithmically, creating 300 root examples and a total of 1,200 sentences. They employed two expansion methods: synonym replacement and sentence structure variation (e.g., using causal links, relative pronouns, or reordering clauses). They specifically avoided LLM-based expansion to prevent introducing model-specific biases into the dataset itself.

Key Insights and Ideas

  • Negation as a subtle bias marker: The study demonstrates that negation isn't just a grammatical tool but can carry significant social bias, mirroring human communication patterns.
  • Architectural differences matter: Autoregressive models, which generate text sequentially, are more sensitive to negation bias than encoder-based MLMs, which process text bidirectionally. This suggests that the way models process language influences their susceptibility to certain biases.
  • Beyond explicit content: This work emphasizes that focusing solely on explicit content or word associations misses crucial dimensions of bias that are embedded in linguistic structure and pragmatic use.
  • Human-like bias in LLMs: The findings suggest that some LLMs can replicate nuanced human biases, indicating that they are not just reflecting biases in training data but may be actively learning and reproducing them.

Key Results

The study found that autoregressive decoder models (like GPT-2, Mistral-7B, and Llama-3.1-8B) exhibited a significant context × form interaction, indicating negation bias. Specifically, these models showed higher perplexity for affirmative descriptions in non-stereotypical contexts compared to negated ones, mirroring human behavior. For example, they were more "surprised" by "The housewife is quite messy" than "The housewife is not tidy." In contrast, masked language models (like BERT variants) did not show this significant interaction, suggesting they are less sensitive to negation's role in modulating stereotypical meaning. The effect was more pronounced in larger autoregressive models, though smaller ones also showed the trend.

Limitations and Future Work

The authors acknowledge several limitations:

  • Perplexity as a measure: Perplexity can be influenced by sentence length and word frequency, and while they attempted to control for these, perfect balancing across all models and tokenizers is challenging.
  • Template-based limitations: The dataset is based on specific templates and stereotypes from prior work, which may not capture all possible biases or harms. The study's focus on Western-centric stereotypes also limits generalizability.

Future work suggested includes:

  • Exploring other linguistic markers: Extending this framework to investigate other subtle linguistic features like abstraction, modality, or nominalization in relation to stereotypes.
  • Controlled ablation studies: Investigating the impact of model size, architecture, and training objectives more granularly, especially on smaller autoregressive models.
  • Analyzing generated language: Moving beyond surprisal to analyze whether LLMs generate language that preferentially uses negation or other biased linguistic devices, thus actively reproducing stereotypes.

Summary updated

BibTeX
@inproceedings{2025negation,
  title = {{The Negation Bias in Large Language Models: Investigating bias reflected in linguistic markers}},
  booktitle = {COLM},
  year = {2025},
}
Figure