Rethinking Semantic Space: Study Challenges 70-Year-Old Language Model With Survival-Driven Framework
Researchers at the University of Vermont find human language is fundamentally organized around power, danger and structure, a finding that could change how machine learning systems analyze text.
Published
Last updated / Fact-checked on
Computational linguists and machine learning researchers have long relied on psychological theories to represent emotion and semantic nuance in automated text systems. For more than seven decades, sentiment analysis, content moderation classifiers and affective computing pipelines have leaned heavily on the valence, arousal and dominance (VAD) model, a framework rooted in mid-20th-century psychology research, including the work of Charles Osgood.
Research from the University of Vermont (UVM), published in Science Advances in May and highlighted by ScienceDaily on Oct. 24, challenges that foundational paradigm. By applying large-scale computational text analysis to billions of word uses, researchers demonstrated that the classical VAD framework leaves critical semantic relationships unaccounted for. Instead, they found, human language is organized around a survival-oriented framework called "ousiometrics," built on three distinct dimensions: power, danger and structure.
What the Research Found
The study, led by Peter Sheridan Dodds with Christopher Danforth, Julia Witte Zimmerman and colleagues at UVM's Computational Story Lab, examined English word usage across large corpora, including digitized books, news reports, social media posts and transcripts of spoken dialogue.
Key findings include:
- Failure of independence: The traditional VAD framework assumes valence (positive vs. negative), arousal (active vs. calm) and dominance (dominant vs. submissive) operate as independent axes. Computational modeling revealed these axes are significantly entangled and fail to capture how humans categorize real-world meaning.
- A new dimensionality: The team identified three alternative, independent dimensions that govern lexical meaning:
- Power: weak vs. powerful
- Danger: safe vs. dangerous
- Structure: structured vs. unstructured
- Explained variance: While the classic VAD model accounts for roughly 72% of the variation in measured word meaning, the new power-danger-structure framework accounts for more than 90%.
- The safety bias: In everyday communication, words linked to safety appeared systematically more often than words associated with danger. The researchers concluded that the long-studied "Pollyanna principle," the universal tendency for human communication to lean toward positive vocabulary, is not primarily about optimism. Instead, the positivity bias is a one-dimensional projection of a deeper survival bias that prioritizes safety and environmental predictability.
Quantitative Comparison: VAD vs. Ousiometrics
| Metric / attribute | Traditional VAD framework | Ousiometrics framework |
|---|---|---|
| Primary dimensions | Valence, arousal, dominance | Power, danger, structure |
| Variance explained | ~72% | >90% |
| Core assumption | Emotion-driven affective reaction | Evolutionary survival and environmental ordering |
| Explanation of positivity bias | Psychological preference for pleasant affect | Functional projection of an underlying safety bias |
| Data scope evaluated | Human-rated word lists | Billions of word uses in large English corpora |
Why It Could Matter for Machine Learning
For natural language processing (NLP) and machine learning practitioners, language representations underpin token embeddings, safety alignment and alignment evaluation benchmarks.
- Refining sentiment and tone classifiers: Automated moderation tools can misclassify defensive, urgent or high-arousal text as malicious or negative, because simple positive/negative valence maps fail to separate safety from affect.
- De-biasing large language models (LLMs): LLMs trained on internet text inherit the human safety bias directly in their probability distributions. Understanding that text corpora over-represent safe, ordered contexts could help researchers better interpret model behavior in hazardous scenarios.
- Interpretability in feature space: By replacing tangled dimensions with axes that explain more than 90% of lexical variance, automated systems may be able to evaluate document tone and conversational dynamics more precisely.
Two of the study's co-authors work at MassMutual Data Science, a sign of industry interest in measuring meaning at scale.
Sources
- Dodds et al., "Ousiometrics: The essence of meaning aligns with a power-danger-structure framework instead of valence-arousal-dominance," Science Advances, May 6, 2026: https://doi.org/10.1126/sciadv.adr4039
- University of Vermont, "Human Language is Biased Towards Safety, Major Study Reveals, Challenging 70-Year Scientific Consensus," May 6, 2026: https://www.uvm.edu/uvmnews/news/human-language-biased-towards-safety-major-study-reveals-challenging-70-year
- ScienceDaily, "A hidden bias in human language challenges 70 years of science": https://www.sciencedaily.com/releases/2026/10/261023023738.htm