For decades, researchers have assumed that meaning could be distilled into three core emotional dimensions: valence, i.e. positive vs. negative; arousal, i.e. excited vs. calm; and dominance, i.e. controlling vs. submissive; collectively known as the VAD framework. Based on pioneering work in the 1950s by Charles Osgood and others, this model has been widely used across disciplines, from psychology and linguistics to sentiment analysis in AI systems.But, a new UVM study finds this theory flawed. Drawing on billions of uses of more than 20,000 words and diverse real-world texts, as well as using modern computational methods, the researchers, with support from the US National Science Foundation, Google, MassMutual and other funders, identified a different set of underlying dimensions. Not only are the VAD dimensions not independent, but they obscure a more fundamental organising principle of language.They found that meaning is best organised by three independent dimensions: power, i.e. weak vs powerful; danger, i.e. safe vs dangerous; and structure, i.e. ordered vs chaotic.At a time when communication technologies are becoming increasingly important, be it in large language models or content moderation, deciphering meaning itself becomes important. Researchers at the University of Vermont have uncovered a new insight into how language works, challenging a cornerstone assumption in psychology, linguistics and artificial intelligence that has stood for more than 70 years.Their study, published on May 6 in Science Advances, introduces ousiometrics, the quantitative study of essential meaning, and suggests that language is organised not around emotion alone, but around a deeper structure shaped by power, danger and order.
How we have measured meaning
How have we been measuring meaning till now?
The VAD framework grew out of the work of psychologist Charles Osgood and his colleagues in the 1950s. Their approach used pairs of opposing words to understand how people perceived meaning. Participants would rate words against scales such as good versus bad, strong versus weak, or active versus passive. Researchers then used these ratings to identify the underlying dimensions that shaped how people understood language.Over time, valence, arousal and dominance became a standard way to measure the emotional associations of words. The framework has been particularly useful in sentiment analysis, which allows computers to classify text according to its emotional tone. A system analysing product reviews, for instance, might identify whether a customer sounds pleased or dissatisfied.However, the Vermont researchers argue that the framework has limitations. Their analysis showed that the three dimensions do not work independently as the model assumes. Instead, they overlap, suggesting that the framework may be capturing aspects of meaning that are related to each other rather than operating as separate units.Instead, the team offers a different system based on power, danger and structure. Power describes the spectrum between weakness and strength, danger distinguishes safety from threat, and structure captures the difference between order and chaos.Take words like leader, earthquake, routine. A leader can summon power; an earthquake can summon danger; routine can summon order and predictability. Every word has emotional associations, but this is not necessarily the most basic aspect of its meaning. The new framework attempts to capture these deeper relationships.
Why language appears to favour safety
One of the study’s most striking findings is that language consistently favours words associated with safety over those associated with danger.Researchers have long observed what is known as the Pollyanna principle, the tendency for positive words to occur more frequently than negative ones. The new findings offer a different explanation for this pattern. The researchers suggest that the bias toward positivity may reflect, in part, a more fundamental preference for describing the world in terms of safety rather than danger.The team examined large collections of English-language texts, including literature, news, Wikipedia, talk radio transcripts and social media. Across these different sources, they found a systematic bias towards words associated with low danger.The distinction is worth making. Positive and safe are not interchangeable, just as negative and dangerous do not mean the same thing. A word can describe something unpleasant without suggesting a threat, while a dangerous situation may be described in positive terms by someone who welcomes the risk. “Safety is a separate dimension of meaning that the traditional model does not fully embrace,” the researchers say.But their results also indicate that the bias for safety trumps any overall preference for words related to power or weakness.
How the researchers mapped the meaning of words
To study these relationships, the team looked at both word types, or unique words, and word tokens, or specific instances of those words in text. That allowed for not only an analysis of the breadth of words in the language, but also how often people actually use them.Scientists have developed ousiograms, visual representations of the semantic relationships between words. These helped them identify patterns in language and compare the proposed framework with the established VAD model.They also proposed an ousiometer, an instrument to measure the change of meaning over time. In one demonstration, the researchers used it to trace patterns across Victor Hugo’s Les Miserables, showing how the language of a long narrative can shift across dimensions such as power and danger.This means we can look at meaning across entire books, news stories or other large bodies of text, rather than one word at a time in isolation. That could help researchers track how themes and tone in a text change over time.
What this could mean for AI
What does this mean for AI and LLMs?
The results could have implications for how artificial intelligence systems read language. Sentiment analysis tools typically classify text as positive or negative, and can also measure the intensity of the sentiment. But a text’s meaning can be more than its emotional tone.A report about a natural disaster, for instance, may be strongly negative, but its meaning also depends on the danger being described, the power of the forces involved and the disruption of the structures people rely on. A simple positive-versus-negative classification cannot capture all these distinctions.A framework based on power, danger and structure could give researchers additional ways to analyse such language. It may be useful to investigate how AI systems recognise threats, interpret descriptions of risk or track changes in public discourse.However, the study does not establish that incorporating these dimensions will automatically make large language models more accurate or improve content moderation. Those applications would need to be tested separately. The research provides a different theoretical framework for understanding meaning, rather than proof that existing AI systems should immediately be rebuilt around it.The study challenges a model that has shaped research for more than seven decades, but it does not make emotion irrelevant to language. Instead, it suggests that the way people organise meaning may be rooted in more fundamental distinctions about power, safety and order. Whether the framework holds across languages and cultural contexts remains an important question for future research.
About The Author
