Language leans to safety, not just positivity: 20,000-word study challenges a 1950s model

Human language leans towards safety more than towards happiness, a team at the University of Vermont says, and the finding questions a way of measuring word meaning that psychologists have used since the 1950s. The study, "Ousiometrics", is published in Science Advances and was summarised by ScienceDaily from the university's release.

The 70-year-old idea

Since psychologist Charles Osgood's work in the 1950s, researchers have described the emotional meaning of a word on three scales: valence (pleasant or unpleasant), arousal (calm or exciting) and dominance (in control or controlled). This VAD model sits behind many tools that score the mood of tweets, reviews and news.

The Vermont team, led by Peter Sheridan Dodds and Chris Danforth, argues that those three scales are not independent of each other, as the model assumes. They propose a different set:

• Power: how strong or weak a thing is;
• Danger: how threatening or safe it is;
• Structure: how ordered or chaotic it is.

The numbers

The researchers scored more than 20,000 English words and tracked billions of uses of them across books, news, social media and speech. Their power, danger and structure framework explained more than 90% of the variation in measured word meaning, against about 72% for the older VAD model.

They then ran their "ousiometer" over a wide spread of English texts: Jane Austen's novels, the Sherlock Holmes stories, the New York Times, Wikipedia, Twitter, talk radio transcripts and an English translation of Les Miserables.

Positive, or just safe?

For decades scientists have noted the "Pollyanna principle": people use positive words more often than negative ones. The new paper says that tilt is really a shadow of something deeper. In the authors' words, the positivity bias "is, in fact, a one-dimensional projection of an underlying safety bias."

"Expressions of safety are crucial to all language," said Dodds, the senior author. Co-author Julia Zimmerman said power, danger and structure "are relevant to every person that's ever lived."

The team is careful about one point: the results fit a link between language and survival, but they do not prove that evolution produced the safety bias.

Why it matters beyond the lab

Sentiment scores built on the old model are used by companies, researchers and AI systems to read public mood, from brand monitoring to election chatter. If the dimensions underneath are wrong, those scores can be off too. The study looked only at English, so whether Urdu, Punjabi or other languages show the same safety bias is an open question for researchers here.

The ten-member team included collaborators from the Santa Fe Institute, the Complexity Science Hub in Austria, UC Berkeley, the University of Adelaide and MassMutual Data Science, with funding from the US National Science Foundation, Google and MassMutual. More research stories are on our science desk, and AI coverage is on the AI desk.