By Roberta Kwok

We tend to become friends with those who are like us—for example, people with the same ethnic or religious background or a similar personality. A new study co-authored by Yale SOM’s Balázs Kovács suggests that invisible as well as visible cues can be part of this process. Specifically, the study examined whether subtle similarities in the way people write or talk, such as the types of words they use, make a difference in determining whether a relationship forms. 

Kovács and Adam Kleinbaum of Dartmouth College’s Tuck School of Business studied online connections between Yelp users and found that people do seem to be drawn to others with similar linguistic styles. A similar, though weaker, trend emerged in a study of in-person friendships at a university. And after a pair of people became friends, their use of language tended to remain more similar over time.

“These are dimensions that you might not think about as important,” says Kovács. But “linguistic stuff matters.”

The findings provide more evidence that people tend to gather in echo chambers, Kovács says. If people cluster with like-minded peers, and then those friends become even more similar over time, society fragments into polarized camps.

“What you would get are pockets of society that are very similar to each other,” he says. Linguistic style likely isn’t the most important factor in creating these connections, but “maybe it’s part of it.”


Read the study: “Language-style similarity and social networks”

Kovács and Kleinbaum started by looking at online relationships. They speculated that language might play a significant role in this setting. After all, people have much less information about potential friends in a digital environment than in person, where they can evaluate other factors such as appearance, facial expression, and tone of voice.

Online, “you mostly have no idea who these people are,” Kovács says. Instead, users mainly assess others based on what they write.

To investigate, the team analyzed 1.7 million reviews on Yelp by nearly 160,000 users in seven metropolitan areas in North America: Phoenix; Las Vegas; Pittsburgh; Urbana-Champaign, Illinois; Charlotte, North Carolina; Madison, Wisconsin; and Toronto. 

The researchers used a linguistic coding framework developed by another team to score each user in 18 categories. For example, they determined how often the person used first-person singular pronouns such as “I,” “me,” and “mine”; negations such as “not” and “never”; and comparison words such as “greater” and “best.”

For each possible pair of people within each metropolitan area, the researchers computed the similarity of their linguistic styles. Then the team checked whether that pair had tagged each other as friends on Yelp.

They found that a one standard deviation increase in similarity was linked to a 26–76% higher likelihood of friendship, depending on the metropolitan area. The effect was strongest in Las Vegas and weakest in Urbana-Champaign, but “the main pattern is the same everywhere,” Kovács says. And higher linguistic similarity in August 2016 was linked to a higher likelihood of a new connection forming between those users over the next five months, suggesting that the matching linguistic styles had helped cause the friendships.

The more similar students’ linguistic styles were, the more likely they were to remain friends. And friends’ use of language tended to converge more over time than if they weren’t close.

The researchers were also curious about whether people tended to adopt their friends’ linguistic tics over time. They found that, on average, Yelp users’ linguistic styles slightly drifted apart from each other from August 2016 to January 2017. But the gap widened less if a pair of people were friends in 2016; in other words, their styles remained more similar.

The online study left some unanswered questions. The researchers didn’t have much information about the users, so they couldn’t control for demographic factors and isolate the effect of language alone. It also wasn’t clear whether their results would hold up for in-person relationships.

So Kovács and Kleinbaum gathered data from 247 students who had entered a university graduate program in the United States. They obtained two sets of writing samples: application essays written before they had met their peers and exam essays written two months into the program. The researchers also collected demographic details and gave the students a personality test. One month and six months after the program started, the students took a survey asking them to list the classmates with whom they spent the most free time.

Again, the team coded each student’s linguistic style and the similarity of every possible pair of people. But this time, the researchers also controlled for factors such as gender, race, nationality, and personality characteristics.

The general pattern was similar to that of the Yelp study, but with a less dramatic result: a one standard deviation higher similarity score was linked to a 5% higher likelihood of forming a friendship. The more similar their linguistic styles were, the more likely they were to remain friends by the end of the study. And when people were friends, their use of language tended to converge more over time than if they weren’t close.

Why was the effect of language weaker for in-person relationships than online? The students likely relied on many other signals such as attractiveness and age to choose their friends. With the Yelp reviews, “they only see text,” Kovács says.

Kovács speculates that the effect of language may go beyond friendship formation. Perhaps people also are more likely to listen to others who have similar linguistic styles—which would affect the flow of ideas and opinions through society.

“It’s a more general phenomenon than just friendship,” he says. “It’s about who influences whom.”