In recent years, Jack Grieve of the department of English and linguistics at the University of Birmingham in England has embraced Twitter as a bountiful lode for looking at language-use patterns. One of his projects examined the regional popularity of profanity in the U.S. (“crap” is big in the center of the country; “f—” turns up more on the coasts). Another study he conducted looked for new word usages spreading on American social media (“baeless” for single, for example, and “senpai” for elder or expert).

Now Grieve and his linguist colleague Isobelle Clarke have turned their analytic expertise to President Donald Trump’s Twitter account. A study published Wednesday in PLOS One shows how the linguistic style of Trump’s 21,739 tweets from mid-2009 to early 2018 (excluding retweets) morphed as his strategy for reaching multitudes of followers changed. (“Linguistic style” here refers to the form of the text, not its meaning.)

The researchers categorized the different styles by scrutinizing the tweets’ grammatical structure. Some of them were merely conversational. Others dispensed advice or campaign rhetoric or simply signaled Trump’s engagement on a particular issue. One tweet style transmuted into another—and then sometimes back—as Trump progressed from his role as peddler of a fake “birther” conspiracy theory to improbable presidential primary candidate to Republican nominee—and then to occupier of the Oval Office. (Such “style shifting”—how an individual’s language varies from one situation to another—is a subdiscipline in linguistics.)

Scientific American asked Grieve about his trek through the Trump tweet corpus and the richness of Twitter as a source for linguists. [An edited transcript of the interview follows.]

What gave you the idea to do this study?

Back in 2017 there was a lot of talk about whether Trump wrote a specific tweet that might have amounted to an admission of obstructing justice. But his lawyer [at the time], John Dowd, then took credit for it. We do research on both social media language and authorship analysis, so we decided to look at that tweet. We found it was relatively unusual in its style, although we couldn’t rule out the possibility that Trump had written it. After that study, we decided to look at stylistic variation on [his Twitter] account more generally—and especially to track changes in style over time.

Why did you think it was important to bring linguistic methods to the study of a president’s tweets?

The account was so important during the election and such a point of contention—we thought this was an area where linguistic analysis could be informative. In general, when most people look at language, they focus on what is said. But looking at the structure of language—how something is said—is what linguists do. We thought that perspective was largely missing from the discussion—and important.

Briefly, how did you conduct the study?

Basically, we downloaded all the tweets that were posted on the account, removed the retweets, automatically identified a wide range of grammatical features found in each tweet (such as various parts of speech) and then conducted a statistical analysis to extract dimensions of linguistic variation, based on which features tended to occur together in tweets. We then interpreted these dimensions stylistically and plotted them over time to understand how the style of the account changed.

What did you find from your research?

Our main takeaway was that Trump’s Twitter account wasn’t just a bunch of random attacks, as is sometimes claimed by the media. Rather we found a lot of stylistic variation—and crucially, we found this variation was, on the whole, fairly systematic over time: the style of tweets changed in clear ways, depending on what was happening in the campaign. For example, we found that a more informal and conversational style was used during the Republican primaries, compared with the general election.

Did you discover that Trump and his campaign were successfully using Twitter as an alternative communication medium?

I think, to a large extent, we knew that was going on. The amount of free media coverage the account generated during the campaign is well documented. We also know it was a big part of how Trump and his campaign chose to communicate with the public. And we know they won. So our aim was to describe how they used Twitter from a linguistic perspective—not to understand if it was effective but to understand why it was effective. And I think we show there was an underlying strategy—that they were shifting the styles of their tweets to achieve certain communicative goals.

Can you give an example?

We found that Trump and his team employed an increasingly disengaged style over the course of the campaign—essentially presenting their opinions as facts and ignoring alternative viewpoints. This style of communication became especially common after the Access Hollywood tape was released and may have helped minimize the damage from that scandal.

Also, is it true you found that you were actually able to discern when unannounced events were occurring, such as Trump's decision to run for president?

Not with any certainty. But we hypothesized that there was a change in his outlook in February 2015—just as his last season of The Apprentice was wrapping up—based on sharp changes in style at that time. In general, all the major stylistic inflection points we identified correspond to major events in the Trump time line. But this one, we couldn’t explain. And right around that time, he was hinting he would run in his tweets—and we know he formed his exploratory committee not long after. So based on these considerations, we proposed that this might be when he personally decided to run.

Through this and previous studies, what benefits do you believe linguistics can bring to disentangling the impact of social media?

Social media is becoming an increasingly important force in society. A lot of news is driven by online communication, and so there is a need for linguistic analysis. But unlike lots of other areas, everyone thinks they’re a language expert—including journalists and data scientists. This has really led to an explosion of amateur linguistics, in my opinion. For example, there are lots of people out there who claim they can distinguish tweets written by Trump from tweets written by his surrogates, but they lack any reliable baseline for Trump’s style on social media. These kinds of issues are obvious to linguists. And of course, we can do much more detailed, accurate and objective analyses. I hope, over the next few years, we’ll see a rise in investigative linguistic research. I think that’s really important, and my hope is that this paper will encourage more linguists to tackle these types of problems.