On May 14th, First Lady Melania Trump had kidney surgery. As time went by, people started to remark that no one had seen her since that date, and by May 29th, conspiracy theories about her whereabouts had gone mainstream. Then on May 30th, the following message was posted on her @FLOTUS twitter account.
I see the media is working overtime speculating where I am & what I'm doing. Rest assured, I'm here at the @WhiteHouse w my family, feeling great, & working hard on behalf of children & the American people!
— Melania Trump (@FLOTUS) May 30, 2018
I’ve read thousands of Donald Trump’s tweets, and this one sounds just like his. Apparently I’m pretty late to the game, because within hours of this tweet being posted, both social media and mainstream journalists started asking whether Mr. Trump had actually written it.
Here’s an excellent article by forensic linguist Claire Hardaker that takes a deep dive into the question. As she put it, the conclusion is “a messy one” because the tweet has both consistencies and inconsistencies with the twitter corpora of both Mr. and Mrs. Trump. So I decided to feed the text through my classifier classifier to see what machine learning has to say.
Did Mr. Trump write Mrs. Trump’s tweet?
For each @realDonaldTrump and @POTUS tweet in my database, I report a final “machine-learning probability” of whether Mr. Trump wrote the text. However, as I explain in my discussion of how the model works, I’m actually running two classifiers. The first returns a probability based on the actual words themselves (text), and the second uses just the grammatical structure. I combine these to determine a total probability, based on how accurate each classifier was on the same validation set. For this exercise, I think it’s important to break out each result on its own, just so you have full disclosure.
Based on the actual text of the tweet, my classifier returns a 77.0% probability that it is consistent with the tweets written by Mr. Trump himself. Based on the grammatical structure of the tweet, my classifier returns a 91.3% probability that it is consistent with Mr. Trump’s tweets. These combine together to give a total probability of 84.3%.
My classifier gives an 84.3% likelihood that this tweet is consistent with those that Mr. Trump has written himself.
Who is writing @FLOTUS tweets?
Could Mrs. Trump have written a tweet that has similar structure and syntax to those of her husband? Certainly. Up until yesterday, I had never read any of Mrs. Trump’s tweets. However, if we are going to ask who wrote a tweet, we have to start with a body of text that we believe was written by that person. I noticed when I pulled up the @FLOTUS twitter page that “This account is run by the Office of First Lady Melania Trump.” That means any number of staff could be writing the tweets just as easily as she could.
To really know what Melania Trump sounds like, we have to go to her personal twitter feed. Since taking office, that page has mostly been retweets from @FLOTUS. Just in case someone was writing her tweets during Mr. Trump’s campaign, we really need to see what Mrs. Trump was writing like before mid-2015.
Most of the @MELANIATRUMP tweets look like this:
— MELANIA TRUMP (@MELANIATRUMP) August 15, 2014
That is, they’re mostly posted photos with very brief titles, using a lot of hashtags and emojis. Short moments of time, captured and shared, with descriptions that match. That’s almost completely inconsistent with the body of tweets in @FLOTUS. Here is a totally random example:
It was wonderful to welcome guests to the @WhiteHouse this evening in celebration of National African American History Month. We look forward to hosting many more events that celebrate our unique and diverse cultures. #BlackHistoryMonth2018
— Melania Trump (@FLOTUS) February 13, 2018
Honestly, I think some of her staff are writing some subset of the @FLOTUS tweets. That seems par for the course, given how political offices work. What first made me think this wasn’t the mystery tweet, but how the work “with” is abbreviated.
In the @MELANIATRUMP tweets, if she abbreviated the word, it was as “w/”. In the @FLOTUS tweets, it switches back and forth between “w/” and “w” (as in the mystery tweet). I’m not a forensic linguist, but as someone who has been writing SMS messages for over 20 years, I know that I don’t switch back and forth between these. My wife doesn’t either. I imagine that many people don’t either.*
Just as with the @realDonaldTrump account, I suspect there are multiple authors for the @FLOTUS account. There have been some news articles recently about how Mr. Trump’s staff writes draft tweets that already sound like him (including grammatical errors). I wouldn’t be at all surprised if someone on Mr. Trump’s staff had written this tweet for whoever on Mrs. Trump’s staff issues @FLOTUS tweets. Or maybe Mr. Trump actually did take her phone and tweet it himself. I doubt we’ll ever know.
Dr. Hardaker was very kind and shared her database of @FLOTUS and @MELANIATRUMP tweets with me. I’m running all of them through my Did-Trump-Tweet-It classifier right now to see if anything interesting pops up. In particular, I’m curious whether it agrees with her follow-on article here. Second, I’m going to try to extract enough text from the @MELANIATRUMP database (i.e. not just 2-5 word photo descriptions) to train a classifier to distinguish between her tweets and those of her husband. If successful, I will post all of those in this archive and you will be able to find them by searching for the @FLOTUS and @MELANIATRUMP tags.
Just to be clear to readers, I’m not a linguist nor an expert in forensic textual analysis. Anything after the machine-learning prediction in this article is my own conjecture and is not based on any rigorous analysis. *For example, Dr. Hardaker noted privately that my differentiating “w” versus “w/” is that we don’t have enough ground-truthed information about her writing to develop her idiolect profile. She may very well use both abbreviations depending on the context, to switch back and forth.