How IBM’s Watson scores the tone of Trump’s tweets


That is Donald Trump’s most-retweeted tweet since he was elected president. He sounds happy, right? Not so much, according to IBM’s Watson. The artificial intelligence software’s Tone Analyzer tool measured the tweet’s tone as more likely to be perceived as sad than joyful, as shown on the chart below.

Let’s try his second most-retweeted, post-election tweet.

Too bad passive-aggressive isn’t an option. Of the five available emotions — anger, disgust, fear, joy and sadness (like the characters in “Inside Out”) — let’s go with joy; he did say “Happy New Year” and “Love” with an exclamation. Right again, says Watson, which scored the tweet as much more likely to be perceived as joyful than any of the other four emotions.


Okay, now his third most-retweeted, post-election tweet.

Well, death proclamations are usually sad, but they’re not usually exclamatory. I can cheat by knowing that Trump followed up this tweet by calling the late Cuban leader “a brutal dictator.” So he’s probably happy about his death. Joy, lock it in. Nope, knock it off, says Watson. The tweet is actually highly likely to be perceived as sad and not likely at all to be considered joyful (or angry or disgusted or scared).


Let’s try his fourth most-retweeted tweet.

Of the five emotional options, I’d say Trump sounds disgusted the most. And according to Watson, I’d be right. But I’d be even more right if I also sensed a little sadness, as IBM’s artificial intelligence technology did.


Let’s finish with number five.

Angry, definitely angry. And definitely wrong, says Watson. While Watson measured this tweet as less likely to be perceived as any of the five emotions, it scored it as much more likely to be perceived as sad than angry, disgusted, scared or joyful.


So what does any of this mean? Am I wrong because I’m not judging the tweet solely on its text, while Watson did? Is Watson wrong because, smart as artificial intelligence software has become, it can’t intuit emotion at a human level? To be honest, I’m not sure, and IBM declined to participate in this story.

Artificial intelligence has become a popular topic in the marketing industry, dominating discussion going into and coming out of this year’s CES. And artificial intelligence software is being pitched to marketers to use for everything from powering their Facebook Messenger bots to analyzing customer service complaints to predicting how people might respond to a brand’s tweet. With artificial intelligence gaining in importance, it seems like as good a time as any to gauge its insight.

According to case studies on IBM’s site, Watson’s Tone Analyzer tool can be used by brands to work out whether a potential tweet is more or less likely to get retweeted and liked, as well as by public speakers to ascertain whether people will like their TED talk. I decided to apply the tool to Donald Trump’s tweets.

In addition to the aforementioned five most popular post-election tweets, I took @realDonaldTrump’s entire corpus since he was formally named the Republican presidential candidate in July 2016, and I ran the tweets’ text through Watson’s Tone Analyzer to study the tone of Trump’s tweets, to see if it had changed in the transition from presidential candidate to president-elect and, if so, how.

In general, Watson scored Trump’s tweets as being primarily joyful, emotional, compassionate, thoughtful and not exerting certainty or inhibition. And aside from being slightly angrier, the president-elect’s Twitter tone hasn’t changed much. Like I said above, I’m not sure what to make of Watson’s results. So I’ll leave it up to you to judge for yourself by checking out the charts below.

But before you do, a quick explanation of my methodology. Because Watson’s Tone Analyzer has a limit on how much text it can process at a time, I separated the tweets into three time periods: 1) from the day after being nominated to the day before the first presidential debate; 2) the day of the first presidential debate until Election Day; and 3) the day after Election Day until yesterday. These time periods seemed to be the clearest stages in which the tone of Trump’s tweets may have changed, from the freshly nominated candidate to the candidate in the final throes of a tumultuous election to the president-elect.

After running the text for each period through Tone Analyzer, I plotted the scores into charts for each tonal category to compare the results. Each chart also features short explanations of the category that are culled from IBM’s site.

The scores are provided on a scale from 0.00 to 1.00. Generally, the higher the score, the more likely the text is to be perceived as having a given tone. For example, a text scoring 1.00 in “Joy” is considered highly likely to come across as joyful, whereas a text scoring 0.00 isn’t very likely to come across as joyful at all. And a text scoring a 1.0 in “Agreeableness” is likely to be perceived caring or humble, whereas a 0.00 is more likely to be perceived as selfish or arrogant.

These categories measure the feelings being expressed, such as how joyful or how angry a text is. Overall, Trump’s tweets were judged to be more joyful than any other emotion and not very angry or disgusted.


These categories measure the personality traits exhibited, such as how sensitive or thoughtful the person who wrote the text is. Overall, Trump’s tweets were judged to be emotional, compassionate and thoughtful.


These categories measure writing style, such as whether a text is analytical and whether it is written with certainty. Overall, Trump’s tweets were judged to not exert certainty or inhibition.


Source link

Facebook Comments


4 + four =