Year of the Election Part II: A Tone Shift in the U.S.

A more human understanding of news can help us look at the presidential race this summer

As we continue through the “super election” year, voters around the world have been casting their ballots to determine who will represent them.

People may vote on policy or party, though in many countries the electorate chooses individuals, with some percentage of the final outcome determined by how voters “feel” about their options. 

Measuring how people feel about things online is a tricky business, with much work in AI focusing on what is called sentiment analysis, looking at text to see how someone feels about a certain subject. Classic examples include running sentiment analysis on movie or Amazon reviews to see if someone likes a film or product.

But a lot of sentiment analysis also takes a decidedly unhuman approach to how people feel, and flattens the entire world of people’s emotions to a simple negative to positive scale, like -1 to +1. It also plays a lot on the perceived sentiment of individual words, which leads to silly situations like a leading Big Tech sentiment analysis tool saying that the headline “Drug company slashes cancer risk” is -0.6 for sentiment despite the fact that any human would understand that the headline is saying something positive. 

With that in mind, Overtone has now built a model that looks at tones in a text, focusing at first on categories including, angry, sad, fearful, happy, funny and hopeful. In the drug company example above, our model says that the text has a 99% probability of being hopeful. This move to more human-understandable tones also means that it can be used in places that require more nuance, like news.

A shift in tone

You can use tone analysis on anything from mentions of your CEO to mentions of a meme. In the UK we differentiated between party leaders who had sad, angry and funny mentions (The one with the most sad mentions leading up to the election lost in a landslide). But you can also look at these sorts of tones over time to see the twists and turns of a race as a campaign unfolds.

The graph below looks at thousands of news paragraphs that mention U.S. President Joe Biden, former President Donald Trump, vice presidential candidate J.D. Vance and Vice President Kamala Harris from late June, the week of the Biden-Trump debate.

Our models rated each paragraph as to the percentage probability of the tones within it, and in aggregate the chart tells the story of the events that have taken place, including the attempted assassination of Trump on July 13 (a spike in fear and sadness in mentions) and Biden dropping out and endorsing Harris on July 21 (a surge in hopefulness in mentions). 

This sort of analysis is definitely useful in the media monitoring space, though we are also excited for publishers to begin to use it to tackle news avoidance, where readers regularly and repeatedly say that one of their biggest problems with news is that it is too negative. By understanding tones, newspapers can send the right article to the right reader based on their user needs, and package different sorts of articles together (much like how in the past the physical newspaper had the big scandalous political stories and the comics all in one bundle).

News vs Social

You can also see in the chart above that the predominant “tone” conveys information in a relatively neutral way, which we call informational. That is to be expected given that the paragraphs were from news articles. 

However, not all platforms are like this. We ran our new tone model on posts from social media that we collected, also mentioning the same four candidates, to compare. What you see is one tone dominating above all the rest, and that is anger. 

This finding is particularly interesting from a business perspective, where both news sites and social media platforms are competing for the same pool of ad dollars from companies that want to place their campaigns next to content where readers are paying attention. It appears to confirm that news has a much less angry tone than social media.

The tone of social media also plays on brand safety concerns those advertisers have and continue to have, especially in turbulent political times. They do not want to be associated with content that is offensive or spouting conspiracy theories, and often would want their message next to something more inspiring, like the Olympics. 

Elon Musk has sued a group of advertisers and the Global Alliance for Responsible Media over controversies about brand safety on X (formerly Twitter), though an analysis with Overtone’s brand safety models also shows that social media tends to be a place where the risk is higher. Of the articles and posts mentioning the candidates, the vast majority in all circumstances on both news and social media didn’t mention conspiracies, or discuss/promote violence or offensive speech. However, the percentages in those risky categories were much higher on social media.

News had less than one percent in the High risk category for each of the three risks, though social saw small but significant portions, including more than 7% of all candidate mentions that were High risk for being offensive. If businesses want to avoid being caught up in angry online discussions, and even work towards rewarding a change of tone in online discussions, supporting news would be a good place to start.

If you’re interested in hearing more, please get in touch.