THE SCIENCE OF POLITICAL POLARIZATION

CHRIS QUIRK

Political polarization in the United States so noticeably on the rise in recent years has, according to the recent findings of a team of Carnegie Mellon researchers, reached the point where the same words can mean different things to those on the left and right sides of the political spectrum. We may now be, to paraphrase George Bernard Shaw, one nation, separated by a common language.

Ashiqur KhudaBukhsh, a project scientist in the School of Computer Science’s Language Technologies Institute, along with Mark S. Kamlet, University Professor of Economics and Public Policy, Founders University Professor Tom Mitchell, and research engineer Rupak Sarkar studied the vocabularies of commenters on network news videos on YouTube, and found that the meanings of particular terms used by political opponents were near precise opposites. “We presume we are speaking in English,” said KhudaBukhsh, “but our political discussions are happening in two different languages.” KhudaBukhsh presented the team’s paper, “We Don’t Speak the Same Language: Interpreting Polarization Through Machine Translation,” at the 35th AAAI Conference on Artificial Intelligence in February 2021.

The algorithms primarily used to perform language translation, like Google’s word2vec, take as their source a 1957 insight by linguist John Wirth, “You shall know a word by the company it keeps.” The algorithms work spatially, setting words in relation to others as they appear in the data set. Given the word “hat,” nearby might be the word “head” or “straw,” and a bit further away you might find “fedora” or “winter.” To translate “hat” into Italian, an algorithm could essentially work backwards from the locations of “testa” (head) or “paglia” (straw) and the other words nearby on the Italian graph to infer that “cappello” equaled “hat.” The insight the research hit on was to separate online speech of left-leaning and right-leaning commenters into separate “languages” and translate terms from one language into the other to find similarities and differences. “I believe it is a novel approach,” said Mitchell. “I don’t think anybody’s thought about machine translation to go from Democrats to Republicans.”

To run their experiment, the team captured and processed more than 86 million comments from videos on YouTube produced by media sources like CNN, MSNBC, Fox News and One America News Network (OAN), an outlet founded by pro-Trump businessman Robert Herring. Then they created a language for each news source. Finally, they compared terms across the languages, effectively translating words from one news source language to another. “When we translate from one news source language to another, in an ideal world not fraught with polarization, of course all the words should translate to themselves from language to language,” said KhudaBukhsh.

The divisions stretched into climate policy, where “solar” in CNN translates to “fossil” in Fox. Some of the results were jarring. For instance, commenters from the left and right had “BLM” and “KKK” as misaligned pairs, meaning Fox commenters see Black Lives Matter supporters similarly to the way CNN commenters see the Ku Klux Klan. “Those differences were disturbing,” said KhudaBukhsh.

Diversity of opinion and spirited debate may be the lifeblood of a healthy political system, but Kamlet says the team’s results point to divisions between political factions that aren’t resolvable via the typical political protocols. “There’s a common way of thinking about policy differences. One person may want nine aircraft carriers for the national defense. Eight would be a little worse, 10 better. Someone else wants six. We can disagree politically, but there’s a nice, natural order.” For Kamlet, the study shows evidence of a much more precipitous social fracture, citing in particular occurrences of what the team calls trigrams, three-word terms that operate as misaligned pairs in the parlance of left and right. Analyzing thousands of occurrences, the team found “black lives matter” and “all lives matter” to be misaligned pairs in the CNN and Fox languages. The trigrams are mathematically more significant than single-word pairs given the unlikelihood [that] three units being so precisely matched across languages. “It’s hard to see how the polarization could be more extreme than it is,” said Kamlet.

In looking to social media for the massive data set needed to produce their study, the researchers were being pragmatic, but also landed in the cauldron of one of the causes of the polarization they surveyed. Divisive speech and extreme content are not just unfortunate byproducts of online life.

Social media companies frequently profit from them, as inflammatory content is catnip for users, and engagement translates into profit. Social media platforms seeking a competitive edge seem to have a hard time spurning the allure of the baser tendencies of our nature. “Sadly, I agree that controversy generates more eyeballs on your site than boring stuff, so social media might be hurting its profit if they produce a more simple discussion,” Mitchell said. “The same thing can be said about news organizations like MSNBC and Fox. I think it’s good for their profit to be extreme.”

Evolving structural issues are also destabilizing the political status quo, said Kamlet, citing gerrymandering as a prime culprit. “Social media can be an echo chamber, but gerrymandering has the same effect of sending people toward the extremes.”

Most developed democracies have an upside-down U-shaped statistical distribution of political opinion — a large group in the center, and long tails to the left and right. “In political science they call this the distributed voter theorem,” he explained. “If the president gets way off one way or the other, there are checks and balances.” Gerrymandering throws the political system out of whack by pushing elected officials farther out along the curve. “There may only be around 20 truly competitive congressional districts at this point, so if you are a Republican running in a safe district, you’re not afraid of a Democrat coming out of the woodwork to beat you, but you’re frightened to death of a challenge in the primary from the right.”

While the team’s research has shined a bright, unflattering light into the crevices riddling the country’s political culture, the broader applications of their work may be just beginning. Mitchell has been ruminating on using the tool to revolutionize polling methods. “This paper is kind of an indicator of a broad movement toward detailed data analysis of very large-scale social media for the purpose of understanding where their agreements and disagreements, and what their opinions are,” he said. “I think it’s fair to ask if, in the coming decade, political polling will be replaced by something that’s more accurate and based on monitoring what tens of millions or hundreds of millions of people are saying. It could be a tool for polling predictions and even candidates in political races to understand in a more precise way what voters are thinking.”

According to Kamlet, given the varied forces that have created the political fissures in our system, it will be necessary to work across the divisions between siloed disciplines to find solutions. “Disciplines can speak their own languages too, and disciplines can be hard to connect,” said Kamlet. “But when you can get domain expertise and technology strengths together like we are able to do, that’s a very potent thing.”

As an alternative to squelching online hate speech, KhudaBukhsh, along with colleagues Shriphani Palakodety, an engineer at Onai, and Jaime Carbonell, who was the Allen Newell Professor of Computer Science until his death last year, studied the potential for boosting positive speech. “Little attention is given to the possibility that non-hate-speech in social media discussions could have a beneficial societal impact,” KhudaBukhsh said. Reviewing online comments in India and Pakistan in 2019, when border skirmishes between the two nations broke out in Kashmir, the researchers found that as the two nations edged closer to war, online speech became less forceful. “Peace-seeking comments heavily outnumbered war-seeking comments, and there was an outpouring of comments and likes in a way I’d never seen before,” said KhudaBukhsh. “I believe social media discussions contain a lot of positive content, and if they don’t get drowned by hate, this world could be a much better place.” ■