Posted on

Looking back at the U.S. Election 2016

It would be easy to only remember the scandals and publicity stunts that have plagued the 2016 U.S. Election. But the race had some substance too, and each candidate had a very different country in mind when campaigning for presidency.

I’ve gathered transcripts from all primary and general election debates (thanks to the American Presidency Project), and computed each candidate’s political footprint: a vector-based representation of political discourse in which each vector represents a word. What is unique about this technique is that it not only detects words most frequently used (word clouds), what emotions they have been accompanied with (using IBM Watson Natural Language Understanding), but also how closely each word is related to one another, thanks to GloVe, a pre-trained word vectors space developed by Stanford University.

Here are below the political footprints of the three main 2016 presidential candidates: Bernie Sander (during the primary election), Hillary Clinton and Donald Trump (both during the general election).

From left to right: Bernie Sanders, Hillary Clinton, and Donald Trump’s political footprints during 2016 U.S. election debates, with the closest words to “people” highlighted.

Let’s have first a closer look at each candidate’s key topics (most relevant terms according to IBM Watson). It is no surprise that Wall Street came first as a key topic for Bernie Sanders. His political footprint shows a campaign focussed on a relatively limited set of terms and keywords.

Bernie Sanders’ key topics detected during 2016 U.S. election televised debates.

Donald Trump talked about trade deals with sadness. He was angry when talking about ISIS, and disgusted when talking about Clinton. His use of his opponent’s name appears to have been more relevant than for the two other candidates.

Donald Trump’s key topics detected during U.S. election televised debates.

Hillary Clinton was very much focussed on new jobs and the affordable care act. Most notably, IBM Watson struggled to identify her emotions, which adds to her reputation of being hard to read. We can see A.I.’s limitations: Hillary Clinton spoke with joy about Donald Trump probably because of her sarcastic tone.

Hillary Clinton’s key topics detected during U.S. election televised debates.

One term that kept coming back in all political footprints was “country”. Why is that and what does the term actually mean? This is when the true power of word vectors comes into play. Political footprints allow us to find what words were closely related to “country” in each candidate’s vocabulary, how relevant they were for them, and what emotion they were accompanied with.

Here are words from Hillary Clinton that were closely related to “country”. We can roughly group them in three categories: foreign affairs (Europe and Iraq), the economy (raising fear), and national themes (America, national, or people).

Hillary Clinton’s topics that were related to the country (U.S. election televised debates).

Bernie Sanders felt visibly bad when talking about country-related issues, in both national and international affairs. Interestingly, his political footprint is the only one where we can see a connection between the country and its citizens.

Bernie Sanders’ topics that were related to the country (U.S. election televised debates).

Not surprisingly, China makes its appearance in Donald Trump’s political footprint.

Donald Trump’s topics that were related to the country (U.S. election televised debates).

Let’s illustrate again the power of political footprints with two terms that have played a central role in Donald Trump and Hillary Clinton’s campaigns: trade (deals) and the (affordable) care (act).

Terms that were closely related to trade (deals) in Donald Trump’s vocabulary where Nafta (used with fear), China, Korea (?), the world (used with sadness), countries (disgust), and other economic terms such as tax, business, and companies.

Donald Trump’s topics that were related to trade deals (U.S. election televised debates).

Hillary Clinton’s care-related vocabulary included terms such as women (used with disgust, probably in the context of gender equality), children, insurance and health (both presented positively).

Hillary Clinton’ topics that were related to the affordable care act (U.S. election televised debates).

All this seems relatively straightforward for anyone having followed the 2016 U.S. elections. The difference is that all the above political footprints have been generated without any human intervention or assumption about what each topic was about: words have been automatically selected using IBM Watson, and grouped using GloVe pre-trained word vectors.

To be clear, political footprint aren’t unbiased: word similarities have been inferred by GloVe using large corpora of text (Wikipedia and Gigaword 5), and will thus reproduce any cultural bias from these corpora. Every language is ultimately biased, every word has a genesis and an history, and political discourses should always be understood in their cultural context. We might think of political ideas as being universal, but they are always rooted in a certain culture, and applied at a certain time of history.

« Le mythe est constitué par la déperdition de la qualité historique des choses : les choses perdent en lui le souvenir de leur fabrication » Roland Barthes

(Myth is the loss of the historical quality of things: in it things lose the memory that they were once made.)

Political footprints is a tool that I hope can help researchers and journalists to reduce their reliance on personal opinions when analysing political discourse. In times when little phrases and social media trends are more commented than any political vision, political footprints’ aim is to help bring political discourse back at the centre of public debate. All instructions to generate political footprints are available on Github