Visualizing 2016 Media Sentiment

Problem

The 2016 American election has been one of the most polarizing. Since popular election for Senators began in 1914, 2016 is the first year that the same party has won both the Senate seat and the Presidential electors for every single state. We are increasingly living in a world of our own creation. We create our own echo chambers and rarely understand the information we consume in its broader context. This is dangerous because we begin to believe that our echo chamber is actually the reality we live in. When we get a glimpse of the reality, it can often come as a huge shock.

This problem is further compounded by the plethora of scandals, embarassing debate performances, and other significant events that stick in people's minds due to the recency heuristic and anchoring biases people use when making decisions. People are more likely to remember recent events and anchor to them when making decisions. In an election cycle with many such significant events, it becomes important to understand how these events influence sentiment.

Solution

We need a solution that addresses the problem of the echo chamber and helps individuals understand how key events affect biases of their news sources. We offer a partial solution to this problem. By visualizing the basic sentiment of your news sources, we can begin to understand the echo chambers we live in. For example, if my favorite news sources are The Washington Post and The Huffington Post, I can begin to see how these news sources discuss the 2016 presidential candidates in the broader media landscape. This tool will not remove the problem of echo chambers. However, this tool will give us a greater awareness of the echo chambers we inhabit. Additionally, we have the opportunity to visually show how key events swayed the media towards or against the candidates inside and outside of our echo chambers.

What is Sentiment?

In this context, sentiment is the general positive or negative tone of an article. The sentiment scores were produced by IBM's Watson Alchemy API. Sentiment scores are centered at 0, which is neutral. Sentiment scores of 1 are very positive, and sentiment scores of -1 are very negative. For example, an article about Hilary Clinton that has a 0.25 sentiment score is a slightly positive article about Hilary Clinton. An important caveat here is that a positive sentiment score does not necessarily mean that the article supports Hilary Clinton. In the same sense, an article with a negative sentiment score about Donald Trump is not necessarily criticizing Donald Trump. However, we can make inferences about a news source's biases based on the sentiment (in aggregate) of the articles they publish about a given candidate.

Methodology

AlchemyData News, is a service provided under IBM's Watson Developer Cloud. The API gathers and analyzes between 250K and 300K English news articles and blog posts every day. We queried AlchemyData News for the two major candidates (Donald Trump and Hilary Clinton) every day during the Fall of 2016. Due to the extensive amount of data, we focused on only the top 25 news producing sources by article volume. To further simplify the visualization, we summarized the data for each news source by day by taking the average of the sentiment scores. We plotted the averages in a scatterplot over time. However, because when taking averages, we lose some of the granularity of the individual data points, we also decided to include histograms that showed the distribution of scores for each news source. We support this website on a Flaskwebserver.