Sentiment analysis of popular media publications

Reid Cobb

By far one of the most divisive candidates in recent times, President Trump was portrayed in a variety of lights depending on the publication. This dataset from Kaggle provides 142,570 articles from 15 different news outlets from 2016-2017, and a few from 2015. The outlets are:The Atlantic, Breitbart, Business Insider, Buzzfeed News, CNN, Fox News, The Guardian, National Review, New York Post, New York Times, NPR, Reuters, Talking Points Memo, Vox, and The Washington Post.

Taking a look at the data presented, I wanted to take a look at how these publications viewed the candidates from a data science point of view. I used the nrc sentiment tool available in R to attach overall emotions to articles. What it does is labels each individual word with the emotion most commonly associated with it.

In addition to this, I looked into n-gram association with key words across different publications and mediums. One particular area that is of interest is comparing predominantly print media to predominantly online media, and if there are any differences in associations.

My overall hypothesis is that sentiments for Trump will be rather negative, particuarly within online publications, because sometimes it appears that they are not held to the same journalistic standards as print publications, and therefore have more leniency to be biased.

Organizing The Data

The data from Kaggle was split into three different csv's, so in order to get them into one dataset, we can just do a simple rbind.

From there, we can either use 'allarticles' for an overall analysis, or we can look at individual publications by filtering them as such:

You will notice that the last two variables created are 'print' and 'web', even though all of our articles are from websites. I decided to split these publications into what they were predominantly known for, i.e. NYT being a newspaper, and Breitbart being a website. The two outliers would be FOX and CNN, as they are cable news networks primarily, so I listed them under 'web', as they do not produce a physical copy of their reports like those in 'print' do.

There are, of course, downfalls to my methodology presented here. One was just listed above. Another one is the 'nrc' sentiment tool. It can only detect sentiments at face value, and not with context, so sarcasm, jokes, or different scenarios may disrupt the data. In addition, it is largely web based, and may not be 100% adept for analysing journalism in terms of sentiment. Another thing to note is that there may be lots of cross contamination between articles of certain candidates, as some may mention both. In this cae, it could potentially raise or lower the data for a candidate, even though the article is primarily not about them. However I still believe that this data is important to look at, you just have to recognize these pit falls and take them for what they are.

Across All Articles

Let's first look at the overall sentiment of articles that contain the keyword, "Trump".

Let's take this from top to bottom. I first filtered all the articles to detect if they had the word, 'Trump' in them through a str_detect.From there we find the nrc sentiments of all the articles, and then do a cbind of the overall sentiments of the articles, and the articles themselves.

Next, we total all of the suentiments together, and then put them in a dataset. The final code is a ggplot of our dataset, which can be found below. I simply repeated this process for all of the sentiment analysis graphs, but just used different initial datasets, i.e. CNN articles vs all articles, and also included 'Clinton' as a string to search for.

As you can see, the sentiment for both is ultimately the same, but with Trump just having more sentiments because he is mentioned more. However, the ratios of positve and negative sentiments are equal between both.

Using the same method, we can also look at articles that contain the keyword, "Clinton", in them.

CNN vs FOX

Another interesting comparison is the sentiment of both of those candidates across the two of the most popular cable networks. Typically, CNN is regarded as left-leaning and FOX is considered right leaning, which makes this an interesting juxtaposition. Now, the downfall with this is that they are primarily cable outlets, meaning it would be more accurate to compare TV transcripts. But, we only have their online articles, so that will have to do.

Here, it largely remains the same as well. However, for Clinton, there is slightly more sadness , and for Trump, there is slightly more anticipation.

Ultimately, we are having the same conclusions for CNN, each seem to have the same sentiments across their articles. Both FOX and CNN appear to have the same sentiments for each person.

Web vs. Print

Once again, for the Web articles, both candidates appear to have the same sentimetns across the board.

This is again reflected in the print articles: there seems to be no discrepency on sentiment levels between Trump and Hillary. Even looking at print vs web trump graphs, they seem to remain largely consistent.

By these analyses, we can conclude that the sentiments associated with each candiate in articles is ultimately even, and that outlets showed no favoritism based off of negative or positive connotations.

N-Grams

I also thought it would be fun to compare the same publicationss above, but with n-grams, or most common word pairings. For every n-gram, there will be two variables executed: one with most common words before the keyword, and one with the most common words after the keyword.

Ok, let's break that down. Essentially, we are taking our dataframe, breaking it down line by line,and telling it to find the most common word pairings with the keyword.

Some of these pairings will be self-evident, such as a first name, 'administration', or 'campaign', but it is still fun to look at particular differences between mediums.

Before Trump

After Trump

Before Clinton

After Clinton

Here we can see that family members are often brought into the political fray. One particular thing to note is that it seems outlets often quoted Trump's twitter, as they have his handle in the top 10 after category. For Clinton, email's were of particular importance, as was 'aide' and 'endoresed'.

Fox vs. CNN

Fox: Before Trump

Fox: After Trump

Fox: Before Clinton

Fox: After Clinton

CNN: Before Trump

CNN: After Trump

CNN: Before Clinton

CNN: After Clinton

Something I spotted here is that CNN seemd to focus on Trump's subsidiaries, such as Trump Towers or his Organization, whereas FOX did not mention these things in their top n-grams. However, FOX did mention the Clinton Foundation, where CNN did not mention that in their top N-grams. This could potentially point to inherant biases each network holds.

Another thing is that CNN has the word 'accused' in 'Before Hillary', meaning they could have potentially been writing about all the accusations Hillary had. Whether they were defending or attacking on this issue is unclear.

Web vs. Print

Web: Before Trump

Web: After Trump

Web: Before Clinton

Web: After Clinton

Print: Before Trump

Print: After Trump

Print: Before Clinton

Print: After Clinton

These findings are similar to the previous categories before. One interesting thing to note is in the Web Before Trump list, number 8 is 'stop', meaning there is a bit of opposition to Trump's Presidency among web publications. Another interesting thing is in Web Clinton After, the mention of the word, 'Cash' suggesting that there may have been a scandal or something of that sort discussed among these websites.

Print is largely the same story, however, they do mention Trump University, the first time the phrase has come up. This may be due to the more investigative journalism that print publications take part in moreso than online publicaitons. Another key phrase found in Clinton's groupings is the term, 'team', suggesting that there may have been a more positive outlook on this candidate from the Print publicaitons.

Conclusion