The Words of the Presidents

Research Process

I started with an extremely broad research question, "what sort of interesting patterns can be found by examining the changes in the text over time?" Understanding the type and nature of these patterns would be an iterative process, with initial findings leading to more detailed areas of inquiry. The most obvious data point for a corpus of the size and nature of PPP was word frequency, so that is what I focused on. After sampling some of the available tools for text analysis, I settled on Voyant Tools; of the various free tools available for text analysis, it is the most scholarly in its range of functions. It allowed for word frequency counting of both single words and phrases, and was quite fast given the size of my corpus.

I used the default stopword list provided by Voyant, with two additions - the words "president" and "administration", which were the first and second most common words in the corpus. There were a few other words that I considered excluding due to frequency, like "united" and "states", but I thought both of those words might show up in phrases that I would be interested in.

Once I had my corpus set up in Voyant, I started playing around with the various tools and visualizations it has.

The default set of tools on the Voyant homepage.

I started with the "Trends" tool, which plots the frequency of words within each segment of the corpus (in my case, a segment being a half-year.) I spent some time just intuitively guessing at words and phrases that might seem interesting, comparing their relative popularity over time. I did come up with a few interesting discoveries using this method; for instance, looking at the countries of Central America:

We see a noticeable drop in usage starting in 1990, with occasional small peaks for individual countries. Despite their geographical proximity, we can see these countries are talked about by the POTUS and White House only when a specific policy or event brings then into the news cycle.

(Note: the asterisk after the name of the countries is a wildcard character that stand for "zero or more occurrences of any letter". So "Hondura*" will match either Honduras or Honduran. I used generalized matching like this whenever possible, as Voyant has no features for lemmatization.)

For the most part, though, just guessing at words that might be interesting proved to be an inefficient process. My methodology was greatly improved by my discovery of the "Terms" tool:

This lists all words found in the corpus by frequency, and the sparkline graph in the Trend column plots the frequency for each segment. This makes it very easy to visually identify words that have a high "peakedness", meaning that their frequency varies significantly between segments. (The documentation for this tool in Voyant says that the peakedness value is actually calculated and displayed in the details for each word, but the documentation appears to be out of date or inaccurate, which is a shame, because being able to quantify this value would have been quite helpful.) This became my main research methodology - examining the Terms tool for words that were outliers in terms of peakedness, cataloging those words, and drilling down on them to identify any patterns.

Page updated

Google Sites

Report abuse