Corpus Analysis of the Carletonian (done by Ben L.M.)
We thought that one of the best ways to find sentiment on Carleton presidents was to look into the Carletonian, first published way back in 1877. For this task, I decided to use one per year, as this still gives nearly 150 total, providing a decent amount of data while also not causing the process to take unreasonably long as I had to analyze each one manually. I also downloaded the last one published for each year (typically in mid-May or June). This was partially because for the last year for each president, the Carletonian talks a more extensive amount about the president’s legacy which would be very useful for this analysis.
Nearly every Carletonian is scanned and available on the Carleton archives website for download, sorted for every decade and year. An example Carletonian on the website looks like this:
One can view the full text on the website or download each one as a PDF. I downloaded each one and then loaded them in a program called Voyant Tools, which does corpus analysis of text documents (or also any document with words, such as the PDFs). For each, I began by primarily using the correlations tool to see correlating words with the president at the time of the Carletonian published. (One caveat with president Strong was that since his last name is a common name, I instead inputted “president,” which is a rough fix but would hopefully be more accurate than if I did Strong.) This works by finding words that appear in a similar order in the publication to the president’s name. Voyant also displays positively associated words as green and negatively associated words as red. I recorded the number of positive and negative words with at least 0.5 correlation for each Carletonian in a spreadsheet. I chose 0.5 rather arbitrarily, but it at least meant that the words had reasonable correlation with the president. This way of gathering data would hopefully paint a picture of sentiment by seeing if one had particularly high amounts of positive or negative words.
However, about halfway through the process, which was towards the end of the Cowling presidency in the early 1940s, I realized that maybe this way of gathering sentiment wasn’t the most concrete. For example, a negative or positive word completely dissociated with the president could simply appear near the president’s name and cause a very high correlation.
As a fix, I decided to do a more subjective analysis of each, where I would look at every instance that the president’s name appeared in each edition and see if they were criticized, complimented, etc. I would then record what I figured was the view in the spreadsheet, where 1 was positive, 0 was neutral, and -1 was negative.
Once finally done with all the Carletonians, I split the spreadsheet into two halves, where the first was the positive/negative words (or for Strong-Cowling) and the second half was for subjective analysis (for Gould-Poskanzer, not Byerly as I figured it’s probably too soon to tell as she’s the current president). I also created a new column for the first part, for a ratio between positive-negative words. This simply divides the number of positive words by the number of negative words for each Carletonian. The only issue with this was when there were no negative words which would create a divide by zero error. To fix this, I just did (0.01 + Pos) / (0.01 + Neg) for each. This is a pretty rough fix, but it fixed the error while also not changing the other ratios too much.
I then downloaded each as a .csv file and graphed using Flourish, a very simple but useful graphing tool. I really didn’t need to change much for simple and clean graphs. I basically just went with the default versions after setting the series correctly. Here are the final products:
Looking at the first graph of Strong, Sallmon, and Cowling, the band-aid fix for the divide by zero error is a bit of an issue with scale (such as the 1901 ratio with Strong being 20, while all the rest is under 5), but it is still possible to tell that on average, all three presidents had a typical ratio above 1 for each edition, indicating a net-positive sentiment.
For the second graph of Gould through Poskanzer, the outlook is volatile for some, but very steady for others. For example, just in Nason’s 8-year presidency, he had an even split of all three sentiments, being 2 negative, 3 neutral, and 3 positive. In comparison, Oden had 8 straight years of neutrality. Even with the substantial differences in trends, every single president had an overall neutral or positive sentiment, as every average of sentiments for each was at least 0.
Overall, nearly every president looked to have a positive sentiment when looking at the Carletonians I analyzed. In retrospect, I would have done the latter subjective analysis for all presidents so that there wasn’t an odd split in the middle, but the correlations still give an okay look at the earlier presidents.
Sources
“The Carletonian.” 1877. Collection PB008. https://archive.carleton.edu/Detail/collections/144132.