Exploring the Morality of Presidential Speech: A Moral Foundations Approach
Advisor: Hal Hansen
Internship Location: University of California, Santa Barbara
Onsite Mentor: Musa Malik
Through the course of this project I will be analyzing presidential debates through a Moral Foundations lens to determine if there exists significant trends between successful presidential candidates. MFT identifies five fundamental moral foundations: care, fairness, loyalty, authority, and sanctity. My final product will be statistically backed results that showcase a trend in one or more of the pillars of Moral Foundations. My initial hypothesis is that successful presidential candidates emphasize the foundation of care. This is based in the fact that successful candidates are better at appealing to the voter base, which can be heavily tied to the pillar of care.
Project Blog
Week 1: Introduction to Moral Foundations Theory
Morality as a subject has implications to a lot of other subjects and can be used to explain real world phenomena. The theory of Moral Foundations (MFT), which divides morality into five pillars: care, fairness, loyalty, sanctity, and authority, allows us to explain the rhetoric of any speaker through morality. Using data mining, natural language processing, and computer vision frameworks, any piece of rhetoric is able to be empirically explained leading to further conclusions about authors and speakers.
Connecting morality to the realm of politics allows social scientists to explore and explain how political officials think. In an election year, I thought it would be insightful to look into presidential rhetoric, especially presidential candidate debates. In the past few election cycles, the popularity of presidential debates has grown exponentially. Especially in the age of mass media, the importance of these debates has also grown significantly with viewers' ideals being subject to change based on how candidates talk about themes and ideas brought up during these debates. Analyzing debates through the lens of the MFT allows trends and patterns to be drawn amongst presidential candidates.
My hope is that I will be able to identify certain trends amongst the pillars of Moral Foundations that can show that successful presidential candidates going back to 1980 emphasize certain common emotions and ideals that lead to their victory. I hypothesize that the pillar of fairness will lead to significant findings. This is mainly due to the fact that successful presidential candidates are usually able to use their fairness ideals to sway voters from the opposite side as well as independent voters to their campaign.
Week 2: Scraping Political Rhetoric from the Web
Welcome to my first weekly update of my project. This past week has been filled with plenty of data mining as well as initial data cleaning for the project. I have come to a general idea of what kind of data I will be using as well as where I will be pulling this data from.
My first point of interest this week was finding a general source from which I would be able to pull presidential speeches from. After some a short amount of surfing as well as conferring with my advisors I was able to find two good sources of presidential speech from which I could pull text from.
The first source is C-SPAN's extensive library of official presidential debate transcripts that can be filtered by year, speaker, text type, and can be searched for specific elements. In addition to transcripts, video links are attached with each transcript. The second source is The Commission on Presidential Debates (CPD) on debates.org which contains another extensive library of presidential speeches dating back to 1960. The CPD contains transcripts for both presidential and vice-presidential debates but lack the filtering options that C-SPAN offers.
Both of these sources work great with my web scraping programs and have been easy to navigate as well. Going forward I will primarily be using the CPD because of its basic format which makes data mining much simpler. However, if I do choose to explore the auditory aspects of debates, I will be going back to C-SPAN's database.
After identifying this general source of data, I then had to scrape the data from each of the webpages as well as make the data suitable for analysis. To do this I use a number of python libraries for browser automation and web scraping. First I had to identify the home pages on which links to each debate transcript are located. After I had this initial link, I had a starting point to then scrape each link. Using Beautiful Soup, a python library that is used to pull data out of XML and HTML files, I was able to build an array of debate links. Using this method I was able to make a list of 56 links or essentially 50 transcripts. After this, I used the Selenium library to pull text from each of the 56 links. Finally, I downloaded these transcripts as .txt files and now I have a working corpus of texts to analyze.
Week 3: Research Into the History of Political Debates
Welcome to my third weekly update of my project. This week, I spent a majority of my time researching more on presidential debates as well as returning to the scoring and dictionary systems that I will be employing throughout my project.
First, looking at the influence of presidential debates, many debates, especially in the digital age have been proven to have mass impact on younger voters. They serve as a dynamic educational tool, allowing young voters to understand key issues and compare candidates directly. Debates also stimulate critical thinking and civic participation, encouraging dialogue among peers and on social media. A debate in particular that was mentioned by a number of articles and papers on the topic is Bill Clinton’s debate with George H.W. Bush. During his debate, Clinton’s personal touch with audience members, especially the youth, made him a promising candidate to voters. In other words the care pillar of his morality resonated deeply with voters and eventually led to his two terms in office.
Furthermore, Clinton’s performance and its aftermath of his eventual election highlight the role of debates in amplifying the personal qualities of candidates, thereby influencing public perception and voter behavior. Debates offer a unique window into the character and values of those seeking office, aspects that are increasingly recognized as pivotal to effective leadership. In the case of Clinton, his debate with Bush played a key role in shaping his public image as a compassionate, relatable, and capable leader, attributes that contributed to his electoral success and his 8 year tenure as President. Clinton played to his strengths well in debates, highlighting his youth in comparison to the almost 70 year old George H.W. Bush.
A majority of the research I found on presidential debates highlights the quality that powerful speakers resonate with the feelings of voters. This has also led me to further believe that the pillar of “care” will provide significant trends when it comes to successful presidential candidates.
Beyond research, I spent time looking at dictionaries previously used for Moral Foundations to get a sense of what dictionary approach I will be using. I also revisited documentation for code regarding the scoring system I will employ as well as playing around with scoring parameters on previous data sets to see how the scores for each of the five Moral Foundations fluctuate.
Week 4: Refining My Scoring System and Unexpected Problems
Welcome to my fourth weekly update of my project. This week was mainly filled with finishing up the refinements to the code I made to the last scoring system last week as well as addressing a problem I found in my dataset this week.
First in fixing the scoring system, based on previous research into the morality of presidential speech, I found that care and fairness would both be pillars of morality that would be important to look into. This stemmed from my reading from last week and how multiple debates showcase the eventually successful presidential candidate using words from care and fairness pillars of morality. To account for this, I had to refine my scoring system to weigh these pillars of care and fairness more.
Using the scoring section of the emfdScore (Extended Moral Foundations) github repository, I am able to change weights of the overall morality score as well as break down individual pillars as I see fit. After my initial tests of each pillar, I will be able to tell whether or not these changes in weights will be required later on or not.
I also found a problem in my data set that persisted for a majority of the week. I found that some of the divisions within text for each presidential candidate were divided unevenly.
The issue that arose in my initial testing was that each break between debaters had words bleeding into the other debaters text block. This was due to an issue in my scraping algorithm reading the “end index” or last word of a debater's specific text box as a later index number. This issue persisted for a majority of the week but by Saturday I was able to divide the text boxes evenly after finding that the end index of each text box was incrementally increasing causing the divisions to be slightly off.
After some final cleaning of the data set including removing stop words which are words that are insignificant in natural language (“is”, “they”, “are”) using the open source library known as spaCy, I finally cleaned my dataset enough to run using the emfd dictionary.
Week 6: Dictionary Applications and Initial Analysis
Welcome to my latest installment where I will go over the most recent findings that I have as well as what I plan to do going forward. This past week, I have been working on analyzing President Biden and President Trump's debate text from 2020 and looking at some of the differences between the Presidents.
First we have to analyze the data. Over the past few weeks, I have collected and stored both Biden and Trump texts that I will be using in .txt files. These .txt files make it easy to run my scoring program on and are easily modifiable. Once these files are in our Colab notebook which is an environment in which we can run data mining techniques, we must also look at our dictionaries.
Our dictionaries for each of the 5 moral foundations are a key part of our analysis section. To build these dictionaries, I used a number of social science techniques. The first is an expert based dictionary building which is a practice in which individuals confer amongst each other on political rhetoric and see which words are most relevant to each pillar. Last summer, two other students and I conducted similar research into other public officials in Congress and we developed a dictionary with the help of our mentors as well. Another approach is using LLM's (Large Language Models) to develop these dictionaries. These LLM's are especially useful because we can synthesize a number of words that are related to a pillar relatively fast. For my dictionary, I used more of an expert-based approach but also took into account some words from our Large Language Models.
After applying our dictionaries, we have the ability to analyze our presidents. The first pair of Presidential candidates I took were our most recent from 2020, President Biden and President Trump. After importing all the required packages, I ran our dictionary against our text files for President Biden. From our initial numbers, President Trump had a higher authority score of 0.221932 compared to Biden's 0.121917. The other pillars would swing between the two Presidents however this initial data on authority is promising. In week 7 I will be looking at this data more in depth as well as other presidential candidates.