Investigating Sociological & Marginalised Themes in Jokes

Proposal

This project aims to study patterns between sociological categories (gender, race, ability, etc) and humor in jokes collected from three websites (primarily Reddit) around 2017. We are specifically interested in identifying patterns around which marginalised groups and themes are represented in the jokes, as well as how those themes relate to the ratings users assign to each joke.
Author
Affiliation

Howell-Xia - (Ahn) Michael Howell, Cat Xia

College of Information Science, University of Arizona

High-Level Goal

To study patterns between sociological categories (gender, race, ability, etc) and humor in jokes collected from three websites (primarily Reddit) around 2017, both in terms of distribution of the sociological themes as well as how they relate to user-assigned ratings. Our end project goal is to create a model that predicts the user rating of a joke based on the sociological categories it falls under and the websites it comes from.

Libraries & packages

import numpy as np
import pandas as pd

Dataset

reddit_url = "https://raw.githubusercontent.com/taivop/joke-dataset/refs/heads/master/reddit_jokes.json"
stupid_stuff_url = "https://raw.githubusercontent.com/taivop/joke-dataset/refs/heads/master/stupidstuff.json"
wocka_url = "https://raw.githubusercontent.com/taivop/joke-dataset/refs/heads/master/wocka.json"

# TODO: update to use CSVs from data folder
# All have columns: id, body
reddit_df = pd.read_json(reddit_url) # Extra columns: title, score
ss_df = pd.read_json(stupid_stuff_url) # Extra columns: rating, category
wocka_df = pd.read_json(wocka_url) # Extra columns: title, category

About our dataset:

  • Total observations? > The dataset includes around 208,000 jokes, collected by Taivo Pungas from three sites: Reddit (~93.7%), StupidStuff (1.8%), and Wocka (4.8%)

  • What time period? > Jokes were collected around early 2017

  • Included columns? > All jokes include id and body

    • Reddit jokes also include: score, title
    • Stupid Stuff jokes also include: rating, category
    • Wocka jokes also include: title, category

  • Citation: @misc{pungas, title={A dataset of English plaintext jokes.}, url={https://github.com/taivop/joke-dataset}, author={Pungas, Taivo}, year={2017}, publisher = {GitHub}, journal = {GitHub repository}}

Note: depending on what we find next, we may go forward with only the Reddit dataset, which is already the bulk of the jokes.

Why we chose this dataset:

  • It includes text data
    • We wanted to work with language-focused and social data, to be able to provide some sociological insights
  • It provides an interesting perspective for analysing social questions
    • Humor (what a culture finds funny) often reflects important insights about what that culture sees as challenging, taboo, and uncomfortable. Marginalised themes - such as someone’s weight or ability - are often at the center of what’s being mocked or reflected on.
  • It is complete
    • This dataset provides enough observations (rows) to be able to surface interesting trends, and the majority of those rows have complete information for their columns.
  • It is available
    • Many of the other datasets we considered and found compelling are no longer available or have a high number of data protections making accessing them difficult, due to the sensitive themes of the data.

Questions

The two central questions we’re looking to answer are:

  1. What are the differences in distributions among social categories, marginalised groups, or themes represented in the jokes?
  • Can we see any patterns around what types of jokes are most prominent on each website?
  • Does the distribution suggest anything about the demographics on each site?
  1. How does each social category/theme relate to the scores and rating - are certain themes rated more highly?
  • What are the distributions of scores per theme?
  • Can we predict a joke’s rating based on its length and theme?
  • Is there major influence from certain users who are more active than others?

Analysis plan

We’ll need to do some preprocessing on the data:

  • Convert JSON to CSV to add to a Google Sheet
  • Consider characters/symbols to remove, lowercasing, stop word removal, and other normalisation of the joke text
  • Create engineered features as new columns (using strict formulas, to start)
    • Find or create a stop word list - use to remove stop words from whole_text_normalised and whole_text_normalised_tokens
    • Come up with word lists for each theme; count how many times word is in the text (we may need to use a smarter approach than simply checking if a word is present, such as using an LLM to annotate themes)

We’ll likely need the following variables, engineered features, and additional data:

  • Some basic info:
    • website (string) - which site the data belongs to
  • Counts/booleans for different social themes: gender, race, etc. - and word lists for each
    • theme_gender - women, men, nonbinary, agender, trans
    • theme_race (edit: ethnicity) - Black, white, Asian, racist
      • Updated to include race-adjacent themes like nationality and language
    • theme_sexuality - gay, homo, straight, sex, sexy, desire, arousal
    • theme_ability - weak, stupid, dumb, slow, disability, spelling, learning
    • theme_age - children, elders, old, young
    • theme_weight - fat, skinny, overweight
    • theme_appearance - ugly, blond, pimple, brunette, fat
    • theme_class - poverty, poor, rich, wealthy
  • whole_text (string) - the first and second parts of the jokes combined
  • whole_text_normalised (string) - to support further NLP analysis
  • whole_text_normalised_tokens (list) - to support further NLP analysis
  • overall_length (int) - count of the words in the text (not normalised, stop words remain)
  • distinct_words (list) - the list of different words in the text (stop words removed)
  • distinct_words_count (int) - the number of different words used

How these will help us answer our questions:

  • Question 1: distributions of themes and representation
    • The social theme counts or booleans will allow us to see distributions per theme, which topics jokes most often reference
  • Question 2: relating themes to user ratings
    • We will be able to see and plot interesting stats (mean, median, standard deviation, etc) of the scores users typically assign per theme
  • Additional insights
    • We can also surface other relevant metrics and patterns using NLP strategies, such as observing which words are most commonly used in jokes, per theme, whether longer or shorter jokes are rated more highly, etc

Model Creation

  • We will create graphs with Seaborn to address the analysis in question 1 and 2.
  • With rating score and word lists, we will create a model that uses classification to predict the rating score of any particular joke from 2017. We can use this model to further understand the underlying social themes around what becomes popular as jokes on Reddit.

Repo organisation:

  • root: contains central files such as README.md, proposal.qmd, presentation.qmd, preprocess_data.ipynb, etc
  • /data: contains the preprocessed dataset (CSV) and any other important data
  • /images: contains plots and other graphics of key interest

Weekly Plan of Attack:

Task Dates Team Member
Project Proposal Aug 8  Both
Data Collection Aug 10  Michael
Data Cleaning and Wrangling Aug 13  Both
Model Training and Evaluation Aug 15  Cat
Initial Powerpoint outline Aug 16  Both
Graph Making and Powerpoint Finalization Aug 18  Both
Powerpoint Recording Aug 19  Both
Last minute cleaning and Submission Aug 20  Both