This repository implements a python program to generate conjoint image-based tweets programatically and at scale.
The purpose of this tool is to facilitate the work for researchers interested in running survey experiments using images as conjoints, instead of text-based implementations.
This is a work in progress and part of my ongoing collaboration with Kevin Munger (Penn State), Katherine McCabe (Rutger University) and Keng-Chi Chang (UCSD) -- who wrote a big part of this code.
In addition to this readme, we prepared two demos explaining how to use our functions. See below for R and Python users:
# import packages
from matplotlib import font_manager
from PIL import Image, ImageDraw, ImageFont
from textwrap import wrap
import os
import re
import datetime
from numpy import asarrayWe wrote a python function that allows researcher to create tweets and quote tweets given a set of inputs. The function has the following parameters:
CreateTweet(): Create tweet using parameters.- Parameters:
author_avatar(str): avatar of authorauthor_name(str): name of authorauthor_tag(str): twitter username/handle of authortext(str): main text of tweetreactions_retweet(str): number of reactions of tweetreactions_quote(str): number of quotes of tweetreactions_like(str): number of likes of tweettime(str/NULL/None): time of tweet in format “2022-07-05 14:34”; if None use current timequote(TRUE/FALSE): whether or not to print quoted tweetquote_author_avatar(str): avatar of author of quoted tweetquote_author_name(str): name of author of quoted tweetquote_author_tag(str): twitter username/handle of author of quoted tweetquote_text(str): text of quoted tweetreply(True/False): whether or not to print quoted tweetreply_author_avatar(str): avatar of author of quoted tweetreply_author_name(str): name of author of quoted tweetreply_author_tag(str): twitter username/handle of author of quoted tweetreply_text(str): text of quoted tweet
- Returns:
- image: Twitter image in PIL Image format
You can get access to the function on this repository. The logic of the function is to split the tweet in many parts, and combine their positions at the end. These components can all be rotated in a conjoint experiment.
A simple example using the CreateTweet.
from conjoint_tweets import *
tweet = CreateTweet()
SaveTweet(tweet, "output/tweet.png", quality = 95)
To generate a quote tweet, you just need to add quote=True
qt = CreateTweet(quote=True)
SaveTweet(qt, "output/tweet_quote.png", quality = 95)
To generate a tweet with reply, you just need to add reply=True
qt = CreateTweet(quote=True, reply=True)
SaveTweet(qt, "output/tweet_quote_reply.png", quality = 95)
image = CreateTweet(
author_avatar="input/avatar/obama_avatar.jpeg",
author_name="Mark Dem",
author_tag="@_dem2022",
text="You’ve only got six days left to vote in the midterms – and while I know you’ve heard this a million times already, this election really is too important to sit out. In many places, your vote could make the difference.",
reactions_retweet="100",
reactions_quote="200",
reactions_like="20K",
time="2022-07-05 14:34",
quote=TRUE,
quote_author_avatar="input/avatar/woman3_clean.png",
quote_author_name="Rep. Jackson",
quote_author_tag="@_Jack",
quote_text="Don't forget to vote. Your participation matters!",
reply=TRUE,
reply_author_avatar="input/avatar/republican.png",
reply_author_name="FrozenPie",
reply_author_tag="@FrozenPie",
reply_text="You are done! Bye, bye Dem! It is time to bring Trump back")
The final step is to generate multiple images for a conjoint setup. This can be easily done with an any type of functional programming technique. We show here how you can generate multiple conjoint tweets with a simple nested loop.
Notice we use the enumerate function so that we can easily recover all the variations of the conjoint directly from the file names.
# import pandas
import pandas as pd
# create a tweet varying author name, text, and number of reactions to the tweet.
for a, author in enumerate(author_name):
for b, text in enumerate(texts):
for c, retweet in enumerate(reactions_retweet):
for d, quote in enumerate(reactions_quote):
for e, like in enumerate(reactions_like):
# generate image
tweet = CreateTweet(
author_avatar="input/avatar/woman_clean.png",
author_name=author[0],
author_tag=author[1],
text=text,
reactions_retweet=retweet,
reactions_quote=quote,
reactions_like=like,
time="2022-07-05 14:34")
# to save
name = ["author_" + str(a+1),
"text_" + str(b+1),
"retweet_"+ str(c+1),
"quote_"+ str(d+1),
"like_" + str(e+1)]
output= '_'.join(name) + ".png"
# save
outdir= "Output/"
SaveTweet(tweet, outdir+output, quality=95)The next step is to embed all the conjoint images in a survey platform, and perform the randomization. We propose a easy solution for this task. Using the R package plumber we develop a simple API to randomize accross all the imges saved on the folder output. Then, we add the API endpoint as an embedded to your survey platform. In our case, we use Qualtrics, and use the webservice function in the survey workflow to send requests to the API.
All the materials for the API are available in the folder api_conjoint_tweets. The api has two simple files:
-
run_api.r: this file is responsible for deploying the api -
plumber.r: this file contains the function to generate the api requests.
We hosted the API at a AWS RStudio server. However, the images and the api can be hosted in any type of server.
Here you can find a nice tutorial writen by Martin Lukac on how to deploy a pumbler API with AWS EC2
If you don't want go through all these steps, fell free to manually upload the images to qualtrics, and add the randomization directly to your survey.
The code is inspired in two main related works.
-
This medium post here describing how to generate tweets’ screenshots in Python.
-
And Alessandro Vecchiato implementation of visual conjoints for Twitter profiles.
This is a work in progress. Suggestions are more than welcome!