Using GPT-2 To Optimize Twitter Bots for Maximum Virtue Signaling

M.C. M.L. 

Professor, Fellow of Machine Learning  

Department of Soundcloud Rapping 

Cranberry-Lemon Univeristy

In the past two years, Deep Learning and Natural Language Processing have revolutionized the way we communicate, allowing computers to create text that is  indistinguishable from human-composed text.  While this has led to the automation of many tasks, including email composition and even the completion homework assignments, NLP has yet to be applied to the task of Virtue Signaling.  Virtue Signaling in this context is defined as posting to Twitter and other social media, to let my friends know how much better I am than them.  This framework is flexible enough to be used by anyone who is interested in appearing tolerant and socially aware to their friends or coworkers, but does not have time to research today’s Social Justice issues.

Nomenclature 

Virtue Signaling: The conspicuous and disingenuous expression of moral values with the intent to enhance one’s own image. 

Woke: Aware of and actively attentive to important facts and issues (especially issues of racial and social justice) 

Readers unfamiliar with these two key terms could benefit the most from the Virtue Signaling Twitter Bot outlined in this paper.

1. Introduction 

Woke Without Work: Many people without insight into social justice issues wish to be a part of the conversation.  In today’s busy world, however, who has time to do the necessary research, talk to the right people, or intentionally experience oppression first-hand?  Most would argue that there is no benefit to being Woke unless everyone knows you are.  In this way, GPT-2 allows users to gain all the benefits of being Woke, with none of the work.

2. Dataset Preparation 

As with many machine learning projects, the most important and time-consuming task is preparing a dataset. We were able to scrape the social media profiles of more than 10,000 U.S. celebrities, the vast majority of which have a passive-progressivist bent. We then cleaned and filtered the data using the following keywords which are shown in the list below.

  1. Race 
  2. Environment 
  3. Harambe 
  4. Wealth Inequality 
  5. Gender 
  6. Corporate Greed 
  7. LGBTQ 
  8. ($N>2)th wave Feminism 

These buzzwords target a particular demographic and if you would like to better virtue signal to conservative friends, we recommend scrubbing your local AM talk radio station for a few days. Due to recent twitter restrictions we were only able to test our GPT-2 twitter bot for a conservative demographic. Each conservative GPT-2 twitter bot attempt was permanently banned for spreading nazi propaganda, anti-vaxxer propaganda, and Armenian genocide denial.

We still ended up with an unmanageable dataset, with each celebrity composing more than 10,000 tweets including any of the above subjects.  In order to further consolidate the dataset, we only included tweets with a hashtag, as these are often an attempt to draw attention to one’s views.  This left us with around 10 million short tweets to work with, which for our purposes and compute budget, was a manageable size.

3. Model Architecture 

The most important step in solving the problem at hand, or rather, at tweet, was to create a loss function. Our loss function is as follows: 

GPT-2 Twitter bot Loss equation

The reason we chose this loss function should be readily apparent. The derivation of the loss function should be a good exercise for students everywhere

The loss function balances similarity to the training set with the number of likes and retweets of the post. The latter being weighted the most heavily because retweets can lead to likes from the retweeter’s own audience.  The loss function is also adjusted using the number of likes on a post, because everyone knows it is more important for the content to be liked than to be accurate. 

4. Results 

The results are highly accurate to the training set and therefore optimized for fielding. They consisted of phrases so tone-deaf and self-aggrandizing that they are instantly offensive to anyone of a persecuted minority group who reads them, but will most likely be liked and retweeted by other virtue-signaling non-oppressed people. Some notable examples include:

Progressive twitter bot result using GPT-2 for maximum Virtue signalling
Progressive twitter bot result using GPT-2 for maximum Virtue signalling
Progressive twitter bot result using GPT-2 for maximum Virtue signalling

In order to judge our algorithm objectively, we did not use likes or retweets as a measure of effective virtue signaling because those could easily be taken in a variety of contexts.  We specifically judged each tweet by the amount of response tweets containing the “fist-bump” emoji, as this is often used by tweeters to indicate solidarity with the poster on a particular social issue.

The following is a graph showing the results of this experiment, arranged by topic:

GPT-2 Virtue Signalling Twitter bot effectiveness result. Fist bumps vs topic

4.1 Low-Level Functions 

In addition to composing text, this software framework will occasionally change the user’s profile picture to a symbol reflecting awareness of Fair Trade (a yellow equals sign), LGBTQ rights (a transparent rainbow overlay) or a photo of a recently deceased Civil Rights martyr.  

5. Conclusions 

In a beta test, five of our new users reported a 10 percent increase in likes, retweets, new followers and use of the fist-bump emoji by followers, indicating that their perceived Wokeness level has increased as a direct result of working with our framework.

In conclusion, our Virtue Signaling Twitter Bot provides an effective solution for users who are interested in appearing well-informed on social justice issues without actually being well-informed or doing anything or solving any problems.

If you enjoyed this article please like, share, and subscribe with your email, our twitter handle (@JABDE6), or our facebook group here for weekly content.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

<span>%d</span> bloggers like this: