Protecting Your Twitter Feed

Pinpoint blocks violent extremist accounts, so that you never have to interact with them in your Twitter feed to start with. Pinpoint is powered by a machine learning model trained off violent extremist messages and posts. Pinpoint is simple to use, all you need to do is authorise the application by clicking the link below, then you’re officially protected by Pinpoint, simple as that!

If you’re redirected to a blank or loading page after authenticating with Twitter, that’s normal and you can leave the page at any time.  If there is not a green button above, then Pinpoint is temporarily unavailable. Please check back later.

 

 

 

How it works

Every 15 minutes Pinpoint will block a subset of extremist accounts it’s identified – this is due to a rate limiting limitation of the Twitter API. Pinpoint breaks the classification of text (social media posts specifically) down into three main categories, these being:

  • Radical Language: How a post is written, including the actual letters and words used as well as the capital letter and violent word frequency.
  • Psychological Signals: The tone of the message such as anger, sadness, anxiety, power, risk, reward, and other sentiments.
  • Behavioural Features: How a user interacts with others in their post such as mentions and hashtags.

 

Training the machine learning model

Pinpoint was trained from messages and posts on the Parler social media network (that were harvested from the site before it originally went offline after the January insurrection in the US). Parler is self-referred to as a freedom of speech social network and is associated with Donald Trump supporters, conservatives, conspiracy theorists, and far-right extremists. In addition to this, the Stormfront internet forum, a neo-Nazi Internet forum and the internets first major racial hate site, was also used for training the classifier. Find out more here.

 

 

The machine learning model

All machine learning models can be measured based off their accuracy and other measures, when using Behavioural Features alone this model has the below scores. In layman’s terms this boils down to it being good, but not perfect.

  • Accuracy: 0.7368404073424881
  • Recall: 0.6270593997684567
  • Precision: 0.7848464582288358
  • F-Measure: 0.6971362094997648

 

 

Legal Stuff

Please Note: If an account is blocked by Pinpoint, that is not to say that the author(s) of this tooling deem such a user to be violent or extremist, instead it is simply a result of the training the machine learning model has undergone and a result of the analysis on the last 10 tweets of the user in question.