Classification of Propaganda on Fragment Level: Using Logistic Regression with Handcrafted Contextual Features

Abstract

Propaganda in the media has become a rising problem, especially after automation. The ease of which propaganda can be created and spread is astonishing. A way to combat this is an automated propaganda detection system. The goal of fine-grained propaganda detection is to determine whether a given sentence uses a propaganda technique, or to recognize which techniques are used on the fragment level. In this paper we try to analyze the effects of contextual features on the fragment level when training a propaganda classifier. Using a logistic regression model I created some handcrafted features that solely depend on contextual information. The results showed no significant impact on the performance. The features based on the propagandistic fragment itself prove to be the top features in this setting. In future research it is recommended to create either more complex contextual features or to create features that are able to discern whether the fragment is Loaded Language or Name Calling .