Intelligently Detecting Information Online-Weaponisation Trends (IDIOT)

Fawzia Zehra Kara-Isitt, Stephen Swift, Allan Tucker

Abstract: This paper discusses a detailed study on existing natural language processing open source and commonly used sentiment analysis toolboxes and looks at how various combinations of those toolboxes’ results can be used to accurately classify a sinister intent in a statement. For example, can the toolboxes’ results for different features, such as Attacks, Toxicity and Aggression be combined together predict an Attacks class with more accuracy than just the Attacks classification alone? Can that combination be used to predict any other intimidating intent within text, and can it also help identify a trajectory of an online threatening trend quicker? The main findings so far conclude that the open sourced and massively used sentiment analysis toolboxes for the English language provided by Python and Java work better for Attacking and Aggressive language, compared to general Toxic language. Also, within this experiment, Support Vector machines, although have the largest overheads and take the longest time, give a more reliable accuracy prediction. Finally, Multi-class aggregates of the toolboxes provide on average a much-improved performance result than just using a single class from a single toolbox.

External IDs:doi:10.1007/978-3-031-23618-1_13