Discussion Manipulation, Language and Domain Dependent Models: An Overview

The amount of antisocial online behavior (AOB) in the political field has been on the rise in recent years. In the research, we are dealing with the reason for confusing AOB with other forms of antisocial behavior. We specifically dealt with NLP in the political field in the Slovak language, with a specific number of prefixes and suffixes. In this paper, we focus on the area of detection of manipulation of political discussions. The data source consisted of previously collected data in the period from April 2018 with an update until November 19, 2022 realized within this research from the site demagog.sk, which we supplemented with new claims. We trained several classification models on preprocessed data and evaluated them. For multi-class classification, the best results were achieved using logistic regression and a support vector machine trained on the resampled dataset—both achieving an accuracy of 0.56 and a macro F1 score of 0.39. In the case of binary classification, the best results were achieved by logistic regression—accuracy 0.7 and macro F1 score 0.56. These models could help detect manipulation in online political discussions.