Online Library

In this paper we present a multimodal approach to categorizing user posts based on their discussion topic. To integrate heterogeneous information extracted from the posts, i.e. text, visual content and the information about user interactions with the online platform, we deploy graph convolutional networks that were recently proven effective in classification tasks on knowledge graphs. As the case study we use the analysis of violent online political extremism content, a challenging task due to a particularly high semantic level at which extremist ideas are discussed. Here we demonstrate the potential of using neural networks on graphs for classifying multimedia content and, perhaps more importantly, the effectiveness of multimedia analysis techniques in aiding the domain experts performing qualitative data analysis. Our conclusions are supported by extensive experiments on a large collection of extremist posts. This research was produced with the aid of VOX-Pol Research Mobility Programme funding and supervision by VOX-Pol colleagues at Dublin City University.

Anashid – Islamic a cappella songs – figure prominently in the propagandaconsumption habits of the 17 jihadists who have caused casualties in the United States since 2013.

VOX-Pol Blog post.

VOX-Pol Blog Post.

Critical assessments of the recent resurgence of right-wing nationalism have rightly highlighted the role of social media in these troubling times, yet they are constrained by an overemphasis on celebrity leaders defined as populists. This article departs from a leader-centric analysis and the liberal frame that still largely informs the assessment of political action, to foreground “fun” as a salient aspect of right-wing mobilization. Building on ethnographic fieldwork among the Hindu nationalists in India, I argue that fun is a meta practice that shapes the interlinked practices of fact-checking, abuse, assembly and aggression among online volunteers for the right-wing movement. Furthermore, fun remains crucial for an experience of absolute autonomy among online users in ideological battles. Providing the daily drip feed for exclusion, fun as a meta practice bears a formal similarity to objectivity in its performative effects of distance and deniability.

This paper reports on a study that is part of a project which aims to develop a multimodal analytical approach for big data analytics, initially in the context of violent extremism. The findings reported here tested the application of natural language processing models to the text of a sample of articles from the online magazines Dabiq and Rumiyah, produced by the Islamic extremist organisation ISIS. For comparison, text of articles found by reverse image search software which re-used the lead images from the original articles in text which either reported on or opposed extremist activities was also analysed. The aim was to explore what insights the natural language processing models could provide to distinguish between texts produced as propaganda to incite violent extremism and texts which either reported on or opposed violent extremism. The results showed that some valuable insights can be gained from such an approach and that these results could be improved through integrating automated analyses with a theoretical approach with analysed language and images in their immediate and social contexts. Such an approach will inform the interpretation of results and will be used in training software so that stronger results can be achieved in the future.