A Novel Conceptual Framework for Measuring Online Extremism

• In Event: Populism Goes Digital

Abstract

Radicalization is a social and psychological process through which individuals experience incremental commitment to or adoption of extremist views and ideologies. Individuals change their belief systems, adopt a radical view, and advocate for action, including the willingness to use violence to achieve a radical societal change. To this end, the Internet has become a central platform through which violent extremists spread and influence others to adopt their views.

Drawing from Berger’s (2018) seminal work, we define extremism as the belief that one’s in-group can never be healthy or successful unless it is engaged in hostile action against an out-group. A commitment to hostile action is the key definitional quality of extremism, although precursor conditions are usually met prior to developing that commitment (Berger, 2018).

We propose a novel framework that conceptualizes the radicalization process as a sequence of components and describes radicalization as a multi-level classification problem that can be detected in unstructured text. Further, we develop a machine-readable “ontology” called “Online Extremism Index and Ontology” (OXY) that contains distilled domain knowledge based on our conceptual framework. OXY will guide and aid in developing context-aware machine learning (ML) models to capture extremist discourse online (Kursuncu et al. 2020).

Concept Building Methodology:
In conceptualizing OXY, we utilize multi-level (i.e., three-level) concept building developed by Goertz (2006; 2020). One empirical advantage of using this methodology is that it explicitly addresses (i.e., theorizes) causal analysis present in concepts (and among their constitutive intensions). Specifically, we identify more nuanced concept intensions using Fuzzy-Set analysis by aggregating them using necessary and/or sufficient conditions (Ragin, 2000). Each conceptualized stage of OXY will include a minimum set of necessary (or sufficient) conditions (Goertz, 2005). Because the Fuzzy-Set approach is not confined to be rigidly labeled as dichotomous or continuous, it will enable researchers to account for multi-dimensionality and causal complexity of OXY’s ontology. Furthermore, the Fuzzy-Set approach will allow us to examine each conceptualized stage of OXY in a holistic fashion (Ragin, 2014).

In this work, we build and utilize OXY based on our conceptual framework and validate its effectiveness for computational classification tasks related to two types of extremism: White Supremacy and Islamist Extremism. Our conceptual framework stipulates that extremist content contains three components (J. M. Berger, 2018). The first two components are preconditions for extremism but are often found in non-extremist content—(1) expression of identification with an in-group and (2) expression of negative views about an out-group. Identification with an in-group is very common, and does not in itself suggest extremist involvement. Negative views toward an out-group are less common, and represent a greater potential for extremism, but may reflect situational conflicts or non-ideological biases. The third component is (3) the belief that an in-group can never be successful unless it is engaged in a hostile action against an out-group. The presence or absence of all three components in a corpus of text will be used as a baseline qualification for extremist orientation. The hostile action can be in explicit forms such as death threats or implicit forms such as harassment, trolling, discrimination, segregation. In this framework of the radicalization process, each of the three components has contextual dimensions based on the type of extremism (e.g., religious or political extremism).


OXY Ontology:
OXY stores and structures such extremism-specific knowledge in a machine-readable format, constituting ground truth and embodying a domain ontology of objects, attributes, and relations in taxonomies, partonomies, and specific content with labeled links. OXY is utilized to infuse domain knowledge in ML models to relate lexical, linguistic, and semantic cues spanning contextual dimensions to classify narratives along the components defined in our conceptual framework. Thus, our approach renders the radicalization process computationally accessible through the components and contextual dimensions. As our findings from prior related research (Kursuncu et al. 2019, 2020) indicated, the infusion of knowledge in OXY within ML models will; (i) reduce false alarm and improve recall without reducing precision, (ii) improve the fairness of the model, (iii) reduce reliance on big data. To the best of our knowledge, OXY is the first ontology for online extremism, and it can be expanded into other social harms and threats. Hence, it will have broader impacts in computational modeling of challenging social and political problems online.

Authors

• Ugur Kursuncu, University of South Carolina

• Dilshod Achilov, University of Massachusetts - Dartmouth

• J.M. Berger, Multifaceted Media Group