RiskTrack Ontology

for On-line Radicalization Domain 

 

Authors: M. Barhamgi, N. Faci, and A. Masmoudi

February 28, 2018

 

 

1. INTRODUCTION 

 

In this page, we present the core of RiskTrack domain ontology. Our ontology defines the radicalization indicators along with a wealth of information that are required for indicator computation. It also defines and incorporates important information about existing terrorist organizations and groups. Our ontology is extensible, i.e., domain experts can extend our ontology with new indicators and domain knowledge. To define our ontology, we have adopted the two-step methodology proposed in [1]. The first step consists in collaboratively designing and organizing the domain ontology by groups of experts. In this step, experts have reused, whenever possible, relevant concepts from existing ontologies from the terrorism domain and completed them with the ones that are relevant for the radicalization domain. The second step consists in integrating the resulting domain ontology in a semantic web tool. The tool built can be used for the ontology maintenance and can be further exploited for the visualization and querying of radicalized users’ data (e.g., personal information, detected radical messages). The first step was carried out in top-down fashion. That is, we identified and defined the key concepts (i.e., the most abstract ones), then we refined them into more concrete concepts that represent specific knowledge or entities. Then, we associated the relevant keywords to each one of the concepts defined.

 

 

2. PRESENTATION OF RISKTRACK ONTOLOGY 

 

We present in Figure-1 shows some of the key concepts of our ontology. For clarity purposes, the figure does not show the properties of defined concepts. The ontology is composed of three main parts: (i) user profile and messages (PART 1), (ii) the radicalization indicators (PART 2), (iii) radicalization and terrorism related concepts (PART 3). We subsequently detail all of these parts.

Figure-1: Overview of RiskTrack Ontology

 

 

PART 1- User profile and messages 

 

This part represents the users of social networks and their personal information. The main concepts are the following:

PART 2- Radicalization indicators 

 

Our ontology defines a set of concepts to represent the radicalization indicators. All indicators are derived of the concept “Radicalization-Indicator” which is directly related to the “Message” concept through the property “isAssociated” (Figure-2).

 

 

Figure-2: Considered radicalization indicators

We considered eight radicalization indicators that we have defined in our previous work [1, 2, 3] and in similar research work [4]. They can be categorized in two types: content-related and style-related. The former refers to the message content regardless of the user. It includes the indicators: ‘perception of discrimination for being Muslim’, ‘expressing negative ideas about the western society’, and ‘expressing positive ideas about Jihadism’, “fixation”, “identification” and “leakage”. The second type refers to the writing style that characterizes each user. It includes two indicators: ‘the individual is frustrated’ and ‘the individual is introverted’. The indicators are defined in Table-1.

Indicator

Description

Frustration [1, 2, 3]

Expressing the fact that the individual does not know what to do with his life. Frustration refers to Irritability, predominantly negative reactions, and anxiety.

Fixation [4, 5, 6]

Fixation is defined as any behavior which indicates an increasing pathological preoccupation with a person or a cause [14]. It can be computed by simply counting the relative frequency of keywords relating to the entities of interest such as persons (e.g., Hitler, Jews), organizations (e.g., al-Qaeda), causes (e.g., Palestine-Israel conflict), etc.

Identification [4, 5, 6]

It represents the identification of an individual with a group (e.g., terrorist group) or cause. It can be expressed through the usage of positive adjectives in connection with mentioning the group or cause.

Fixation [4, 5, 6]

Fixation is defined as any behavior which indicates an increasing pathological preoccupation with a person or a cause [14]. It can be computed by simply counting the relative frequency of keywords relating to the entities of interest such as persons (e.g., Hitler, Jews), organizations (e.g., al-Qaeda), causes (e.g., Palestine-Israel conflict), etc.

Introversion [1, 2, 3]

Expressing the fact that the individual does not like being the center of attention (e.g., when the individual has only few interventions in group conversations, uses short sentences or of moderate length and has few interpersonal relationships ).

Perception of discrimination for being Muslim [1, 2, 3]

The individual expresses a sensation of being discriminated because of his or her religion.

Negative ideas about the Western society [1, 2, 3]

The individual expresses negative opinions about western countries and blames them for crises and wars in the Muslim world.

Positive ideas about Jihadism [1, 2, 3]

The individual expresses positive ideas and opinions about Jihadism and terrorist acts and organizations.

Leakage [4, 5, 6]

Leakage can be defined as the communication of intent to do harm to a third party. Leakage usually signals research, planning and/or implementation of an attack [6].

Table-1: Radicalization Indicators

Each indicator is associated with a set of keywords that have been defined by the experts of Risk-Track (i.e., by linguistic experts, sociologists and criminologists). In Risk-Track the values of indicators could be computed in several ways as follows:

  • Based on their keywords. : The value of the indicator is computed by counting the frequency of their associated keywords then normalizing the counts (this is applicable to most of indicators.);
  • Based on the writing style : For example, the indicator ‘the individual is introverted’ that refers to the introversion personality trait, does not have keywords and is computed based on the length of sentences and the use of ellipses in messages.
  • By applying semantic inference. : Messages could be associated with some indicators when a set of conditions are met. This association and the satisfaction of conditions could be modelled as inference rules. In this project we use SWRL (Semantic Web Rule Language) inference rules.

Before going further, it is worth here to distinguish between an indicator as a concept and its quantified value. For example, the following message “Support and love for IslamicState from Kashmir” reflects a positive stance towards the “Islamic State” which is one of the Jihadist terrorist organization, and therefore it relates directly to the indicator “Positive Ideas about Jihadism” and should be annotated with that concept. However, the value of an indicator represents the likelihood of a given individual exhibiting an indicator. In a perfect world, the value of the “Positive Ideas about Jihadism” indicator for the individual who posted the message “Support and love for IslamicState from Kashmir” would be “1”. However, given the imperfection of existing data mining algorithms, that value would be less than “1”, and it should be computed using the whole set of messages posted by the individual, not only “Support and love for IslamicState from Kashmir”. In the following, we focus on when a message should be annotated with an indicator. The annotation of a message with an indicator means that the message should be included when computing the indicator value.

 

Inferring indicators by semantic inference

  • Inferring the indicator “Positive Ideas About Jihadism” : In order to annotate a message with the “Negative Ideas about the western society” indicator, two disjoint conditions must be satisfied. The first condition is that the message content should express a negative intention that can be demonstrated by negatives actions or emotions. The second condition is the presence of some keywords related to western people or society (for example: West, US, USA, Europe, etc.). “I detest western society” is an example of a message that satisfies the specified conditions. In other words, the message should be already associated (i.e., annotated) with the concepts “Negativity” and “West”. To ensure the satisfaction of those two conditions (for annotating a message with the concept “Negative idea about the West”), we model them using an inference rule written in the Semantic Web Rule Language (SWRL) as follows.

Rule-1 states that if a message (represented by the variable “?m”) is associated with a concept representing a negativity (by referring to one of the negativity concepts or their associated keywords in its text content) and mentions one of the concepts representing the West, then it should be considered as expressing a negative idea about the West, i.e., associated with the indicator “Negative Ideas about the western society”.

  • Inferring the indicator “Negative ideas About the West” : To associate (i.e., annotate) a message with the “Positive Ideas About Jihadism” indicator, the message content should contain positive verbs, adjectives or adverbs that indicate user’s support to an idea or to an entity. The message content must also refer to keywords or concepts related to the “Jihadism” concept. “Jihadism” relates to several concepts in the ontology such as, “Terrorist Organisation”, “Weapon”, “Tactic”, “Terrorist Personality”, etc. To ensure the satisfaction of the aforementioned two conditions (for annotating a message with the concept “Positive Ideas about Jihadism”), we model them using an inference rule written in SWRL as follows:

Rule-2 states that if a message (represented by the variable “?m”) is associated with a concept representing a positivity (by referring to one of the positivity concepts or their associated keywords in its text content) and mentions one of the concepts representing or associated with the concept “Jihadism”, then it should be considered as expressing a positive idea about the Jihadism, i.e., associated with the indicator “Positive Ideas about Jihadism”. For example, the message “Support and love for IslamicState from Kashmir” exposes a positive stance towards a terrorist organization and would be annotated with “Positive Ideas About Jihadism”, as the “Islamic State” is a concept (or more accurately an instance of a concept) that relates to Jihadism.

  • Inferring the indicator “Perception of discrimination for being Muslim” : To annotate a message with the “Perception of discrimination for being Muslim” concept, two conditions must be satisfied. The first is that the message should include Islamic keywords (keywords related to the “Islamism” concept) while the second is the message expresses discrimination feelings (by including keywords related to the “Discrimination” concept). To ensure the satisfaction of those two conditions, we model them using an inference rule written in SWRL as follows:

For example, the message “Poor Muslims are oppressed by US governments” would be annotated with this high-level concept.

  • Inferring the indicator ““Frustration and negative Feeling” : A message is annotated with this concept when it includes a negative content expressed through negative feelings (e.g., frustration, hate, guilt, and fault) or negative actions (e.g., kill, explode, etc.). The following SWRL rule models those two conditions.

Note that more semantic inference rules will be defined and included to cover the complete list of radicalization indicators.  

 

PART 3- Radicalization and terrorism related concepts  

 

Our ontology reuses and extends the definition of concepts from several existing terrorism ontologies. Specifically, we reused and extended the following concepts:

  • The “Terrorist Organization” concept: We reused this concept from the Adversary Intent ontology [5] and refined it with several sub-concepts and instances. For example, the following are all instances of the “Religious Terrorist Organization”: “Islamic State of Iraq and Syria”, “Hezbollah”, “Al-Nusra Front”, “Al-Aqsa Foundation”, “Qaeda”, to name just a few. We have defined the concept of “Terrorist Personality” to represent high-profile individuals that have been associated with a terrorist organization (e.g., Osama Ben Laden, Abu Bakr Al Baghdadi, etc.). We have also associated these concepts with relevant keywords and abbreviations.
  • The “Weapon” concept: We reused this concept, along with its different sub classes, from the Terrorism ontology in [7]. We have also associated these concepts with relevant keywords and abbreviations. This part defines also important concepts that can be used to characterize the content of messages such as “Discrimination”, “Tactics” (to represent the tactics used for a terrorist attack), “Negative Feelings”, “Negative Actions”, “Positive Feelings”, etc.

 

 

REFERENCES 

 

[1] R. Lara-Cabrera, A. Gonzalez-Pardo, D. Camacho, “Statistical Analysis of Risk Assessment Factors and Metrics to Evaluate Radicalisation in Twitter”. Future Generation of Computer Systems - FGCS. Online, 4th November 2017. doi.org/10.1016/j.future.2017.10.046

[2] R. Lara-Cabrera, A. Gonzalez-Pardo1, K. Benouaret, N. Faci, D. Benslimane, and D. Camacho, Measuring the Radicalisation Risk in Social Networks. IEEE Access 2017.

[3] A. Masmoudi, M. Barhamgi, N. Faci, D. Benslimane, D. Camacho. “An Ontology-based Approach for Mining Radicalization Indicators from Online Messages”. The 32-nd IEEE International Conference on Advanced Information Networking and Applications - T10-Internet of Things and Social Networking. Crakow, Poland, May 16-18, 2018.

[4] F. Johansson, L. Kaati, and M. Sahlgren, “Detecting linguistic markers of violent extremism in online environments,” Combating Violent Extremism and Radicalization in the Digital Era, pp. 374–390, 2016.

[5] P. Mullen, D. James, J.R. Meloy, M. Pathé, F. Farnham, L. Preston, B. Darnley, B., & Berman, J. (2009). The fixated and the pursuit of public figures. Journal of Forensic Psychiatry and Psychology, 20, 33-47.

[6] K. Cohen, F. Johansson, L. Kaati, & Mork (2014). Detecting linguistic markers for radical violence in social media. Terrorism and Political Violence, 26(1).

[7] F. Baader, D. Calvanese, D.L. McGuinness, D. Nardi, and P.F. Patel-Schneider, The Description Logic Handbook: Theory, Implementation, and Applications.: Cambridge University Press, 2003.