=Paper=
{{Paper
|id=Vol-2182/paper_4
|storemode=property
|title=Contextual
Semantics for Radicalisation Detection on Twitter
|pdfUrl=https://ceur-ws.org/Vol-2182/paper_4.pdf
|volume=Vol-2182
|authors=Miriam Fernandez,Harith
Alani
|dblpUrl=https://dblp.org/rec/conf/semweb/FernandezA18
}}
==Contextual
Semantics for Radicalisation Detection on Twitter==
https://ceur-ws.org/Vol-2182/paper_4.pdf
Contextual Semantics for Radicalisation Detection on
Twitter
Miriam Fernandez and Harith Alani
Knowledge Media Institute, The Open University, United Kingdom
{m.fernandez, h.alani}@open.ac.uk?
Abstract. Much research aims to detect online radical content mainly using
radicalisation glossaries, i.e., by looking for terms and expressions associated with
religion, war, offensive language, etc. However, such crude methods are highly
inaccurate towards content that uses radicalisation terminology to simply report on
current events, to share harmless religious rhetoric, or even to counter extremism.
Language is complex and the context in which particular terms are used should not
be disregarded. In this paper, we propose an approach for building a representation
of the semantic context of the terms that are linked to radicalised rhetoric. We
use this approach to analyse over 114K tweets that contain radicalisation-terms
(around 17K posted by pro-ISIS users, and 97k posted by “general” Twitter users).
We report on how the contextual information differs for the same radicalisation-
terms in the two datasets, which indicate that contextual semantics can help to
better discriminate radical content from content that only uses radical terminology.
The classifiers we built to test this hypothesis outperform those that disregard
contextual information.
Keywords: Radicalisation Detection, Semantics, Feature Engineering, Twitter
1 Introduction
In an increasingly digital world, radical individuals continue to successfully exploit
social media platforms for their recruitment campaigns across the world. Particularly, the
so-called Islamic State of Iraq and the Levant (ISIL/ISIS) is one of the leading terrorist
organisations on the use of social media to share their propaganda, raise funds, and
radicalise and recruit individuals.
While law enforcement agencies, experts, and volunteers are actively working to
target this problem (see for example Ctr-sec1 , where volunteers report the existence
of ISIS propaganda in social media and they claim responsibility for closing more
than 200,000 Twitter accounts), the vast amount of traffic generated in social media
makes it nearly impossible to manually identify radical content and violent extremism
online. Researchers and governments are therefore investing in the creation of advanced
information technologies to identify and counter extremism through intelligent large-
scale analysis of online data.2
However, existing methods to automatically identify radical content online mainly
rely on the use of glossaries (i.e., lists of terms and expressions associated with religion,
?
Copyright held by the author(s)
1
https://twitter.com/CtrlSec
2
http://www.voxpol.eu/identifying-radical-content-online/
war, offensive language, etc.). These methods are not always effective and we continue
to observe that many who use radicalisation terminology in their tweets are simply
reporting current events (e.g., “Islamic State hacks Swedish radio station’, or sharing
harmless religious rhetoric (e.g., “If you want to talk to Allah, pray. If you want Allah
to talk to you, read the Qur’an’’, or even countering extremism (“armed jihad is for
defence of muslim nation. Not for establishment of khilafah.”).
Reliability is indeed a well-known concern in automatic radicalisation detection. A
popular example is the campaign launched by the hacker community Anonymous as a
response to the Paris attacks, where they claimed taking down more than 20,000 Twitter
accounts linked to ISIS. Those accounts included, among others, the ones of the U.S
ex-president Barack Obama, the White House, the BBC news, to name just a few.3 One
important source of inaccuracy of automatic radicalisation detection approaches is their
reliance on the appearance of terms and expressions without considering their context.
In other words, although the content may contain terms that are indeed associated with
common radicalised rhetoric (“Islamic State”, “Allah”, etc.), these words do not convey
radicalisation meanings all the time. Instead, this meaning is only ’active’ in particular
situations or contexts. We hypothesise that, in order to provide effective mechanisms to
identify radicalised content, and to develop intelligent algorithms to identify and counter
extremism, it is vital to not only consider radicalised terminologies and expressions in
an isolated manner, but to take into account the surrounding contextual information.
Considering this, our work investigates the following two research questions:
– Are there significant variances between the semantic contexts of radicalisation
terminology when this terminology is used to convey ’radicalised’ meaning vs. when
it is not?
– Can we improve the effectiveness of radicalisation detection by making it more
context-relevant, thus increasing its accuracy and consistency?
The notion of context is not new and has been long acknowledged as being of high
importance in a wide variety of fields, such as computational linguistics, automatic
image analysis, information retrieval or personalisation, to name a few [16]. The research
presented here focuses on the role of context for radicalisation detection in social media,
and more particularly Twitter. Our approach for context modelling exploits semantic,
ontology-based contextual information from DBpedia and Wikidata,4 as well popular
classification Taxonomies including IPTC media topics, IPTC newscodes and the Internet
Advertising Bureau QAG segments.5 Among the possible knowledge representation
formalisms, ontologies and knowledge bases present a number of advantages as they
provide a formal framework for supporting explicit, machine-processable semantic
definitions, and facilitate inference and derivation of new knowledge based on already
existing ones.
To answer the above research questions, we analysed over 114K tweets mention-
ing terms linked to radicalised rhetoric (17K originated from pro-ISIS users and 97K
3
http://www.bbc.co.uk/newsbeat/article/34919781/anonymous-anti-
islamic-state-list-features-obama-and-bbc-news
4
http://wiki.dbpedia.org/,https://www.wikidata.org
5
http://cv.iptc.org/,http://www.iab.com/guidelines/iab-quality-
assurance-guidelines-qag-taxonomy/
originated from “general” users). We study how contextual semantic information differs
for the same terms in the two datasets. Aiming to use this notion of context to better
discriminate radical content we proposed an approach to incorporate it into existing clas-
sification methods, obtaining a 4.5% improvement in precision when using contextual
information. By conducting this work our paper provides the following contributions:
– A contextual representation for radicalisation detection on Twitter based on ontolo-
gies and knowledge bases.
– Demonstrate that contextual information differs between content originating from
pro-ISIS accounts vs. content originating from general Twitter accounts, even though
both contents use radicalised terminology.
– Show that contextual semantic information can help to enhance existing keyword-
based radicalisation detection mechanisms.
The following sections are structured as follows. Section 2 describes related work.
Section 3 shows our proposed approach to automatically model and capture the semantic
context of terms and expressions. Sections 4 and 5 discuss the evaluation and results of
studying the semantic context divergence across terms and the results of incorporating
contextual information for radicalisation detection. An in-depth discussion of our findings
and the limitations of this study is reported in Section 6. Section 7 concludes.
2 Related Work
Researchers from the areas of counter-terrorism and cyber-security have begun to exam-
ine the radicalisation phenomenon and to analyse the social media presence and actions
of extremist organisations [1] [9]. These works include computational approaches devel-
oped towards the analysis and detection of radicalisation.
Among the works developed towards analysing the online radicalisation phenomenon,
we can highlight the works of Klausen [11], Carter [7], Chatfield [8], Vergani [17] and
Rowe [13]. These works examine the type of language and communication strategies that
ISIS members and supporters use to radicalise and recruit others. The works highlight:
(i) the direction of the communication flow, from the terrorist accounts, to the fighters
based in the insurgents’ zones, to the followers in the west [11], (ii) the frequent use of
certain terms in the conversations (islamic, Allah, fight, Mujahideen, ISIS, etc.) [7], (iii)
the frequent mentioning of international media, regional Arabic media, IS sympathisers
and IS fighters [8], (iv) the use of expressions related to achievement, affiliation and
power with a focus on emotional language, (v) the frequent mentioning of death, female
and religious related terms [6], and (vi) the high relevance of social-homophily on
the diffusion of pro-ISIS terminology. These works are not focused on the automatic
detection of online extremism, but rather on investigating and understanding the online
radicalisation process in different channels, including Twitter [11, 7, 8, 13], Facebook
[7], YouTube [6] and Dabiq; the official ISIS magazine [17].
Among the works concerned with the detection of extremist content and Tweets,
we can highlight the works of Berger [4, 5], Agarwal [2], Ashcroft [3] and Saif [14].
These approaches tend to: (i) rely on the use of known radical users as seeds and expand
through the network of followers to detect other potential radical users and content [4][5]
or, (ii) build classifiers trained on data collected based on radicalisation glossaries, where
users whose tweets contain certain words would be regarded as in the “radicalised” set
[1][3][14]. Sometimes these collected tweets are additionally filtered based on: (i) lists
of known radicalised users [3][14], (ii) based on whether the user account has been
suspended [5] or, (iii) validated by annotators, who, in many cases are non-experts on the
topic [2]. It is also unclear how the manual annotations are performed and what criteria
are followed by the annotators when categorising users and data [5].
The lack of grown truth datasets for radicalisation detection, and the difficulties
for obtaining them (particularly when relying on glossaries of radicalisation terms) are
well known problems in this field (see Section 6). To the best of our knowledge no
previous studies have focused on understanding how the context of radical terminology
semantically differs when used to convey a radical message vs. when not, which is one
of the key contributions of this work.
3 Semantic Context
3.1 Modeling Semantic Context
Context is a difficult notion to grasp and capture in a software system. In this work, we
focus our efforts on the notion of semantic context for radicalised content detection. The
information objects in our model are therefore a set of posts P , and the units for which
we want to extract the context are a set of terms t ∈ T , where T is a glossary/lexicon
used for radicalisation detection (examples of these terms can be seen in Section 4.2).
We define the semantic context of a term t as the set of categories Ctp , topics Ttp ,
entities Etp and concepts (entity types) ETtp that emerge from the posts where the
term appears tp ⊂ P . Multiple semantic services have been built over time to enable
the annotation of text with concepts, properties and values defined according to domain
ontologies and stored in knowledges bases (KBs) [12]. For this work, we selected
TextRazor6 as the service to extract the semantic elements (categories, topics, entities
and types) that we will use to compose the context for each term. For every annotation,
TextRazor provides a confidence score. More details on how annotations are extracted
and how this confidence score is calculated can be found on TextRazor’s website. Our
semantic context model is therefore composed by four dimensions:
1. Categories: The set of Categories C for the posts P is extracted considering three
different taxonomies: IPTC media, IPTC news codes and the Internet Advertising
Bureau (IAB) QAG segments. Examples of relevant IPTC categories include: {id:
16009001 label: unrest, conflicts and war>war>civil war}, {id:20000677 label:
religion and belief>religious belief>Islam}. Examples of IAB taxonomy categories
include: {id: IAB11 label: Law, Govt & Politics}.
2. Topics: High-level topics are assigned to posts P based on Wikipedia. These topics
include their corresponding link to wikidata and wikipedia pages and categories.
Examples of relevant topics include: {wikidataId: Q7283 label:Terrorism} and
{wikidataId: Q42418 label: Taliban}.
3. Entities: The set of Named Entities E for the posts P is extracted based on DBpedia
and Wikidata. Examples of relevant entities include: {wikidataId: Q41183 label:
Aleppo types:[Place, PopulatedPlace, Settlement, City]}, {wikidataId: Q133207
label: Muslim Brotherhood types:[Agent, Organisation, PoliticalParty]} .
6
https://www.textrazor.com
4. Entity Types: Entity types refer to the DBpedia concepts (types) of the previ-
ously extracted entities. These concepts are also linked to wikidata. E.g., the types
for the entity: Muslim Brotherhood include: {dbo:Agent wikidataid: Q24229398},
{dbp:Organisation wikidataid:Q43229} and {dbo:PoliticalParty, wikidataId:Q7278}7
Once we have extracted the different semantic elements for every post p ∈ P we use
these elements to extract the semantic context of every radicalised term t ∈ T . In our
work, we extract two different semantic contexts for each term: SCR(t) and SCN R(t).
The first semantic context SCR(t) refers to the context extracted from a subset of tweets
posted by pro-ISIS users. The second semantic context SCN R(t) refers to the context
extracted from a subset of tweets posted by non pro-ISIS users.
Let Ptr ⊂ P be the set of tweets posted by pro-ISIS users where the term t appears.
SCR(t) is then defined as the set of categories Ctr , topics Ttr , entities Etr , and entity
types ETtr that emerge from the semantic annotations of Ptr . Note that the same
category, topic, entity, or entity type may emerge multiple times in the tweets associated
with the term. We therefore define the vectors of unique categories, topics, entities and
entity types as:
Pp=|Ptr |
p=1 conf score(cip )
V ctr = {c1 , c2 , ....cn }, ci ∈ Ctr and val(ci ) =
|Ptr |
Pp=|Ptr |
p=1 conf score(tip )
V ttr = {t1 , t2 , ....tm }, ti ∈ Ttr and val(ti ) =
|Ptr |
Pp=|Ptr |
p=1 conf score(eip )
V etr = {e1 , e2 , ....eo }, ei ∈ Etr and val(ei ) =
|Ptr |
Pp=|Ptr |
p=1 conf score(etip )
V ettr = {et1 , et2 , ....ets }, eti ∈ ETtr and val(eti ) =
|Ptr |
Note that the value for each contextual element is computed based on the confidence
score of the annotations. For example, if we are computing the context for the term
Allah, and this term appears in five posts, three of them annotated with the category
religious belief, the value of the category religious belief in the context of the term Allah
is computed as the sum of the confidence scores of the three annotations, divided by the
total number of posts in which the term Allah appears, in this case five. By doing this,
the value incorporates both, the frequency of the semantic element (i.e., in how many
annotations it appears) as well as the confidence of those annotations. Note also that
confidence scores values are between [0, 1]. By normalising by the number of posts we
ensure that the value for every element (ci , ti , ei , eti ) in the context of the term is always
between [0,1] as well. We then define the two semantic contexts for each radicalised
term t ∈ T as:
SCR(t) = {V ctr , V ttr , V etr , V ettr }, which is built based on Ptr ⊂ P , the set of
radical tweets, or tweets posted by pro-ISIS users.
SCN R(t) = {V ctnr , V ttnr , V etnr , V ettnr }, which is built based on Ptnr ⊂ P ,
the set of non radical tweets, or tweets posted by non pro-ISIS users.
A reduced example of these contexts for the term Andalus is displayed in Table 1.
7
dbo refers to DBpedia Ontology
Table 1: Semantic contexts for the term andalus. SCR(andalus) vs. SCN R(andalus).
Scores for each element are not added to the image for better visibility
3.2 Calculating Semantic Context Divergence
To assess the divergence of semantic contexts among radicalisation terminology, we
compute for each term t ∈ T the similarity between SCR(t), and SCN R(t). We
maintain the four defined contextual dimensions (categories, topics, entities and entity
types), and use the cosine similarity metric (cos)8 to compute the similarity between the
two semantic contexts for each dimension. The semantic context divergence for each
term is then defined as: SCD(t) = {1 − cos(V ctr , V ctn r ), 1 − cos(V ttr , tV tn r ), 1 −
cos(V etr , V etn r ), 1 − cos(V ettr , V ettn r )}
3.3 Exploiting Semantic Context for Radicalisation Detection
There are multiple ways in which the proposed semantic context model could be incor-
porated into existing radicalisation detection approaches. It could particularly be helpful
for defining rules associated to current radicalisation lexicons, so that posts and users are
not automatically categorised as ’radical’ if they contain or use certain terms. Instead,
the use of those terms could be weighted based on their contextual information, similarly
to existing contextually enhanced lexicon-based sentiment analysis approaches [15].
In this work however, our purpose is not to generate a novel method for radicalisation
detection, neither to improve current state of the art approaches for radicalisation detec-
tion, but to study whether semantic information could help to enhance keyword-based
methods by making them more context relevant. To test this hypothesis we have build
classifiers based on n-grams and test them against classifiers that incorporate as addi-
tional features the extracted contextual information. To train these classifiers we make
use of a dataset of posts P , along with their class labels C =(pro-ISIS, non pro-ISIS).
Features are extracted in the following way. For each posts p ∈ P we create two vectors:
the unigram vector and the semantically enhanced vector.
8
https://en.wikipedia.org/wiki/Cosine_similarity
Vunigramp = (w1 , w2 , ....wn ), where wi is the tf ∗ idf value of the unigram.9
Vsemanticp = (w1 , w2 , ....wn , c1 , c2 , ...cm , t1 , t2 , ...to , e1 e2 , ...ep , et1 , et2 , ....etq ),
where (c1 , c2 , ...cm ), (t1 , t2 , ...to ), (e1 e2 , ...ep ), (et1 , et2 , ....etq ) are the semantic con-
text elements (categories, topics, entities and entity types) extracted for p. To generate
these vectors we consider only semantic elements for which the confidence score of their
annotations is higher than 0.5.
4 Experimental Setup
4.1 Dataset
We use two publicly available datasets to study radicalisation, from Kaggle datascience
community. The first dataset contains 17,350 tweets from 112 distinct pro-ISIS ac-
counts.10 collected based on a three-month period study.11 To ensure that this dataset
contains only users that are pro-ISIS, we checked the profiles of the 112 accounts using
the Twitter API. We assume that if the account is no longer available (i.e., has been
blocked), then most likely the account did indeed belong to a radicalised individual. Only
two of those accounts were still active at the time of writing, one belongs to a journalist,
and the other one to a researcher who focuses on Jihadi groups. We deleted these two
profiles and their tweets from the dataset, thus obtaining a final dataset of 110 users and
16,949 tweets.
The second dataset was created as a counterpoise of the previous dataset. It contains
122K tweets from 95,725 distinct users collected on two separate days 7/4/2016 and
7/11/2016. Tweets were collected based on the following keywords (isis, isil, daesh,
islamicstate, raqqa, Mosul, ’islamic state’).12 Many of these accounts have now been
blocked. To ensure that this dataset contains only users that are not pro-ISIS (they could
be anti-ISIS or neutral), we collected the profiles of the 95,725 users by making use of
the Twitter API. We found that 76,819 of them were still active at the time of writing.
We assume that, if the account is still active even though two years have passed since it
was shared on Kaggle, then the profile does not belong to a radicalised individual. To
assess this hypothesis, a random subset of 40 profiles was selected and manually assess
by two annotators (authors), who agreed, for all the 40 profiles, that these accounts do
not show signs of support for ISIS. The final dataset after removing the content of the
closed accounts contains 97,287 tweets from 79,819 active users.
Note that in the cases of both datasets, tweets contain radicalised terminology, but,
while in the first dataset that terminology was used by radical individuals to spread
their message, in the second dataset, ’common’ Twitter users are the ones using that
terminology to, for example, report or discuss related news and events.
While most tweets in these datasets are written in English, few of them are written in
arabic. TextRazor recognises a variety of languages.13 However, the annotator could not
extract information for 3,249 tweets from the pro-ISIS dataset and 2,573 from the non
9
https://en.wikipedia.org/wiki/Tf-idf
10
https://www.kaggle.com/fifthtribe/how-isis-uses-twitter
11
http://blog.kaggle.com/2016/06/03/dataset-spotlight-how-isis-
uses-twitter/
12
https://www.kaggle.com/activegalaxy/isis-related-tweets
13
https://www.textrazor.com/languages
pro-ISIS dataset. This was due to the use of unknown languages (e.g., arabic script) or
because the tweets contained very few pieces of information (for example, just a URL).
4.2 Radicalisation Terminology
Radicalisation lexicons have been developed by experts, and have also been created from
content generated by ISIS, such as the Dabiq 14 and Inspire15 magazines. In this work
we make use of a lexicon containing 305 entries, including 556 terms, expressions and
variances created in a previous work [10]. An example can be seen in Figure 1
Fig. 1: Examples of entries in the radicalisation glossary
4.3 Semantic Context Extraction
To extract the categories, topics, entities, and entity types that define our semantic
context (see Section 3.1) we have made use of the TextRazor semantic annotator. To
compute/build our context we are only considering annotations for which the confidence
score is higher than 0.5. Categories, topics, entities, and entity types that emerge from
the text but with a confidence score lower than 0.5 are discarded. Context is only built
for radicalised terms that appear in a minimum of five tweets per dataset.
5 Evaluation Results
5.1 Semantic Context Divergence
The first step when conducing our analysis was to distinguish between: (i) the terms
that do not appear in any of the datasets, (ii) the terms that appear in one dataset but not
in the other and, (iii) the terms that appear in both datasets, which will be the ones for
which we will extract and compare contextual information.
– Pro-ISIS: 48 terms, (15.7%) have been found only in the Pro-ISIS dataset. Examples
of the terms appearing with higher frequency include: Sahwat - Awakening Councils,
derogatory term denoting militias/groups who are allegedly supported by the West -
(e.g. “Sahwat spent three months battling IS in #Aleppo with the support of #US
planes to create a ”Safe zone”)” , or taghut - terminology denoting a focus of worship
other than Allah (e.g., “After the recent killings of Sheikh faris Zahrani and other
Sheyukh by the taghut of al saud AQAP has warned this [url]”16 ).
– Non Pro-ISIS: 17 terms (5.6%) have been found only in the Non Pro-ISIS dataset.
Examples among the terms appearing with higher frequency include: Fundamentalist
-a person who believes in the strict, literal interpretation of scripture in a religion-
14
https://en.wikipedia.org/wiki/Dabiq_(magazine)
15
https://en.wikipedia.org/wiki/Inspire_(magazine)
16
Sheikh Faris bin Ahmed Jamaan al-Showeel al-Zahrani was on Saudi Arabia’s list of 26 ’most-
wanted’ suspected terrorists. On 2 January 2016 Sheikh Zahrani was executed by the Saudi
state along with 46 others convicted of terrorism
(e.g., “IF ISIS IS NOT ISLAMIC TERRORIST OR ISLAMIC EXTREMIST OR
ISLAMIC FUNDAMENTALIST OR NOT ISLAMIC AT ALL,WHY SHARIA
LAW NEEDED??”), or “Zindiq -Muslim individuals who are considered to hold
views or follow practices that are contrary to central Islamic dogmas- (e.g., “#ISIS
are takfiri, khawarij, zindiq & the real enemy of #Islam.. They did a frontal attack
on #Islam. [username]”).
– Pro and Non Pro-ISIS: 146 terms (47.9%) appear in both datasets. Common terms
include Allah -God -, Amaq - A news agency linked to ISIS -, Islamic State, Khilafa,
etc. See for example the term Allah used by the pro-ISIS group “Our Prophet, has
ordered us to fight you till you worship Allah Alone or give Jizya” vs. a use by the
non pro-ISIS group “May ALLAH the Almighty destroy ISIS and rid humanity of
their disgusting actions. #Istanbul [username]”.
– Not found: 94 terms (30.8%) do not appear in any of the datasets. Examples of
these terms are: Millah - which refers to any variety of religion, except the “true”
religion of Islam, or Tatarrus - the practice of using civilians as human shields.
We have further analysed the use of the 146 terms appearing in both datasets to better
understand the semantic context divergence when the terms are used by the pro-ISIS vs.
the non pro-ISIS group. As reported in section 3.1 we consider four different dimensions
of semantic context: (i) categories, (ii) topics, (iii) entities and (iv) entity types. Each
of these contextual dimensions is extracted for each of the 146 terms under analysis by
considering the tweets in which the terms appear within the non-pro-ISIS and pro-ISIS
datasets.
Our experiments show that, while there exist a semantic contextual divergence among
terms, this divergence is not equally distinctive across the different contextual dimensions
(see Figure 3). The more discriminative contextual dimension is Entities. Out of the
terms analysed, 95 (65%), displayed contextual divergence > 0.25 based on entities.
This is lower for categories (29 terms, 20%), topics (12 terms, 8.2%) and entity types
(8 terms, 12.33%). Entities are more fine-grained/specific descriptors of the context,
while topics, categories and types provide higher-level conceptualisations. The Venn
diagram showing these intersections is displayed in Figure 2. In total, out of the 146
terms analysed, 100 (68.5%) showed a context divergence >= 0.25 for at least one of
the semantic context dimensions. Those terms that exhibit low context divergence for
all dimensions refer to locations (such as Afghanistan or Iraq), as well as historic and
religious-related terms (e.g., aqdah, khawarij). Interestingly, this group of terms also
includes proper nouns used for the radical groups including ISIL, Taliban, or Al-Qaeda.
These terms appear in a very high number of tweets and are therefore associated with a
high number of topics, categories, entities and entity types in both groups. An indication
that, while they may be useful to detect social media conversations around jihadism, they
may not be that helpful when discriminating radical content among those conversations.
Among the contextual divergences we have also observed that some contextual
elements (either categories, topics, entities or entity types) appear in one context but
not in the other one, and that certain semantic elements, although they appear in both
contexts, they show more ’strength’ in one of them. Note that value for a semantic
element in the context of a term is computed based on the frequency and confidence of
the semantic element in the annotations (see Section 3.1)
Let’s take as an example the term fatwa (a ruling on a point of Islamic law given
by a recognized authority). Among the entities that appear in SCR(f atwa) but not in
SCN R(f atwa) we found: Jihad, Al-Qaeda, Syria, Ayatollah, Prophets and messengers
in Islam, Mujahideen, etc. Among the entities that appear in SCN R(f atwa) but not in
SCR(f atwa) we found: Terrorism, Sacrilege, Kurt Westergaard, and Charlie Hebdo
(who published controversial cartoons of figures and elements of Islam) and were the
target of terrorist acts. Among the entities that appear in both, but stronger in one context
than the other, the entity Muslim is used with more strength in the non pro-ISISl context,
while the entity Sunni Islam appears with higher strength in the pro-ISIS one. All these
semantic variances can therefore help us to better understand when these terms may be
conveying a radical connotation rather than descriptive meanings.
Fig. 2: Venn Diagram showing the terms whose context divergence is > 0.25 for any of
the four semantic context dimensions
5.2 Radicalisation Detection
As mentioned in Section 3.3, while semantic contextual information can be incorporated
in a variety of ways to enhance current radicalisation detection approaches, in this work
we are testing its effectiveness against the use of unigrams for radicalisation classification.
Our dataset for training is composed the 13,700 tweets of the pro-ISIS dataset for which
we could obtain semantic annotations (see Section 5.1) (labelled as pro-ISIS) and a
random sample of 13,700 tweets from the non pro-ISIS dataset (labelled as non pro-ISIS).
Fig. 3: Divergence across the four different context dimensions for those terms showing
a context divergence > 0.25 for at least one of the dimensions.
We tested multiple classifiers including Naive Bayes (NB), SVM and decision trees
(J48) and compared the use of n-gram only features vs. the use of n-grams enriched with
contextual semantics. The top results were obtained with SVM in both cases (see Table
2). Results are reported based on 10-fold cross validation. As we can see in the table,
the semantically enhance classier outperforms the unigrams-only classifier in precision,
recall and F1-measure. This indicates that the use of semantics can indeed help providing
additional meaning to terms in order to enhance radicalisation detection.
Table 2: Keyword-based vs. Semantic-based radicalisation detection
Features Precision Recall F-measure
Uni-grams Only 0.816 0.801 0.822
Semantic Context Enrichment 0.859 0.843 0.851
6 Discussion and Future Work
Accurate detection of radicalised content and users online is a major problem faced
by current governments and policing organisations. This is due to the fact that media
agencies, journalists, political figures, and religious non-radical individuals may be using
the same terminology that radical individuals use to spread their ideology. The problem
is not only technically challenging but also ethically, since the wrong association of an
individual to a radical movement can have serious implications.
In this work, we explored the use of semantic information to better understand the
contextual variances in which radicalisation terms are used when conveying ’radicalised
meaning’ vs. when not. We also looked at whether keyword-only radicalisation detection
could be enhanced by incorporating contextual semantic information. While our work
indeed shows that semantics can provide a more fine-grained understanding of online
radicalisation, this research has various limitations.
The most difficult challenge confronted by anyone who studies online radicalisation
is the lack of gold standard datasets. Existing datasets are rarely verified by experts to
ensure that false positives (i.e., content or user profiles that have been wrongly associated
to radicalisation) are identified. Annotating this data is also not trivial, since annotators
should have a certain degree of background knowledge on religion and politics to
correctly annotate such content. We are currently working with law enforcement agencies
and experts to be able to achieve such gold standards.
As explained in Section 5.1, we checked to see whether a Twitter account was
blocked or not to consider the account to be a pro-ISIS or not. However, Twitter could
wrongly block an account, or block it for reasons other than radicalisation. Therefore, a
better approach is to manually annotate these accounts. To be on the safe side, instead of
including blocked accounts as part of the pro-ISIS dataset we decided to discard them.
Another issue is the imbalanced number of users and tweets in the pro-ISIS and
the non pro-ISIS datasets. While we have balanced the dataset to create our classifiers
(see Section 5.2), we have decided to maintain all tweets to study the semantic context
associated to radicalisation terms. Our assumption here is that, the more tweets we have,
the more complete can be the semantic context we extract. Terms appear in average more
than 100 times in each dataset and, as mention in Section 3 we have discarded those
terms for which we have less than 5 tweets in each dataset to build their context.
The pro-ISIS and the non pro-ISIS datasets were collected in different time periods.
Global information (i.e., news and events around the world) do vary over time, and
leaders and locations that are relevant now may not be in a few months. This could
particularly affect the contextual divergence obtained by looking at entities (persons,
organisations, locations), which is indeed the highest for all terms across the four
considered contextual dimensions. Ideally one should analyse two datasets collected in
the same time period to avoid the potential time influence.
Finally it is also important to highlight that, while the tweets contained in the pro-
ISIS dataset have been written by pro-ISIS users, that does not necessarily mean that all
these tweets have radical content. Similarly the non pro-ISIS dataset may also contain
some tweets of radical nature. Therefore, while our results show that pro-ISIS and
non pro-ISIS users present divergences in their use of radicalisation terminology, we
cannot claim that these terms are used differently in radical vs. non-radical content. The
same remark applies to our classifier. Our classifier shows that semantic information
can help us to better differentiate content coming from pro-ISIS accounts, but we can
not claim that our classifier improves the detection of radical content. Regarding the
creation of this classifier, it is also important to highlight that our aim here was not to
enhance current state of the art approaches for content-based radicalisation detection,
but to investigate whether the use of semantics could help to enhance current methods
by providing additional contextual information.
Regarding our proposed approach for semantic context representation, we consider
that if an entity and a radicalisation term appear in the same tweet they are context-
related. However a more fine-grained definition of context could be considered by taking
into account the specific relation between the term and the entity as expressed in the text.
For simplicity, we have not considered relations in this study. Regarding the computed
contextual values, although the performance of semantic annotators have significantly
improved over the years, annotations are not always accurate, specially when it comes to
social media text (which is generally short and ill-formed). Similarly, Knowledge Bases
may also contain incomplete or non-updated data, or even mistakes. While our model
already takes into account the notion of ’confidence’ within the annotations, this notion
could be expanded by considering the quality (accuracy, completeness) of the KBs used
to generate the annotations.
Despite all the above mentioned limitations, every step towards the effective detection
of radicalised rhetoric may have important societal implications. We hope that the
presented study will serve as basis for future work within and across the Semantic Web,
the Social Web, and the policing research communities.
7 Conclusions
In this paper we propose an approach for building a representation of the semantic context
of the terms that are linked to radicalised rhetoric. We report on how the contextual
information differs for the same radicalisation-terms in two datasets generated by pro-
ISIS and not pro-ISIS users. We also build classifiers to test whether contextual semantics
can help to better discriminate radical content. Our results show that the classifiers used
to build this hypothesis outperform those that disregard contextual information.
Acknowledgment
This work was supported by the EU H2020 project TRIVALENT (grant no. 740934)
References
1. Agarwal, S., Sureka, A.: Applying social media intelligence for predicting and identifying
on-line radicalization and civil unrest oriented threats. arXiv preprint arXiv:1511.06858
(2015)
2. Agarwal, S., Sureka, A.: Using knn and svm based one-class classifier for detecting online
radicalization on twitter. In: International Conference on Distributed Computing and Internet
Technology. pp. 431–442. Springer (2015)
3. Ashcroft, M., Fisher, A., Kaati, L., Omer, E., Prucha, N.: Detecting jihadist messages on
twitter. In: Intelligence and Security Informatics Conference (EISIC), 2015 European. pp.
161–164. IEEE (2015)
4. Berger, J., Strathearn, B.: Who matters online: Measuring influence, evaluating content and
countering violent extremism in online social networks. International Centre for the Study of
Radicalisation and Political Violence (2013)
5. Berger, J.M., Morgan, J.: The isis twitter census: Defining and describing the population of
isis supporters on twitter. The Brookings Project on US Relations with the Islamic World
3(20), 4–1 (2015)
6. Bermingham, A., Conway, M., McInerney, L., O’Hare, N., Smeaton, A.F.: Combining social
network analysis and sentiment analysis to explore the potential for online radicalisation. In:
Int. Conf. Advances in Social Network Analysis and Mining (ASONAM’09) (2009)
7. Carter, J.A., Maher, S., Neumann, P.R.: # greenbirds: Measuring importance and influence in
syrian foreign fighter networks (2014)
8. Chatfield, A.T., Reddick, C.G., Brajawidagda, U.: Tweeting propaganda, radicalization and
recruitment: Islamic state supporters multi-sided twitter networks. In: Proceedings of the 16th
Annual International Conference on Digital Government Research. pp. 239–249. ACM (2015)
9. Correa, D., Sureka, A.: Solutions to detect and analyze online radicalization: a survey. arXiv
preprint arXiv:1301.4916 (2013)
10. Fernandez, M., Asif, M., Alani, H.: Understanding the roots of radicalisation on twitter. In:
Proceedings of the 10th ACM Conference on Web Science. ACM (2018)
11. Klausen, J.: Tweeting the jihad: Social media networks of western foreign fighters in syria
and iraq. Studies in Conflict & Terrorism 38(1) (2015)
12. Rizzo, G., Troncy, R.: NERD: evaluating named entity recognition tools in the web of data. In:
ISWC 2011, Workshop on Web Scale Knowledge Extraction (WEKEX’11), October 23-27,
2011, Bonn, Germany. Bonn, GERMANY (10 2011)
13. Rowe, M., Saif, H.: Mining pro-isis radicalisation signals from social media users. In: Int.
Conf. Weblogs and Social Media (ICWSM). Cologne, Germany (2016)
14. Saif, H., Dickinson, T., Kastler, L., Fernandez, M., Alani, H.: A semantic graph-based
approach for radicalisation detection on social media. In: European Semantic Web Conference.
pp. 571–587. Springer (2017)
15. Saif, H., He, Y., Fernandez, M., Alani, H.: Contextual semantics for sentiment analysis of
twitter. Information Processing & Management 52(1), 5–19 (2016)
16. Vallet, D., Castells, P., Fernández, M., Mylonas, P., Avrithis, Y.: Personalized content retrieval
in context using ontological knowledge. IEEE Transactions on circuits and systems for video
technology 17(3), 336–346 (2007)
17. Vergani, M., Bliuc, A.M.: The evolution of the isislanguage: a quantitative analysis of the
language of the first year of dabiq magazine. Sicurezza, Terrorismo e Società= Security,
Terrorism and Society 2(2), 7–20 (2015)