Knowledge Graph Analysis of Russian Trolls

Chih-yuan Li

, Soon Ae Chun

and James Geller

Department of Computer Science, New Jersey Institute of Technology, Newark, NJ 07102, U.S.A.

City University of New York, College of Staten Island, New York City, NY 10314, U.S.A.

Keywords: Relationship Analysis of Troll Tweets, Entity Extraction, Triple Extraction, Sentiment Analysis.

Abstract: Social media, such as Twitter, have been exploited by trolls to manipulate political discourse and spread

disinformation during the 2016 US Presidential Election. Trolls are users of social media accounts created

with intentions to influence the public opinion by posting or reposting messages containing misleading or

inflammatory information with malicious intentions. There has been previous research that focused on troll

detection using Machine Learning approaches, and troll understanding using visualizations, such as word

clouds. In this paper, we focus on the content analysis of troll tweets to identify the major entities mentioned

and the relationships among these entities, to understand the events and statements mentioned in Russian Troll

tweets coming from the Internet Research Agency (IRA), a troll factory allegedly financed by the Russian

government. We applied several NLP techniques to develop Knowledge Graphs to understand the

relationships of entities, often mentioned by dispersed trolls, and thus hard to uncover. This integrated KG

helped to understand the substance of Russian Trolls’ influence in the election. We identified three clusters of

troll tweet content: one consisted of information supporting Donald Trump, the second for exposing and

attacking Hillary Clinton and her family, and the third for spreading other inflammatory content. We present

the observed sentiment polarization using sentiment analysis for each cluster and derive the concern index for

each cluster, which shows a measurable difference between the presidential candidates that seems to have

been reflected in the election results.

1 INTRODUCTION

Since the activities of Russian internet trolls were

discovered in the 2016 US Presidential elections, the

influence of trolls has been studied (Linvill et al.,

2019). Social media, e.g., Facebook and Twitter, have

become influential platforms of political discourse,

but were also misused by “trolls” who manipulated

the political exchanges. Trolls are users who create

social media accounts in order to post or retweet

misleading messages to negatively influence the

political process.

Different definitions of "troll" exist. Mojica

(2016) focuses on the intentions of the user, while

Kumar et al. (2014) use the term "trolling" when a

user posts and spreads disinformation. Addawood et

al., (2019) defined trolls as user accounts whose sole

purpose is to sow conflict and deception, and Jachim

et al. (2020) consider trolls as users who identify

themselves with a group that wants to cause

disruption and trigger conflict in discourse. The

Russian Trolls (RTs) were 2,752 Twitter handles

(accounts) identified by Twitter that were allegedly

tied to the Internet Research Agency (IRA), known to

be a troll farm sponsored by the Russian government,

attempting to sow discord among Americans and

influence the 2016 US election by spreading

disinformation. The networks of these trolls posted

inflammatory tweets, such as claims that Democrats

are practicing witchcraft. Some trolls created bogus

personae pretending to be BLM activists and posted

aggressive tweets. Trolls also connected with

influencers, e.g. celebrities, to manipulate them and

to amplify their malicious intents.

With the rapid data sharing on social media, the

impact of trolls can be quite damaging. Thus, the

troll-related research challenges include: (1) methods

for troll detection for distinguishing troll posts from

non-troll posts, and (2) the in-depth analysis of the

content of troll posts to further uncover the underlying

entities and their relationships in distributed posts by

different trolls. With the advances of Machine

Learning, models for the detection of trolls or their

posts were trained with high accuracy (e.g., Chun et

al., 2020).

Li, C., Chun, S. and Geller, J.

Knowledge Graph Analysis of Russian Trolls.

DOI: 10.5220/0010605403350342

In Proceedings of the 10th International Conference on Data Science, Technology and Applications (DATA 2021), pages 335-342

ISBN: 978-989-758-521-0

335

In this paper, we present new approaches of in-

depth analyses of the Russian Troll dataset, linking

disparate tweets to understand the involved entities,

and their semantic relationships. We applied several

NLP techniques including Named Entity Recognition

and Triple Extraction to derive relationships between

entities and construct a Knowledge Graph based on

the triples for the semantic analysis of troll tweets.

We also performed unsupervised clustering to

discover major clusters of troll tweets and to construct

a knowledge graph for each cluster that consists of

entities, the relations between the entities, and the

sentiments expressed by the troll tweets to get more

in-depth insights into the operations of RTs. We:

• Integrate the distributed microblog posts (tweets)

by different trolls into a Knowledge Graph to

understand the entities and their relations/events

between entities. The Knowledge Graph allows to

focus on the semantic relationships of entities

existing in different trolling posts that are not

directly visible and therefore usually ignored.

• The semantic approaches include NER and triple

extraction, leading to an understanding not only of

entities, but also of asserted statements by trolls.

• The interactive visualization of these entity-

relationship triples (Subject, Predicate, Object)

allows the users to uncover prominent events or

claims by trolls.

• The sentiment analysis performed on major troll

tweet clusters uncovers the comparative

polarizations existing in different troll clusters.

These methods can be applied to any other electoral

Twitter dataset for insights.

Section 2 explores previous work on troll

research, e.g., troll detection and discovery,

especially of right- and left-wing trolls. In Section 3,

we detail the dataset used in this work and the text

preprocessing steps. In Sections 4 and 5, we introduce

the methods used, e.g., Named Entity Recognition,

Text Clustering, Sentiment Analysis and Triple

Extraction, followed by the results, findings, and

future work.

2 RELATED WORK

A number of approaches using AI and Machine

Learning (ML) models have been used to classify troll

and non-troll tweets. (Chun et al., 2019) trained

several ML models and applied them to decide

whether a given tweet is a troll tweet. (Addawood et

al., 2019) identified linguistic cues as potential

markers of deceptive language to distinguish between

troll and non-troll tweets. (Monakhov, 2020)

proposed a quantitative measure for detecting troll

contents, which focused on certain sociolinguistic

limitations of troll speech, and discussed two

algorithms that both require only 50 tweets to

distinguish whether a message is ‘genuine’ and ‘troll-

like.’ In (Jachim et al., 2020), two automated

reasoning mechanisms for detecting and evading

trolling detection are presented, TrollHunter and

TrollHunter-Evader. While the former reached an

accuracy of 98.5% identifying trolls, the latter

undermined the performance of the former by 40%,

by manipulating the text and hashtags in the tweets.

(Seah et al., 2015) detected troll users from the

sentiments of the textual content. (Ghanem et al.,

2020) identified trolls by studying the effect of a set

of text-based features, including affective ones, and

proposed ML models that take into account topic

information. (Cambria et al., 2010) used Sentic

Computing, a new paradigm for the affective analysis

of natural language text, to extract semantics and

sentics from web-posts and hence protect web-users

from getting emotionally hurt by malicious posts.

To deal with troll contents that are not written in

English, (Miao et al., 2020) detected troll tweets in a

bilingual English and Russian corpus. (Mutlu et al.,

2016) measured the awareness level of users in

Turkey and around the world in regard to terrorism,

based on the results of troll detection. To prevent

trolls from influencing public opinion with fake

information, methods for discovering or deactivating

suspected troll accounts on Twitter have been studied.

(Im et al., 2020) developed ML models to predict

whether a Twitter handle is a Russian Troll, and the

findings imply that many RTs are likely still active

today.

The features and semantic patterns of troll tweets

have been explored. In (Chun et al., 2020), they not

only worked out troll detection, but also successfully

classified specific tweets as coming from left trolls or

right trolls. (Atanasov et al., 2019) automated the

analysis of different behavioral patterns (Left, Right,

Newsfeed) observable in the online traces of trolls, by

using ML in a realistic setting, in a supervised

learning scenario and in a distant supervision

scenario. (Iqbal et al., 2020) found patterns and topics

in tweet contents and categorized the trolls as left

trolls or right trolls.

Dynamic Exploratory Graph Analysis was

proposed (Golino et al., 2020) to discover latent

topics in the left troll and right troll tweets. Common

topics posted by right trolls include support for

Donald Trump and defending political agendas

aligned with Trump’s proposed policies, including

pro-gun, pro-police, anti-terrorism, and anti-Islam

DATA 2021 - 10th International Conference on Data Science, Technology and Applications

336

policies, etc. These tweets also "expose" and attack

Hillary Clinton. As for left trolls, the main topics are

Black Lives Matter (BLM), activities against police

brutality, and support of black culture and music.

Another study (Etudo et al., 2020) showed that the

timing of tweets about police brutality by RTs

coincided with periods of increased BLM activities.

(Badawy et al., 2018) found that RTs had a mostly

conservative, pro-Trump agenda and conservatives

amplified trolls’ messages much more often than

liberals.

As we have been looking into political trolls’

behavior, political psychology is also worthy of

exploration. (Alizadeh et al., 2019) revealed that

extremists show lower positive emotions and higher

negative emotions than partisan users, but their

differences in certainty are not significant. Moreover,

while left-wing extremists express more anxiety than

liberals, right-wing extremists were lower than

conservatives on the scale.

3 DATASET

The dataset we used is derived from the RT tweets

made available by NBC News (Popken, 2018). Under

the House Intelligence Committee investigation into

how RTs have influenced the 2016 US Election,

Twitter released almost 3,000 Twitter handles

believed to be associated with Internet Research

Agency (IRA). These accounts as well as their data

were suspended and deleted by Twitter. A team at

NBC News was able to reconstruct a dataset

consisting of a subset of the deleted data for their

investigation and were able to show how these troll

accounts went on the attack during the election

period. The dataset is freely available and includes

203,482 tweets from 454 Twitter handles.

We used this dataset to understand the underlying

entities and relationships among the trolls. In the first

step of preprocessing, we removed URLs, Twitter

handles and Non-ASCII characters from the tweets,

using the ‘re’ package of Python (McKinney, 2017).

Repetitive punctuations and spaces were also

eliminated. Then we separated camel case words and

"sticky" numbers and letters and used other data

cleansing steps as needed. Table 1 shows an example.

After standard text pre-processing, we were left with

201,366 troll tweets stored in a Python list with

2,624,037 words and 15,586,691 characters.

Table 1: Tweet before and after preprocessing.

Before Preprocessing After Preprocessing

IGetDepressedWhen someone says

he doesnÃƒÂ¢Ã‚Â€Ã‚Â™t have

any options except for voting for

Trump

I Get Depressed When

someone says he doesnt

have any options except

for voting for Trump

4 METHODS

4.1 Named Entity Recognition

Named Entity Recognition (Named-entity

recognition, n.d.) finds and classifies named entities

in text into pre-defined categories, such as person,

organization, location, date, time, quantity, etc. We

utilized spaCy (spaCy 101, n.d.) to assign named

entity labels to our tweets. SpaCy is a powerful, free

library for Natural Language Processing (NLP) in

Python. Its features include linguistic and more

general ML functionalities. There are features that

require statistical models (spaCy models, n.d.) to be

loaded so that spaCy will be able to predict linguistic

annotations, e.g., whether a word is a noun or a verb.

Different languages are supported by spaCy. ML

models also differ in format of input data, accuracy,

size, speed, etc. Users can choose the model

depending on individual cases and the input.

Normally a small or default model is a proper option

to start with. There are 18 types (Data formats, n.d.)

that spaCy can recognize, such as person, event,

location, etc. We used the (displaCy Named Entity

Visualizer, n.d.) to visualize the labeled entities.

Types are color-coded.

4.2 Troll Tweet Clustering

Clustering (Koch, 2020) is used for grouping troll

tweets based on their text into groups that contain

similar objects. We used K-Means Clustering (k-

means clustering, n.d.). As it requires numerical data

for similarity and distance measures, we used TF-

IDF, which uses term frequency and inverse

document frequency. To implement TF-IDF and

clustering, we used scikit-learn tools. We also

performed preprocessing for text clustering by

removing stop words, numbers, and punctations.

4.3 Sentiment Analysis of Trolls

To understand how Twitter users feel about politics,

we measured sentiments expressed by the collected

tweets. We used the sentiment analysis tool in the

Stanford NLP (Sentiment Analysis, n.d.) package,

Knowledge Graph Analysis of Russian Trolls

337

which uses fine-grained analysis based both on words

and labeled phrasal parse trees to train a Recursive

Neural Tensor Network (RNTN) model. The model

computes the sentiment expressed by a sentence,

based on how words compose the meaning of longer

phrases. Then the sentiments of the nodes

(representing phrases) in the parse tree are composed

to predict a sentiment value for the whole sentence.

Previous work has indicated that it is possible to

achieve an accuracy for fine-grained sentiment labels

above 80% (Socher et al., 2013).

The RNTN model works with a single sentence at

a time. To analyze tweets that include more than one

sentence, we converted periods, question marks and

exclamation marks into semicolons. Then tweets are

labeled either as "Very negative," "Negative,"

"Neutral," "Positive" or "Very positive" (Table 2).

Table 2: Sentiment analysis of example trolls.

Tex

Sentiment

Sore loser Obama turns to Russian hacking to

delegitimize Trump's triumph;

Very

negative

Hillary Clinton And you have turned the Middle

East into a living hell;

Negative

Trump Signs Obamacare Executive Orde

Neutral

In light of Hillary's FBI investigation you can

change your early vote in these states;

Positive

real Donald Trump is brilliant and has amazing

insight; He's going to be a fantastic President;

Very

positive

We analyzed the degree of negative polarization

toward a candidate. The higher the negative

polarization, the bigger the negative influence the

readers were exposed to with regards to the candidate,

thus, became more likely to vote against him/her. To

characterize the degree of negative polarization in

normalized sentiments, we used the concern index,

that we previously defined (Ji et al., 2013), as follows:

CI = N/(N+P)

N is the count of negative tweets + very negative

tweets; P = #(positive + very positive tweets).

4.4 Triple Extraction

To understand the semantic relations between two

entities, we use a triple which codifies a statement

about the entities in the form of subject–predicate–

object (Hitzler et al., 2014). To determine the

relations between entities in tweets (e.g., between two

people), we extracted such triples from tweets. We

used the Stanford Open Information Extraction

system (Stanford Open Information Extraction, n.d.)

for this. Multiple triples can be extracted from one

tweet through repeated split operations to capture

several relations in a tweet. Tweets are parsed into

sentences. Traversing a dependency parse tree

recursively, the algorithm predicts at each step

whether an edge should yield an independent clause

(Angeli et al., 2015).

Figure 1: Donald Trump (Subject) relates to Hillary Clinton

(Object) with different relations (excerpt from Neo4j

visualization, post-edited for readability).

Clauses are shortened into fragments, and

fragments are segmented into OpenIE triples. For

example, triples extracted from “Born in Boston, he

is a US citizen” would be (he, Born in, Boston) and

(he, is, US citizen). The triples (typically) retain the

core semantics of the original sentence. A set of

triples is stored as a CSV file with three columns.

4.5 Knowledge Graph for

Relationships

A large triple set with information about the same

entity spread out over many rows is not conducive to

comprehension. We used Knowledge Graph and

visualization to collect all information about one

entity (one concept as one node) at one place, for

understanding the entity-to-entity relationships.

A Knowledge Graph represents a collection of

interlinked entity descriptions. Inside the graph there

are entities such as real-world objects and events or

abstract concepts stored, and they are interlinked with

relations to form a network. As the semantic web is

developing, Knowledge Graphs are often associated

with linked data, focusing on the connections between

entities and concepts. (Ehrlinger & Wöß, 2016; Soylu

et al., 2020)

In our work, the transformation from text to a

Knowledge Graph (KG) is achieved as follows. Every

subject and every object becomes a node in the graph.

Every relation becomes a link, forming larger graphs

with overlapping nodes in several steps. For example,

two triples (Anthony Wiener, Criticize, Hillary

Clinton) and (Hillary Clinton, Delete, Email),

together can form a larger KG by sharing "Hillary

Clinton."

In a visual representation of such a graph in Neo4j

nodes appear commonly as ovals and links as arrows

pointing from the subject node to the object node.

DATA 2021 - 10th International Conference on Data Science, Technology and Applications

338

When a specific entity occurs in many rows of the

CSV file (in our case, e.g., "Hillary Clinton"), we

want to represent all instances by the same node.

Neo4j offers a setting that automatically merges

identical entities into one node. The graph data is

accessible by its own query language, called Cypher

(Cypher (query language), n.d.). Having a graph

representation makes it possible to answer many

questions of interest about a tweet set with little visual

effort. For example, we can determine which entities

have many relations, what kinds of relations they are,

and whether more of these relations are outgoing or

incoming. For instance, Figure 1 shows the Neo4j

relations between Donald Trump and Hillary Clinton,

such as "defeats," "leads," and "catches up to", etc.,

(entity labels are represented in textboxes for

readability). Collecting all these relations next to each

other in a graph provides a rich picture of how two

entities in the real world connect to each other,

according to the opinions of Twitter users. Similarly,

the conceptual distance between two entities can be

seen in a graph as the minimal number of links that

have to be traversed to get from one entity to the

other.

Figure 2: Named Entity Recognition and Label

Visualization.

5 RESULTS

5.1 Named Entities in Trolls

The six most common entity types recognized in our

dataset are shown in Table 3. Figure 2 shows

examples of the occurrences of different types of

entity labels.

Figure 3 shows the counts of person entities

extracted, with the two presidential candidates

topping the frequencies of being mentioned, and

“Hillary Clinton” appearing more often than “Donald

Trump.” This confirms the hypothesis that this dataset

is election-related. According to the aggregated

statistics "Clinton" appears more often than "Trump,"

thus we can say that Clinton (and family) are more of

a target of Russian Trolls than Trump (and family).

Table 3: Six of 18 Named Entity types.

Entity Type Count

Person 99,273

Org 60,596

(

Geo

olitical location

)

30,466

Cardinal 30,354

Date 23,727

Norp (Nation/Religion) 17,455

5.2 Clustering of Trolls

In the K-Means algorithm, the value of K depends on

how many clusters we want to partition the set of all

tweets into. This requires a trial-and-error approach.

In the beginning we set K=6, however, results were

not intuitive. As we stepwise reduced K to 3, this

dataset was naturally grouped into three clusters. One

resulting cluster of 16,851 tweets is mainly about

Clinton. Another cluster of 30,111 tweets is about

Trump. The largest cluster with 154,404 tweets

relates to Obama and other topics. This result enabled

us to separately look into how each of these three

person entities was portrayed by RTs. We collected

the top 15 terms of each cluster to obtain a general

understanding of the topic (Table 4). We can easily

identify in Cluster 1 the Trump-related tweets, Cluster

2 contains Clinton-related tweets and Cluster 3 is on

Obama and others. Figure 4 shows the word clouds

for Clinton and Trump.

Figure 3: Top 6 Frequencies of ‘PERSON’ entity.

Table 4: Top 15 terms in each cluster.

Cluste

Frequent terms

1 (Trump)

trump, donald, real, president, vote,

politics, say, america, clinton, maga, win,

media, make, obama, pence

2 (Clinton)

hillary, clinton, trump, email, campaign,

vote, prison, president, trust, obama, thing,

debate, say, politics, crooked

3 (Others)

obama, word, make, people, day, like,

thing, say, know, love, news, black, life,

new, want

Knowledge Graph Analysis of Russian Trolls

339

Figure 5: The sentiment distribution in Trump cluster (red),

and in Clinton cluster (blue).

We extracted 428,729 triples from our dataset of

201,365 troll tweets. The triples directly and

concisely show entities, e.g., persons or

organizations, and the relations between them. Table

5 shows examples of extracted triples with entities

from the Clinton and Trump clusters, including

relations between Clinton and email, between Clinton

and foundation, between Trump and media, and

between Trump and presidency.

Table 5: Example triples extracted from the tweets.

Example Triples from Clinton cluster

subject relation object

Hillary Clinton

delete

Clinton order destruction email

Clinton foundation is most corrupt

enterprise in

political history

Clinton foundation employ

Muslim

brotherhood

official

Example Triples from Trump cluster

Trump lash out a

media

Trump trash

mainstream

media

Trump is worst president

Trump

Keep workplace

protection fo

LGBTQ

Americans

The Knowledge Graph of triples (visualization of

the graph is omitted) shows the frequent relations

existing between "Trump" and "media." With main

relations such as "trash," "lash out at," and "attack,"

Russian Trolls tried to create the impression that there

existed conflicts between Trump toward and the

media.

6 CONCLUSIONS AND FUTURE

WORK

An elaborate combination of Knowledge Graph and

Natural Language Processing methodologies, such as

Named Entity Recognition, (subject, predicate,

object) triple extraction, and sentiment analysis, has

been applied in this paper to further understand the

semantic relationships among entities in trolls. The

Knowledge Graph approach has been conducted to

further understand the events or statements expressed

in the triple relationships among different entities in

troll tweet sets from the 2016 Presidential Election.

The trolls targeted one candidate, Hillary Clinton, and

her family, by repeatedly accusing her of the "email-

gate scandal" and of misuse of the Clinton foundation,

etc. The concern index in the Clinton cluster of troll

tweets was the highest, and over 10% higher than in

the other two clusters, which shows that Russian

Trolls had used many more negative terms portraying

Clinton than Trump.

We plan to further validate the effectiveness of the

Knowledge Graph approach to integrating micro-

blogging posts that are often not connected to each

other, as they are coming from many different

accounts. We seek to uncover more of the underlying

semantic relationships among entities in such

Knowledge Graphs.

Other future work includes: (1) applying the

Knowledge Graph approach to other election and troll

datasets, such as tweets from Alt-Right and Alt-Left

groups; (2) building more accurate Machine Learning

models that are location and language sensitive (e.g.,

by country) and comparing their influences; (3)

distinguishing between professional trolls and

amateur trolls; (4) an analysis of statistical

significance of differences observed; (5) semantic

deduplication (i.e., which "Trump" is "Donald

Trump" and which is another member of the Trump

family); (6) extension to non-Russian trolls; and (7) a

deeper analysis of insights that can be gained from the

KGs.

ACKNOWLEDGEMENTS

This work partially supported with grants from NSF

CNS 1747728, NSF CNS1624503, and NRF-Korea:

2017S1A3A2066084.

REFERENCES

Addawood, A., Badawy, A., Lerman, K., and Ferrara, E.

(2019). Linguistic Cues to Deception: Identifying

Political Trolls on Social Media. Proceedings of the

International AAAI Conference on Web and Social

DATA 2021 - 10th International Conference on Data Science, Technology and Applications

340

Media, 13(01), 15-25. Retrieved from

ojs.aaai.org/index.php/ICWSM/article/view/3205

Alizadeh, M., Weber, I., Cioffi-Revilla, C. et al. Psychology

and morality of political extremists: evidence from

Twitter language analysis of alt-right and Antifa. EPJ

Data Sci. 8, 17 (2019). https://doi.org/

10.1140/epjds/s13688-019-0193-9

Angeli, G., Premkumar, M. J. J, Manning, C. D. (2015).

Leveraging Linguistic Structure for Open Domain

Information Extraction. Proc. of the 53rd Ann. Mtg. of

the ACL and the 7th Int. Joint Conference on Natural

Language Processing (V. 1) (pp. 344–354). Beijing,

ACL.

Atanasov, A., Morales, G., & Nakov, P. (2019). Predicting

the Role of Political Trolls in Social Media. ArXiv,

abs/1910.02001.

Badawy, A., Ferrara, E., and Lerman, K., (2018)

"Analyzing the Digital Traces of Political

Manipulation: The 2016 Russian Interference Twitter

Campaign," IEEE/ACM Int. ASONAM, pp. 258-265,

2018.

Cambria, E., Chandra, P., Sharma, A., Hussain, A. (2010).

Do Not Feel The Trolls. CEUR Workshop Proceedings.

664.

Chun, S. A., Holowczak, R., Dharan, K. N., Wang, R.,

Basu, S., & Geller, J. (2019). Detecting political bias

trolls in Twitter data. In A. Bozzon, F. J. D. Mayo, & J.

Filipe (Eds.), WEBIST 2019 - Proc. of the 15th Int.

Conf. on Web Information Systems and Technologies

(pp. 334-342).

Cypher (query language). (n.d.). Retrieved from Wikipedia:

https://en.wikipedia.org/wiki/Cypher_(query_language

)

Data formats. (n.d.). Retrieved from spacy.io:

https://spacy.io/api/data-formats#named-entities

displaCy Named Entity Visualizer. (n.d.). Retrieved from

explosion.ai/: https://explosion.ai/demos/displacy-ent

Ehrlinger, L. and Wöß, W. (2016). Towards a Definition of

Knowledge Graphs.

Etudo, U., Yoon, V.Y., Yaraghi, N. (2019). From Facebook

to the Streets: Russian Troll Ads and Black Lives

Matter Protests. HICSS.

Fivethirtyeight, Russian-troll-tweets, https://github.com/

fivethirtyeight/russian-troll-tweets/ Retr. 1/ 2019.

Ghanem B., Buscaldi D., Rosso P. (2020). TexTrolls:

Identifying Trolls on Twitter with Textual and

Affective Features. In: Proc. Workshop on Online

Misinformation- and Harm-Aware Recommender

Systems (OHARS), Co-located with RecSys 2020,

CEUR Workshop Proceedings.CEUR-WS.org, vol.

2758, pp. 4-22

Golino, H., Christensen, A., Moulder, R., Kim, S., Boker,

Steven. (2020). Modeling latent topics in social media

using Dynamic Exploratory Graph Analysis: The case

of the right-wing and left-wing trolls in the 2016 US

elections. 10.31234/osf.io/tfs7c.

Hitzler, P., Lehmann, J., Polleres, A. (2014). Logics for the

Semantic Web, Editor(s): Jörg H. Siekmann, Handbook

of the History of Logic, North-Holland, Volume 9,

Pages 679-710.

Im, J., Chandrasekharan, E., Sargent, J., Lighthammer, P.,

Denby, T., Bhargava, A., Hemphill, L., Jurgens, D., &

Gilbert, E. (2020). Still out there: Modeling and

Identifying Russian Troll Accounts on Twitter. 12th

ACM Conference on Web Science.

Iqbal, S., Keshtkar, F., Chun, S. A. (2020) Extract Semantic

Pattern from Trolling Data, FLAIRS-33 (pp. 509-514).

Iqbal, S., Chun, S. A., Keshtkar, F. (2020) Using

Computational Linguistics to Extract Semantic Patterns

from Trolling Data.Proceedings of IEEE 14th

International Conference on Semantic Computing

(ICSC 2020): 369-374

Jachim, P., Sharevski, F., Treebridge, P. (2020).

TrollHunter [Evader]: Automated Detection [Evasion]

of Twitter Trolls During the COVID-19 Pandemic.

New Security Paradigms Workshop (pp. 59-75). New

York, NY: ACM.

Ji, X., Chun, S. A., and Geller, J., "Monitoring Public

Health Concerns Using Twitter Sentiment

Classifications," 2013 IEEE International Conference

on Healthcare Informatics, Philadelphia, PA, USA,

2013, pp. 335-344, doi: 10.1109/ICHI.2013.47.

Kersting, J., Geierhos, M. (2020). Neural Learning for

Aspect Phrase Extraction and Classification in

Sentiment Analysis. The 33rd International FLAIRS

(pp. 282-285).

K-means clustering. (n.d.). Retrieved from Wikipedia:

https://en.wikipedia.org/wiki/K-means_clustering

Koch, K., 2020. A Friendly Introduction to Text Clustering,

https://towardsdatascience.com/a-friendly-introduction

-to-text-clustering-fa996bcefd04, Retrieved Jan. 29,

2021.

Kumar, S., Spezzano, F., Subrahmanian, V.S., 2014.

Accurately detecting trolls in slashdot zoo via

decluttering. In Proc. of ASONAM ’14, 188–195,

Beijing, China.

Lewinski, D., Hasan, M. R., “Russian Troll Account

Classification with Twitter and Facebook Data”, arXiv

e-prints, 2021.

Linvill, D., Boatwright, B., Grant, W., Warren, P. (2019).

“The Russians are Hacking my Brain!” investigating

Russia's internet research agency twitter tactics during

the 2016 US presidential campaign. Computers in

Human Behavior. 99. 10.1016/j.chb.2019.05.027.

Miao, L., Last, M., Litvak, M. (2020). Detecting Troll

Tweets in a Bilingual Corpus. Proc. of the 12th

Language Resources and Evaluation Conf. (pp. 6247–

6254). Marseille, France: European Language

Resources Association.

McKinney, W., 2017. Python for Data Analysis, Data

Wrangling with Pandas, NumPy, and IPython. O'Reilly.

Mojica, L. G., 2017. A Trolling Hierarchy in Social Media

and a Conditional Random Field for Trolling (Richard

Socher, 2013) Detection, arXiv:1704.02385v1 [cs.CL].

Monakhov, S. (2020) Early detection of internet trolls:

Introducing an algorithm based on word pairs / single

words multiple repetition ratio. PLoS ONE 15(8):

e0236832. https://doi.org/10.1371/journal.pone.0236

832

Knowledge Graph Analysis of Russian Trolls

341

Mutlu, B., Mutlu, M., Oztoprak, K., Dogdu, E. (2016).

Identifying Trolls and Determining Terror Awareness

Level in Social Networks Using a Scalable Framework.

10.1109/BigData.2016.7840796.

Named-entity recognition. (n.d.). Retrieved from

wikipedia: https://en.wikipedia.org/wiki/Named-entity

recognition

Popken, B "Twitter deleted 200,000 Russian troll tweets.

Read them here.," (2018). Available: https://www.nbc

news.com/tech/social-media/now-available-more-200-

000-deleted-russian-troll-tweets-n844731.

Seah, C. W., Chieu, H. L., Chai, K. M. A., Teow, L., Yeong,

L. W. "Troll detection by domain-adapting sentiment

analysis," 2015 18th International Conference on

Information Fusion (Fusion), Washington, DC, USA,

2015, pp. 792-799.

Socher, R., Perelygin, A., Wu, J., Chuang, J., Manning, C.

D., Ng, A., Potts, C. (2013). Recursive Deep Models for

Semantic Compositionality Over a Sentiment

Treebank. Proc. of the 2013 Conf. on Empirical

Methods in Natural Language Processing (pp. 1631–

1642). Seattle, Washington.

Soylu, A., Corcho, O., Elvesaeter, B., Badenes-Olmedo, C.,

Yedro, F., Kovacic, M., Posinkovic, M., Makgill, I.,

Taggart, C., Simperl, E., Lech, T., Roman, D. (2020).

Enhancing Public Procurement in the European Union

through Constructing and Exploiting an Integrated

Knowledge Graph.

Roeder, O., 2018. Why We’re Sharing 3 Million Russian

Troll Tweets https://fivethirtyeight.com/features/why-

were-sharing-3-million-russian-troll-tweets/, retrieved

6/3/2019.

Sentiment Analysis. (n.d.). Retrieved from stanford:

https://nlp.stanford.edu/sentiment/

spaCy 101: Everything you need to know. (n.d.). Retrieved

from spacy.io: https://spacy.io/usage/spacy-101

Stanford Open Information Extraction. (n.d.). Retrieved

from https://nlp.stanford.edu/software/openie.html

Trained Models & Pipelines. (n.d.). Retrieved from

spacy.io: https://spacy.io/models

Wikipedia contributors, 2019. Internet Research Agency. In

Wikipedia, The Free Encyclopedia. https://en.wikipe

dia.org/w/index.php?title=Internet_Research_Agency

&oldid=900092717, Retrieved June 3, 201.

DATA 2021 - 10th International Conference on Data Science, Technology and Applications

342