ABUSE OF POWER: COORDINATED ONLINE HARASSMENT OF FINNISH GOVERNMENT MINISTERS 978-9934-564-97-0
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
978-9934-564-97-0
@
!%^#@ D!
B@*h
F*%#
ABUSE OF POWER: COORDINATED
ONLINE HARASSMENT OF FINNISH
GOVERNMENT MINISTERS
Published by the
NATO Strategic Communications
Centre of Excellence
ISBN: 978-9934-564-97-0 Project manager: Rolf Fredheim Authors: Kristina Van Sant, Rolf Fredheim, and Gundars Bergmanis-Korāts Copy-editing: Anna Reynolds Design: Kārlis Ulmanis Riga, February 2021 NATO STRATCOM COE 11b Kalnciema Iela Riga LV1048, Latvia www.stratcomcoe.org Facebook/stratcomcoe Twitter: @stratcomcoe This report was completed in November 2020, and based on data collected from March to July 2020 This publication does not represent the opinions or policies of NATO or NATO StratCom COE. © All rights reserved by the NATO StratCom COE. Reports may not be copied, reproduced, distributed or publicly displayed without reference to the NATO StratCom COE. The views expressed here do not represent the views of NATO.
PAGE LEFT
BLANK
INTENTIONALLY
The main topics triggering abusive messages were the COVID-19 pandemic, immigration, Finnish-EU relations, and socially liberal politics. Executive Summary This report is an explorative analysis a period spanning the state of emergency of abusive messages targeting Finnish declared in response to the COVID-19 ministers on the social media platform pandemic.1 Twitter. The purpose of this study is to understand the scope of politically This report is informed by the findings of motivated abusive language on Finnish three recent Finnish studies, one of which Twitter, and to determine if, and to what investigated the extent and effects of online extent, it is perpetrated by inauthentic hate speech against politicians while the accounts. To this end, we developed a other two studied the use of bots to influence mixed methodology, combining AI-driven political discourse during the 2019 Finnish quantitative visualisations of the networks parliamentary elections. The first study, delivering messages of abuse with a released by the research branch of the Finnish qualitative analysis of the messages in government in November 2019, found that a order to understand the themes and triggers third of municipal decision-makers and nearly of abusive activity. We collected Twitter half of all members of Finnish Parliament data between 12 March and 27 July 2020, have been subjected to hate speech online. 1 Finland declared a state of emergency on 16 March 2020 that was in force for three months until 16 June 2020, although the Emergency Powers Act remained in force through the end of June. 4 �����������������������������������������������������������������������������
The two studies tracking inauthentic activity languages and either not generally focused
during the 2019 parliamentary elections on Finland or used to push certain causes
identified bot interference but concluded in multiple languages. We repeatedly came
that the impact of these bots on Finland’s across a cluster of accounts throughout
political environment appeared limited. Based our monitoring period that posted the same
on these findings, and on our comprehensive messages about animal cruelty and climate
literature review, we developed two change. These accounts predominantly post
hypotheses: in English and appear in some cases to be
automated or semi-automated. However,
1. We expect to observe abusive they represent a very small part of the
language targeting Finnish conversation. Likewise, a small cluster of
politicians, with female politicians automated accounts amplified messaging by
receiving gendered abuse; a number of right-wing voices. Again, there
was a degree of coordination here, but these
2. We expect to observe low levels amplifications looked more like attempts
of coordinated inauthentic activity at self-promotion rather than systematic
in the Finnish information space, manipulation of the information space. If
with increased levels of inauthentic large-scale inauthentic coordination exists
activity during periods of political in the Finnish information environment, we
significance. are either looking in the wrong place, or it is
so sophisticated or so small in scale that it
Our quantitative and qualitative analyses evades our detection methods.
confirmed both hypotheses and yielded
multiple findings. Our investigation We found that the main topics triggering
demonstrated that the messaging directed abusive messages were the COVID-19
at Finnish government officials is largely pandemic, issues of immigration, Finnish-
free from automated activity. When it comes EU relations, and socially liberal politics.
to abusive messaging, we find a number of We observed that female Finnish ministers
users singularly focused on harassing the received a disproportionate number of
government. While both left- and right-leaning abusive messages throughout our monitoring
communities engaged in abusive activity, the period. A startling portion of this abuse
bulk of abusive messaging originated from contained both latent and overtly sexist
clusters of right-wing accounts. language, as well as sexually explicit
language. Although we found large volumes
Overall, we observed very low levels of both of offensive and abusive messaging, we did
bot and coordinated activity. The majority of not observe threats of physical violence.
bots we identified were operating in foreign
���������������������������������������������������������������������������� 5
Social media has become an essential platform for political engagement, granting citizens unprecedented access to their government representatives. Introduction Lipstick brigade. Lipstick girls. Feminist challenges for states navigating the complex quintet. Tampax team. These are all phrases relationship between freedom of speech and used on Twitter to refer to the current coalition protection from harmful discourse, as online in Finland, in which all five party leaders are hate speech and abusive messaging have women, led by Prime Minister Sanna Marin become increasingly recognised as socio- of the Social Democratic Party. When the political issues. Social media has become an remarkably young and female leadership essential platform for political engagement, came into power in December 2019, they granting citizens unprecedented access to made international headlines as pioneers of their government representatives. Twitter gender equality in governance. Their election in particular has provided candidates and also provoked online resistance in the form of constituents with an informal channel of abusive messages. Many assumptions about communication, through which citizens can their political inexperience were accompanied share feedback and politicians have the by sexist and misogynistic language. ability to engage with these concerns directly. Social media platforms provide individuals However, this unfettered access to politicians with virtually limitless opportunities for online, combined with the anonymous communication and self-expression. This nature of social media platforms, has led potential, though transformative, has raised to government officials being targeted with 6 �����������������������������������������������������������������������������
abusive messages. This virtual vitriol can in response to the COVID-19 pandemic and
take many forms: it can be threatening, several weeks after it was lifted.
misogynistic, racist, vulgar, and so on.
For governments, online harassment is a The report is structured as follows. The
growing concern, as it can have the effect of literature review engages with the scholarly
discouraging participation in public service, literature discussing definitions and
particularly among women. Simultaneously, methods of detecting abusive language on
the rise in fake account activity in online social media platforms, abuse of politicians
political discourse is equally concerning, online, misogyny online, and the use of bots
as recent examples highlight the impact for political purposes on Twitter. Having
inauthentic activity can have on public established this framework, we will describe
opinion and political participation. our methodological approach for analysing
the data, which combines social network
This study is an analysis of how abusive analysis, bot detection, hate speech detection,
messaging intersects with the activity of fake and narrative analysis. This combination of
Twitter accounts in the political sphere. In quantitative and qualitative approaches is
this explorative analysis, we will be focusing designed to identify instances when accounts
on the state of politically motivated online coordinate to send abusive messages to
abuse in Finland. Specifically, we will be politicians. The study continues with a social
analysing messages directed at Finnish network analysis that informs the basis of the
ministers between 12 March and 27 July qualitative analysis. We conclude our study
2020, encompassing the three months with a discussion of our findings, conclusions,
Finland maintained a state of emergency and policy recommendations.
���������������������������������������������������������������������������� 7
The Proliferation of Online Abuse
Defining hate speech and abusive (Waseem et al, 2017) and is the most
language frequently used phrase for describing the
phenomenon of insulting user-generated
The concept of hate speech, considered content (Schmidt and Wiegand, 2017).
an umbrella term for abusive user-created Broadly, hate speech is defined as any
content, does not have a single formal communication that disparages a person
definition. Rather, hate speech is interpreted or group on the basis of a particular trait,
as a collection of overlapping terms, such as race, color, ethnicity, gender, sexual
including designations such as cyberbullying, orientation, nationality, or religion, among
abusive language, and hostile language other characteristics (Nockleby, 2000).
Explicit Implicit
“Go kill yourself”, “You’re a sad little f*ck” (Van “Hey Brendan, you look gorgeous today. What
beauty salon did you visit?”
(Dinakar et al., 2012),
“@User shut yo beaner ass up sp*c and hop
your f*ggot ass back across the border little
Directed
“(((@User))) and what is your job? Writing
n*gga” (Davidson et al., 2017), cuck articles and slurping Google balls?
#Dumbgoogles” (Hine et al., 2017),
‘Youre one of the ugliest b*tches Ive ever
fucking seen” (Kontostathis et al., 2013). “you’re intelligence is so breathtaking!!!!!!”
(Dinkar et al., 2011).
“I am surprised they reported on this crap “Totally fed up with the way this country has
who cares about another dead n*gger?”, “300 turned into a haven for terrorists. Send them
missiles are cool! Love to see um launched all back home.” (Burnap and Williams, 2015),
Generalised
into Tel Aviv! Kill all the g*ys there!” (Nobata
et al., 2016),
“most of them come north and are good at
just mowing lawns” (Dinakar et al., 2011),
“So an 11 year old n*gger girl killed herself
over my tweets? ^_^ that’s another n*gger off
the streets!!” (Kwok and Wang, 2013). “Gas the skypes” (Magu et al., 2017).
Table 1: Typology of abusive language (Waseem at al, 2017)
8 �����������������������������������������������������������������������������
Davidson et al (2017) differentiate hate of hate speech as a political tool.
speech from offensive language, defining it Accusations and prosecution of politicians
as, “language that is used to express hatred for hate speech is an established reality
towards a targeted group or is intended in modern democracies, especially with
to be derogatory, to humiliate, or to insult the rise of anti-immigration parties in
the members of the group.” In an attempt Western Europe in recent decades (van
to synthesise varying definitions of hate Spenje and de Vreese, 2013). Extreme-
speech, Waseem et al (2017) proposed a right political parties in Spain were found
typology establishing a distinction between to imply discrimination on their Facebook
explicit and implicit abusive language. pages, which was then exacerbated by
Explicit abuse is clearly derogatory, their followers who used hate speech in
including language that contains racist or the comment sections (Ben-David and
sexist slurs, while implicit abuse is less Matamoros-Fernandez, 2016). But what is
overt, often obscured by the use of sarcasm the state of abuse directed at politicians?
and other ambiguous language, making it Little academic work exists regarding the
more difficult to detect qualitatively and extent of abusive messages addressed to
with machine learning approaches. When politicians online (Gorrell et al, 2018).
discussing abusive language in this report,
we will refer to this typology (Table 1). Social media has become an essential
platform for political engagement, voter
mobilisation, electoral campaigning, and
Targets of online abuse intimate communication between political
candidates and the public. Despite the
Existing studies on hate speech have typically opportunity for individual interaction,
focused on specific forms of abuse, such Theocharis et al (2016) found that
as racism and homophobia, as well as on politicians prefer to engage in broadcasting-
rhetoric shared by hate groups and content style communication. The authors
circulated on radical forums (Silva et al, 2016). hypothesised that this is the case because
For the purposes of this study, we will discuss politicians are reluctant to invite the vitriol
literature that examines hate speech directed of citizens empowered by the anonymity
at politicians, misogynistic hate speech, and of Twitter. Their results lend support to
abuse targeting female politicians. this theory, as they found that candidates
with more engaging messages are also
Political abuse more exposed to criticism and harassment
online (Theocharis et al, 2016). Ward and
The scholarly debate surrounding hate McLoughlin (2020) identify four explanatory
speech and politicians is dominated by themes for abuse of British MPs: mental
studies investigating the employment illness among members of the public;
���������������������������������������������������������������������������� 9
the nature of social media; the increase in to 29 reported being sexually harassed
political polarisation and extremism; and online, a figure that is more than twice
social problems around identity issues, such the percentage of men in the same age
as race, religion, and gender. bracket (Pew Research Center, 2017).
Between December 2016 and March
A study released by the research branch of 2018, Amnesty International conducted
the Finnish government in November 2019 qualitative and quantitative research into
explored the nature and extent of hate speech women’s experiences on social media
targeting Finnish politicians. The study, platforms. During their investigation, the
which constituted the first in Finland on how authors found that Twitter fosters a toxic
societal decision-making may be influenced and unregulated underbelly of violence
by hate speech, found that a third of and abuse against women (Amnesty
municipal decision-makers and nearly half of International, 2018).
all members of Finnish Parliament have been
subjected to hate speech online. Additionally, Mantilla (2013) distinguishes
two-thirds of policymakers surveyed believe “gendertrolling”, the targeting of women
that hate speech has increased in recent with abuse online, from generic trolling.
years. Hate speech directed at public Gendertrolling is defined by the following
servants may have a negative impact on features: gender-based insults; vicious
political participation, as 28% of municipal language; credible threats; the participation,
officials who were targeted with hate speech often coordinated, of numerous people;
expressed that the experience decreased their unusual intensity, scope, and longevity of
willingness to participate in decision-making attacks; and reaction to women speaking
(Knuutila et al, 2019). The findings of the out (Mantilla, 2013: 564–65). Mantilla
Finnish government provide the foundational argues that gendertrolling systematically
basis for this project. targets women to prevent them from fully
occupying public spaces, particularly
Online misogyny traditionally male-dominated arenas
(ibid., 569).
Research has repeatedly found that
women are subjected to more online abuse, A 2016 Inter-Parliamentary Union (IPU)
bullying, hateful language, and threats report discusses the three characteristics
than men (Bartlett et al, 2014). A 2017 that distinguish violence against women in
survey by the US-based Pew Research politics: (1) Women are targeted because
Center found that women are much of their gender; (2) the abuse itself can be
more likely to experience severe types highly gendered, as exemplified by sexist
of gender-based or sexual harassment abuse and threats of sexual violence; (3)
than men. In fact, 21% of women aged 18 its impact is to discourage women from
10 ����������������������������������������������������������������������������The targeting of women with gender-
based abuse online, particularly women in
positions of power, has become a global
phenomenon.
becoming or continuing to be active in found that female MPs were more likely to
politics. The IPU identified a number of receive tweets that stereotyped them or
factors that exacerbate the vulnerability questioned their position as government
of certain women parliamentarians to representatives. While they identified clearly
gender-based abuse. These aggravating gendered patterns in the data, the authors
factors include belonging to the political concluded that there are fewer differences
opposition, being under 40 years old, and in how female and male politicians are
belonging to a minority group, where sexism addressed than previously expected. In
is often compounded by racism. Alarmingly, their explorative analysis of instances of
the IPU found that this phenomenon exists sexist hate speech and abusive language
to varying degrees in every country (Inter- against female politicians on Twitter in
Parliamentary Union, 2016). Japan, Fuchs and Schäfer (2020) found
that negative attitudes expressed towards
In 2017, Amnesty International carried female politicians have become a common
out a small-scale investigation into sexist trend on the platform, especially towards
and racist abuse faced by women in UK controversial or more prominent female
politics on Twitter. In the run-up to the 2017 politicians.
election, Amnesty International researchers
found that Labour MP Diane Abbott The targeting of women with gender-
received almost half of all abusive tweets based abuse online, particularly women in
and that black and Asian women MPs positions of power, has become a global
received 35% more abusive tweets than phenomenon. The anonymous nature of
white women. Notably, online abuse did social media platforms, such as Twitter,
not adhere to party lines: women from all has empowered individuals to engage in
UK political parties were targeted by sexist abusive discourse online. Recent studies
hate speech (Dhrodia, 2017). Southern and have identified patterns in gendered abuse
Harmer (2019) conducted a comparative on social media and have attempted to
study of the experiences of less prominent explain their ubiquitous prevalence. The
male and female MPs online in which they possible impacts of sexist abuse against
��������������������������������������������������������������������������� 11women politicians are vast, ranging from sentence structure of derogatory social
psychological distress to discouragement media posts to identify instances of hate
from participating in public service. Given speech with a high rate of precision. In
that 11 of the 19 ministers appointed their article, Burnap and Williams (2015)
to the current Finnish government are developed a supervised machine learning
women, we expect to observe gender-based classifier of hateful content on Twitter
abuse targeting these ministers and the for the purpose of monitoring the public
functioning of the government as a whole. reaction to highly emotive events, such as
a terror attack.
H1 We expect to observe Fuchs and Schäfer (2020) adopted corpus-
abusive language targeting based discourse analysis (CDA), a mixed-
Finnish politicians, with female methods approach that combines critical
discourse analysis with corpus linguistics,
particularly keyword analysis based on
frequency and occurrence patterns. In
order to investigate misogynistic hate
Automated hate speech detection speech, Hewitt et al (2016) gathered
thousands of tweets using a range of sexist
As the volume of user-generated content terms, disregarding irrelevant commercial
on social media platforms has surged, so messages, messages in foreign languages,
has the amount of hate speech circulating or completely unintelligible tweets, and
online and, consequently, the demand for manually coded the remaining sample
hate speech detection tools (Schmidt and using a simple binary model. Badjatiya et al
Wiegand, 2017). However, both manual and (2017) applied deep learning architectures
automated detection methods are hindered to the problem of identifying hate speech
by the lack of a clear definition of hate on Twitter, which they define as being
speech and the often-ambiguous nature able to classify a tweet as racist, sexist,
of verbal abuse. Previous studies have or neither. Despite their variance, most
identified hateful and antagonistic content techniques for detecting hate speech on
through various qualitative and quantitative social media platforms–namely Twitter–
methodological approaches. Among them incorporate machine-learning-driven data
is the bag-of-words (BoW) approach, a collection and identification algorithms
technique of natural language processing with in-depth qualitative analysis to
that uses words within a corpus to classify address the shortcomings of automated
hate speech but is prone to misclassification identification and enhance understanding
(Greevy and Smeaton, 2004). In addition to of the content of hateful messages.
keywords, Silva et al (2016) leveraged the
12 ����������������������������������������������������������������������������Social Media Companies Struggle to Keep up with Hate
!%^#@
Speech
Companies such as Microsoft, Twitter, YouTube, and Facebook have signed the EU’s Code of
Conduct on Countering Illegal Hate Speech Online (European Commission, 2020), which compels
the companies to remove any post containing hate speech within 24 hours. Despite the huge
resources and data availability of these tech giants, social media companies to this day find it
hard to automate hate speech detection. One indication of this is their decision to publicly release
datasets to allow the general public to contribute to the challenge (Vidgen and Derczynski, 2020;
Davidson et al, 2017; de Gibert et al, 2018; Ahlgren et al, 2020).
Recently Facebook has invested heavily to proactively detect 88.8% of the total hate
in measures to control toxic content and speech content they remove, up from 80.2%
inauthentic behaviour of platform users the previous quarter thanks to progress in two
(Ahlgren et al, 2020). As an example, key areas: deeper semantic understanding
Facebook previously mainly outsourced of language (the detection of more subtle
its content moderation to a relatively small and complex meanings) and broadening the
group of reviewers around the world. capacity of AI tools to understand content
However, the volume of potentially abusive (holistic understanding of content) (Dansby
messaging was such that Facebook is et al, 2020). Even though almost 90% sounds
currently investing heavily in Artificial impressive, this is relatively low compared
Intelligence solutions to automate the to simpler classification tasks (for instance
process. Despite recent advancements in automatically detecting pornography), and
AI, the current level of development is far certainly much too low to hand the process
behind necessary efficiency levels as it must over to the machines entirely. In the case
be able to understand content holistically of content moderation, the ~10% of data
(the way we perceive). missed may reach and harm a significant
number of platform users, especially
To address this problem and to boost AI considering that Facebook currently has
development in this direction, Facebook 2.7 billion active users worldwide (Clement,
launched a hate speech challenge to detect 2020). In this case, the emphasis should not
meme-based political hate speech using a be on the number of harmful posts removed,
data set of 10,000+ new multimodal examples but on the total number of missed posts that
(Kiela et al, 2020). Facebook recently may potentially cause harm.
claimed that its current AI solution is ableAutomation and Political
Contestation Online
The role of bots and trolls in political humans over a diverse range of social
discourse media platforms, have been used to
infiltrate political discourse, manipulate the
Bots—short for software robots—are stock market, steal personal information,
computer programs that perform tasks and spread disinformation (Ferrara et
automatically and have been a staple of al, 2016). Political bots are social bots
online activity since the advent of computers. used as tools for politics and propaganda,
While much bot activity is benign, they such as posting carefully staged photos
can also be utilised by economically or and well-crafted responses in pursuit of
politically motivated actors to carry out political objectives (Howard et al, 2018).
malicious activities such as launching The use of bots can be classified as
spam, distributed denial-of-service (DDoS) inauthentic behaviour. Facebook defines
attacks, click fraud, cyberwarfare (Kollanyi inauthentic behaviour as the use of
et al, 2016), and the manipulation of public accounts, pages, groups, or events to
opinion (Woolley and Howard, 2016). The mislead social media platform users and
label troll describes persons who start the platforms themselves about:
arguments online to elicit an emotional
response, typically outrage. This practice is t he identity, purpose, or origin of the
widely known as trolling. Political trolling entity that they represent;
and astroturfing are terms that refer to the the popularity of content or assets;
artificial promotion of messages online the purpose of an audience or
in order to manufacture a false sense of community;
popularity or support for these messages the source or origin of content;
(Bradshaw and Howard, 2018).
or to evade enforcement of a platform’s
As contemporary social media ecosystems Community Standards.
have increased in scope and complexity,
the demand for bots that convincingly Furthermore, coordinated inauthentic
mimic human behaviour has risen in activity is defined by Facebook as the use
parallel. Social bots, scripts designed of multiple assets, working in concert to
to produce content and to interact with engage in inauthentic behaviour, where
14 ����������������������������������������������������������������������������the use of fake accounts is central to the Political use of bots on Twitter
operation (Facebook, 2020).
Recent academic studies are replete
Social media companies themselves estimate with instances where social bots have
that about 5–8% of accounts are fake or bot been used as instruments to carry out
accounts, but scholars tend to see these political campaigns or to shape the
estimates as conservative. For instance, political conversation on Twitter. Wooley
Varol et al (2017) estimate that bot accounts (2016) identified three main ways political
make up 9–15% of all Twitter users. However, bots have been employed: to demobilise
studies of non-English data sets have pointed opposition, to disseminate pro-government
to even higher bot concentrations. Filer messages, and to inflate follower counts.
and Fredheim (2015) found that Russian- Political actors and governments use
language discussions are at times conducted bots to manipulate public opinion,
virtually exclusively by bots. In their 2016 choke off debate, and muddy political
Bot Traffic Report, the security company issues (Howard and Kollanyi, 2017).
Imperva estimated that over half of online Political bots can engage in weaponised
traffic was attributed to bots, nearly 30% of communication—the strategic use of
which were categorised as ‘bad bots,’ such as communication as an instrumental tool to
impersonators and spammers. gain compliance and avoid accountability
(Mercieca, 2019). Such campaigns have
In recent years, Facebook, Twitter, and been documented in Argentina, Australia,
other social media platforms have received Azerbaijan, Bahrain, China, Iran, Italy,
intense criticism for their treatment Mexico, Morocco, Russia, South Korea,
of users and their personal data, their Saudi Arabia, Turkey, the United Kingdom,
insufficient responses to pervasive issues the United States, and Venezuela (Wooley,
such as hate speech, the dissemination 2016).
of mis- and disinformation by bots, and
their systematic evasion of critical and The two most widely studied cases of
independent oversight (Bruns, 2019). Ünver political bot interference both occurred in
(2019) argues that the debate at the heart 2016, during the US presidential election
of the bot problem is whether technology and the UK-EU Referendum. In a study of
companies are deliberately, or at least the 2016 US presidential election debates,
passively, facilitating negative political Kollanyi et al (2016) found one third of pro-
messaging. Despite strengthening their Trump tweets were driven by bots and highly
efforts against inauthentic activity, social automated accounts, suggesting bots were
media platforms have not been able to get used to amplify the popularity of pro-Trump
malicious bot activity under control (Bay messages on Twitter. The authors later
and Fredheim, 2019). found that political bot activity reached an
��������������������������������������������������������������������������� 15all-time high during the 2016 presidential In Finland, several Aalto University studies
campaign. By election day, the gap between have investigated the use of bots to influence
highly automated pro-Trump and pro-Clinton political discourse during the 2019 Finnish
messaging was 5:1, indicating a deliberate parliamentary elections. Rossi (2019)
and strategic attempt to sway the outcome developed a machine-learning tool for
of the election (Kollanyi et al, 2016). detecting bots on Twitter to analyse whether
Howard and Kollanyi (2016) found that bots inauthentic accounts were used to influence
generated a noticeable portion of all traffic voters. According to this model, over 30%
surrounding the UK Brexit referendum. of followers of Finland’s top politicians
These messages predominantly supported were classified as bots, indicating that they
the Vote Leave campaign. may have been used to increase the Twitter
following of certain politicians. However,
Since 2016, the deployment of bots apart from artificial follower inflation, the
during important and divisive national and impact of these bots on Finland’s political
international political moments has become a environment appears limited. Researchers
staple of the 21st century information space. working on the ELEBOT project funded by the
In 2019, during the Hong Kong protests, Finnish Ministry of Justice concluded that
Twitter and Facebook took action against the volume of Twitter-bots and their influence
China for using hundreds of fake accounts to on the Finnish elections were minimal. The
sow political discord. Twitter announced that researchers visualised communities whose
it was suspending nearly a thousand Chinese members were linked and identified the three
accounts, citing a “significant state-backed largest communities. The smallest of the
information operation” (Baca and Romm, three communities, which included around
2019). Similarly, during mass demonstrations 7% of all users and 8% of all bots, primarily
seeking political change in Lebanon, it was used language relating to immigration and to
discovered that accounts tweeting a pro- the right-wing populist Finns Party. Overall,
government hashtag had a higher likelihood authentic Twitter users rarely engaged with
of displaying automated behaviour (Skinner, tweets produced by bots (Salloum et al,
2019). 2019).
16 ����������������������������������������������������������������������������H2 e expect to observe low levels of coordinated inauthentic activity in the Finnish
W
information space, with increased levels of inauthentic activity during periods of
political significance.
Methodological Approach
For the purposes of this report, we used before Prime Minister Marin declared a
data collected from Twitter. From a state of emergency in Finland due to the
technical perspective, this decision is easily COVID-19 pandemic. Monitoring continued
justifiable: other social media platforms until 27 July 2020, encompassing the
are increasingly closed to research of this entirety of the state of emergency and the
type. Additionally, Twitter is a platform period immediately after the declaration
on which it is easy to engage directly and was lifted.
publicly with people outside of mutual
friend relationships. However, one drawback For this research we used a range of methods
of using Twitter data is the platform’s to attempt to infer coordinated inauthentic
relative unpopularity in Finland. According behaviour from observational data:
to the We Are Social #Digital2020 report
for Finland, 95% of the total population is 1. Bot detection—useful for finding
active on the internet. Based on data from possible inauthentic behaviour.
SimilarWeb, the three most-visited websites 2. Community detection using social
are Google, YouTube, and Facebook. There network analysis—useful for
are approximately 3.3 million active social identifying clusters of users with
media users in Finland, 773,000 of whom similar behavioural patterns.
can reportedly be reached with adverts 3. Narrative estimation—identifying
on Twitter (Kemp, 2020). According to the subjects of conversation.
Media Landscapes, the most popular 4. Abusive language detection—
social networks used weekly for news are identifying possibly harmful
Facebook (34%), YouTube (9%), Twitter (6%), material.
and Suomi24 (5%) (Jyrkiainen, 2017). 5. Timeline comparison—identifying
users with similar posting patterns.
We began monitoring Finnish activity on
Twitter on 12 March 2020, a few days
��������������������������������������������������������������������������� 17While none of these methods directly for our quarterly Robotrolling report, which
identify coordinated behaviour, our monitors English- and Russian-language
assumption was that taken together, they messaging about the NATO presence in the
would reveal pockets of similar activity, be Baltics and Poland. The algorithm works by
that in terms of account type, community calculating how similar behaviour of new
position, subject of messaging, or level of (unseen) users is to observed examples of
abusiveness. Such pockets of activity could automated activity. To make this calculation,
through qualitative analysis be verified it draws on the metadata provided by
as examples of coordinated inauthentic Twitter, and a range of calculated metrics.
behaviour. Should the analysis fail to These include, for instance, proportion of
identify clusters of inauthenticity, it might be retweets, frequency of posting, average
because they are uncommon in the Finnish- number of posts per day, standard deviation
language Twitter conversation, that the of message posting, and so on.
clusters of activity are very small (only two
or three accounts), or because our choice of
methods was inappropriate. Abusive language detection
When analysing abusive messaging in the
Data collection Finnish information space, the language
barrier poses a challenge. One possible
We collected messages on the social solution to this problem is to translate the
media platform Twitter mentioning one dataset to English via machine translation,
or more members of the Marin cabinet. and then apply mainstream models
Data collection was conducted using the optimised for the English language to the
Twitter Search API, which Twitter describes translated dataset. However, this approach
as focused on relevance rather than would result in the loss of vital information,
completeness. Consequently, one limitation as translation services may distort the
of this study is that some messages were original text and its meaning. Therefore, in
not captured during collection. The data order to preserve the original text of abusive
collection script was run periodically messages, we tried to avoid translation.
during the observation period. As data was
collected, we automatically anonymised all We decided to conduct hate speech analysis
personal identifiable information. in the original Finnish language. To do so,
we relied on the prepared dataset used in
After data collection we calculated the bot- Knuutila et al (2019). This dataset contained
likelihood for each account in the dataset. approximately 2,000 tweets identified by the
The algorithm used for this process was researchers as one of two classifications,
trained on data collected from 2017 to 2020 either abusive or neutral. Using this data,
18 ����������������������������������������������������������������������������we trained a neural network to recognise The Finnish language BERT model we
unseen examples of both categories. used was provided by TurkuNLP, a group
of researchers at the University of Turku
Bidirectional Encoder Representations (Virtanen et al, 2019). The training and
from Transformers (BERT), a concept classification tasks were carried out
pioneered by Google AI, is a Transformer- using the Python library ktrain (Maiya,
based machine learning technique for 2020) which supports model training
natural language processing (NLP). It learns and deployment from the Transformer
contextual relations between words in order framework—a Natural Language Processing
to create a model of a language. Unlike so- (NLP) for Pytorch and TensorFlow 2.0. The
called bag-of-words models where language model achieved excellent precision on the
is stripped of context and reduced to counts given training corpus even after only a
of signifiers, the BERT model includes few training epochs. We were concerned
contextual information to separate the that it might be prone to overfitting, and
meaning of otherwise identical words (e.g. unable to generalise to unseen data.
the word ‘model’ could refer to a person in However, after performance analysis and
a photoshoot, it could be a verb, it could training we assessed that the model was
be a noun referring to an abstract object, able to capture the tweets that contain
it could denote the make of a car, etc.). In abusive language with swear words
each of these cases, the word ‘model’ would allowing us to conduct the quantitative
be positioned differently within a vector analysis. Messages predominantly
space. This contextualised positioning in containing expletives yielded very high
vector space allows the comparison and abuse probabilities (0.97–0.99). Less
even generation of text. In the context of blunt examples of abuse were classified
abusive language, this allows the researcher within the 0.6–1.0 probability range. In
to determine that the phrase ‘he is a total order to capture the abusive portion of
failure’ is closer in meaning to ‘what a the dataset we decided to implement a
moron!’ than to ‘a failure of leadership’. We classification threshold of 0.6 to act as
use the model to assess whether novel a filter. Nonetheless, it is worth noting
sentences are in some way contextually that the model has a higher accuracy for
similar to the examples of abusive language classifying explicit abuse than for implicit
in the training data, even in cases when a abuse (see Table 1).
different vocabulary is chosen.
��������������������������������������������������������������������������� 19NexPlore or NView: A multilingual narrative explorer tool
For our analysis we wanted to understand whether abusive language was more common for
some topics than for others. To explore possible correlations between subject matter and
hate speech, we developed a bespoke exploratory tool that implements state-of-the-art NLP
text similarity search methods and algorithms. The narrative estimation tool incorporates the
similarity search project, FAISS developed by Facebook AI research lab (Johnson et al, 2017),
and ScaNN developed by Google scientists (Guo et al, 2020). The Facebook algorithm was
implemented in order to find similarities of each tweet to predefined topics. The algorithm
iterates through all individual tweets and computes the FAISS distance to all topics of interest.
As a result, one might use post-processing for clustering or simply to visualize the results.
Tweets found to be too distant can be extracted separately and used for additional NLP
analysis to estimate new narratives not considered before. We found this approach particularly
useful when graph analysis methods are used. Estimated distances of predefined narratives
serve as weights and features when creating complex graphs and analysing communities.
Another useful application is to monitor selected topics over time to follow the activity of users
engaging in those topics. When superimposed with a news timeline one might follow and
estimate the reactions of communities.
The Narrative estimation tool also offers Similarity search requires text translation
search-engine type functionality for the to vector space. Currently various text
complete dataset where a fast tweet encoders are available. One might use
search using ScaNN finds the top 100 best BERT sentence transformers (Reimers
matching tweets corresponding to and Gurevych, 2019) for English-language
rather abstract search queries instead of texts or even a language-agnostic model
searching for exact text matches. Speed is in the case of a multi-language dataset
achieved using efficient search algorithms (Yang and Feng, 2020). The rapid pace of
by incorporating item indexing similarly development of NLP models and methods
to the way it is done in databases. This will create increasingly powerful text
rapidly gives analysts insight into large analysis tools for analysts to use regardless
collections and allows them to quickly of which languages are involved. Future
estimate whether provisional narratives or implementations will incorporate additional
keywords of interest are within the dataset. functionality such as synthetic (generated)
text detection probabilities.The leadership in Finland is notably
young, female, and left leaning.
Case Study Background
Since December 2019, Finland has been monitoring period, eleven/twelve of the
governed by a centre-left coalition led nineteen Finnish ministers were women.
by Prime Minister Sanna Marin of the The leaders of the coalition also hold
Social Democrat Party. Marin’s election ministerial posts: Andersson serves as
drew international attention, as she Minister of Education, Ohisalo is Minister
simultaneously became the world’s of the Interior, Kulmuni served as Minister
youngest prime minister and the head of of Finance until her resignation on 5 June
a unique coalition in which all five party 2020, and Henriksson is Minister of Justice.
leaders are women. The four additional Marin’s political agenda prioritises climate
coalition party leaders, three of whom change, equality, and social welfare (Specia,
are in their 30s, are Li Andersson of the 2020).
Left Alliance, Maria Ohisalo of the Green
League, Annika Saariko of the Centre Party, Marin took over the premiership from
and Anna-Maja Henriksson of the Swedish Prime Minister Antti Rinne, who resigned
People’s Party of Finland. Throughout the in December 2019 after the Centre Party
months that we gathered data, Katri Kulmuni expressed that it had lost confidence in him
was initially serving as leader of the Centre over his controversial handling of a postal
Party but was replaced by Annika Saariko on workers’ strike. Although Rinne tendered his
5 September 2020. government’s resignation, President Sauli
Niinisto requested Rinne’s cabinet continue
The new leadership in Finland is notably on as a caretaker government, allowing the
young, female, and left-leaning. During our coalition to remain intact (YLE, 2019). The
��������������������������������������������������������������������������� 21coalition was originally formed following at the time consisted of the Finns Party,
parliamentary elections in April 2019, when the Centre Party, and the National Coalition
the Social Democrats narrowly defeated Party. Finnish state media YLE reported that
the right-wing populist Finns Party with just both Prime Minister Juha Sipilä and head
17.7% of the vote. The close victory resulted of the National Coalition Party Petteri Orpo
in a centre-left coalition and paved the way announced that their respective parties
for Rinne to become the first leftist Prime would no longer cooperate with a Finns
Minister in nearly 20 years (Reuters, 2019). Party led by Halla-aho due to his convictions
of hate speech for comments made about
The 2019 elections also highlighted the Islam and Somalis (YLE, 2017). Meanwhile,
increasingly fragmented nature of Finnish Finland’s traditional political parties
politics. The right-wing populist Finns struggled, as the Centre Party, conservative
party, which campaigned on a eurosceptic National Coalition Party, and left-leaning
and anti-immigration platform, fell a mere Social Democrats combined received just
6,800 votes short of winning first place. The 49% of support from voters. This was one of
party is led by Jussi Halla-aho, who holds the poorest election outcomes for the Social
controversial political views concerning Democrats, while the Centre Party polled
climate change and immigration policy. its lowest general election result ever (YLE,
Halla-aho’s election in 2017 caused tension 2019). As a result, the current coalition is
within the ruling three-party coalition, which navigating in a polarised political climate.
22 ����������������������������������������������������������������������������Descriptive Statistics:
Comparison with Robotrolling
The algorithm we used to assess bot-
likelihood for each account in the Finnish
dataset was trained on data that tracks
English- and Russian-language messaging
about the NATO presence in the Baltic
countries and Poland. Before we share the
results of our analysis, we will provide a 20000 40000 60000 80000 100000 120000
comparative overview of these two datasets.
During our observation period of the Finnish
Twitterspace, the majority of tweets were
Robotrolling
shared by human accounts (50%) and
anonymous accounts (45%), with just 3% of
messages originating from bot-like accounts
(bot, troll, hybrid). When we apply the abuse
filter, the amount of bot messaging remains
the same, while anonymous activity jumps
to 59% and human activity decreases to 35%.
Compared to Robotrolling figures for the 200 400 600 800 1000 1200
same time period, we can see that the Finnish
information space fosters roughly 1/5 of the
bot activity observed in both the Russian-
and English-language networks discussing
NATO in the Baltics and Poland. Our initial
assessment is that the Finnish online space is
a ‘cleaner’ space in which a greater number of
legitimate actors are operating online. 100 200 300 400 500
Number of accounts by type
human institutional anonymous
bot troll/hybrid news
��������������������������������������������������������������������������� 23Social Network Analysis
In order to understand communication patterns on Finnish Twitter, we mapped the connections
between users to form a network visualisation. The positioning of each user within the following
figures is relative to all the other users in our dataset. As a result, users who mention the same
domains, use the same hashtags, and retweet the same users tend to be grouped closer
together. Mapping the data in this way creates an approximation of the interests represented
within the dataset. For example, users who frequently discuss issues related to climate change
or education policy will tend to cluster.
The graphs are coloured to illustrate trends disproportionately responsible for such
in the data. This allows us to see whether messaging. Similarly, we map the calculated
accounts identified as sources of abuse likelihoods that each account is automated or
are randomly spread across the network, anonymous. While there are many legitimate
or whether particular communities are uses for automation on social media, and
animal cruelty
Figure 1: Social Network visualisation of the data collected. Figure A (left) shows users by language (pink
is Finnish), whereas figure B (right) shows the users by account type. Anonymous accounts are in yellow;
automated accounts in pink. Users are positioned close to accounts that share the same hashtags, retweets,
and links.
24 ����������������������������������������������������������������������������there are reasons why users may prefer to
remain anonymous, areas of the graph where
such users cluster may indicate coordinated
inauthentic activity. We then project other
connections onto this configuration, including
characteristics such as which users mention
each other, whether these mentions are
categorised as abusive, and what topics are
discussed.
In Figure 1a, users depicted in the
constellation to the left have been assigned a
colour to differentiate them by language. We
note that the majority of users discussing Figure 2: Finnish language accounts by type
Finnish politicians engage in Finnish (pink,
48%) and English (green, 29%). Figure 1b
shows the same network pattern but is cluster with a high proportion of anonymous
coloured according to account type—human accounts: one on the left and one on the
(blue), anonymous (yellow), and automated right. Both communities feature patches
(pink). This reveals that most of the of blue, reflecting the high proportion of
potentially suspicious activity is primarily anonymous account activity. The accounts
conducted in languages other than Finnish. on the right-hand side are centred around
Although examining foreign-language bot Jussi Halla-aho, the leader of the right-
activity is helpful for comparative purposes, wing populist Finns Party. The community
it is of little relevance for this analysis. For situated on the left side of the graph,
the remaining graphs, we are only including featuring a smaller blue area, is populated
Finnish-language users. by left-leaning Twitter users. Overall, the
data do not suggest widespread use of
automated accounts within the Finnish-
Finnish-language Twitter language conversation.
Figure 2 depicts the Finnish-language One minor exception to this assertion is a
conversation during our monitoring period. handful of users from the area on the right
The graph is coloured according to account whose messaging is regularly retweeted by
type, with pink representing human users, automated accounts.
blue—anonymous users, red—institutional
accounts, and black—automated users. We To visualise the proportion of abusive
identified two principal communities in this messaging, we overlay the calculation of
��������������������������������������������������������������������������� 25Figure 3: sources of abusive messaging
average abuse levels onto this network arbitrary and is shown only for illustrative
map. In the result, Figure 3, accounts that purposes—the borders between the
tend to engage in systematic abusive communities is to some degree random.
messaging are coloured red. As is The individual ‘bubbles’, which feature
demonstrated by the cloud of white nodes in several visualisations, vary in size
and edges, the vast majority of users do depending on the number of messages
not use abusive language when engaging sent. In Figure 4, we can see that the blue
with current government officials. The bulk and yellow communities broadly coincide
of abusive messaging is clustered in the with the left- and right-wing activists
lower right corner, among the online right- who exhibit disproportionate levels of
wing community. This finding is relatively anonymous and abusive activity.
unsurprising, as the Finns Party is currently
part of the opposition to Marin’s centre-left When we recreate the graph mapping
coalition government. A small collection by mentions, opposed to hashtags,
of abusive messaging was produced by domains, and retweets, we observe
the left-wing community, in large part in that the yellow and blue communities—
response to anti-government rhetoric. ideologically opposed users—exchange
a large volume of messages (Figure 5).
To illustrate the interactions between Restricting the connections between
communities, we split the graph into users to those sharing abuse allows us to
groupings calculated by a modularity see the accounts that receive the highest
algorithm. The number of communities is volume of abusive messaging—the Twitter
26 ����������������������������������������������������������������������������Figure 4: Community clusters
Figure 5: a) interactions between clusters, b) abusive messages between clusters
��������������������������������������������������������������������������� 27accounts of Prime Minister Sanna Marin, show a higher concentration of inter-
Minister of the Interior Maria Ohisalo, community engagement as well as a
and Minister of Education Li Andersson. correlation between community interest
and messaging about the topics. It is
This reveals an additional dynamic: the immediately clear that discussions about
targets of yellow community users do not government incompetence are riddled with
respond directly to the abuse. Instead, it abusive language, particularly coming from
appears that other users positioned close to the right-wing community, shown in yellow,
the targets in the map appear to respond on and left-leaning community, shown in blue.
their behalf, presumably in their defence. Levels of abuse are similarly high among
users engaging in sexist and homophobic
discourse. Discussions of Topic 3, Racism
Topics of abuse and Islamophobia, garnered similar levels
of activity from the yellow community and
We further analysed the data to understand reduced activity from the blue community.
which themes in Finnish politics attracted Abusive messaging about COVID-19, Topic
the highest levels of abuse from online 5, appeared to largely originate from the left-
users. We identified six prominent topic leaning community directed at right-wing
areas that accounts engaged with users. Topic 5 about education exhibited
throughout our observation period. These virtually no abusive messaging
topics are:
G
overnment Corruption and Was there coordination of inauthentic
Failure abusive messaging?
S
exism and Homophobia Around 7% of the messages shared on Finnish
Twitter during our monitoring period were
R
acism and Islamophobia identified as abusive and over 5,000 users
sent at least one abusive message. However,
G
overnment Handling of COVID-19 a handful of users shared high volumes of
such messages. For instance, during the
E ducation (in the context of 138-day observation window, one user sent
COVID-19) 520 messages, of which 199 were classified
as abusive. Education Minister Li Andersson
In the graphs on page 30, the same was the primary subject of the messaging,
network graph is coloured to show the receiving 87 abusive messages from this
proportion of abusive messaging about one user. Additionally, this user directed 46,
each topic. The darker areas of the graph 44, 35, and 34 messages at Interior Minister
28 ����������������������������������������������������������������������������Bad Government Sexism Racism Covid19 Education
500
400
300
200
100
2020-04-01 2020-05-01 2020-06-01 2020-07-01
Figure 6: total number of messages by theme
Ohisalo, Prime Minister Marin, Minister of are operated automatically, by the same
Family Affairs and Social Services Kiuru, person, or even in coordination with one
and former Minister of Finance Kulmini, another.
respectively.
Beyond the example of these four users, we
Like this user, the subsequent three most observed a tendency where the most abusive
prolific posters of abusive content averaged messages come from users who in their twitter
more than one such message daily. None activity are singularly focused on harassing
of these users were identified as part of a the government. These accounts may be
community in the social network analysis, fake—certainly the owners of the accounts
because they never retweeted other users, are not generally easily identified—but they do
shared links to news stories, or even not appear to be tightly, if at all, coordinated.
commented on specific hashtags. Instead, Thus, when it comes to abusive messaging of
more than 93% of their posts were directed this kind, the story told in the data is less about
at (@) other twitter users, in large part messages being posted from coordinated
government ministers. That said, there is accounts, but rather a stream of abusive
nothing about the posting patterns of these messages coming from a few accounts.
four hyperactive users that indicate they
��������������������������������������������������������������������������� 29You can also read
