Divided We Rule: Influencer Polarization on Twitter During Political Crises in India
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Divided We Rule: Influencer Polarization on Twitter During Political Crises in
India
Saloni Dash, 1 Dibyendu Mishra, 1 Gazal Shekhawat,1 Joyojeet Pal 1
1
Microsoft Research
Bengaluru, India
t-sadash@microsoft.com, v-dibmis@microsoft.com, t-gshekhawat@microsoft.com, joyojeet.pal@microsoft.com
arXiv:2105.08361v1 [cs.SI] 18 May 2021
Abstract handles of politicians from two major national parties in In-
dia – the national incumbent BJP (Bharatiya Janata Party)
Influencers are key to the nature and networks of information and the opposition party INC (Indian National Congress).
propagation on social media. Influencers are particularly im-
portant in political discourse through their engagement with
We do this by examining influencer engagement on four is-
issues, and may derive their legitimacy either solely or partly sues that have dominated the news and seen divisive political
through online operation, or have an offline sphere of exper- activism during 2019-2021. In the order of their onset, they
tise such as entertainers, journalists etc. To quantify influ- are as follows –
encers’ political engagement and polarity, we use Google’s • Abrogation of Article 370: The central government’s
Universal Sentence Encoder (USE) to encode the tweets of
unilateral abrogation of article 370 of the Indian Con-
6k influencers and 26k Indian politicians during political
crises in India. We then obtain aggregate vector represen- stitution, which had secured autonomy for India’s only
tations of the influencers based on their tweet embeddings, Muslim majority state, Jammu & Kashmir. The region,
which alongside retweet graphs help compute the stance and which is also part of a major international dispute, was
polarity of these influencers with respect to the political is- broken into two union territories, with no consultation
sues. We find that while on COVID-19 there is a confluence with the state’s residents. Hotly debated actions following
of influencers on the side of the government, on three other the abrogation included cutting of communication lines in
contentious issues around citizenship, Kashmir’s statehood, Kashmir, imprisoning politicians, imposing a curfew, and
and farmers’ protests, it is mainly government-aligned fan increasing armed forces on the ground.
accounts that amplify the incumbent’s positions. We propose • Citizenship Amendment Act (CAA): The Act, which
that this method offers insight into the political schisms in
disqualified Musims from qualifying for refugee status
present-day India, but also offers a means to study influencers
and polarization in other contexts. when entering from India’s neighboring countries, was
enacted alongside the government’s plans to conduct a na-
tionwide exercise to identify “illegal foreigners” living in
Introduction India by requiring historical documentation of residency.
As social media raises concerns of polarizing politics, Since such an exercise would automatically exclude peo-
worldwide, questions arise on the role of public figures ple of certain faiths from scrutiny and intensify attention
who wield the ability to influence the political discourse. In towards others, it stoked fears of lost citizenship among
parts of the west, there is an inherent tension between what India’s Muslims and led to widespread protests all around
are largely seen as liberal mainstream celebrities (Demaine the country.
2009) and an emerging group of hyper-partisan micro- • COVID-19: We studied the tweets surrounding the ongo-
celebrities (Lewis 2020). Meanwhile in India, the role of in- ing COVID-19 pandemic in India up until March 2021,
fluencers has been complex, especially among mainstream before the second wave of the crisis began. The first wave
offline celebrities who have leaned more towards the nation- period was marked by the government announcing a na-
alist incumbent, which has a much better organized online tionwide lockdown, which triggered an exodus of low-
social media presence (Lalani, Kommiya Mothilal, and Pal income migrant workers from Indian cities to hinterlands,
2019), while public figures speaking against the government often on foot, since transport lines were closed. The mi-
have faced online harassment, trolling, and attacks to their grant crisis and economic impacts, as well as the health
livelihoods (Singh 2020; Menon 2020). concerns and practices were part of the contentious issues
In this paper, we study the role of influencers in ampli- in these conversations.
fying political polarization in India, by quantifying the par- • Farmers’ Protests: The Indian Farm Bills which were
tisan engagement of Twitter handles of Indian entertainers, passed in parliament in September 2020, targeted pric-
sportspersons, journalists and other public figures alongside ing and subsidies and allowed the entry of corporates into
certain domains of crop trading. This led to protests by
Copyright © 2021, Association for the Advancement of Artificial farmers, particularly in states of North India where the
Intelligence (www.aaai.org). All rights reserved. new configurations of crops and purchase prices brought
threats to livelihoods. platform celebrities (those with no other known sources of
Our analysis builds on past work about Twitter-driven po- renown other than for their social media presence) or media
litical outreach strategies in India, particularly by the rul- houses rather than mainstream journalists, entertainers etc.
ing party BJP and its leader, Prime Minister Narendra Modi (refer to Data for the categories of influencers in this study).
(Pal 2015; Sinha 2017; Neyazi, Kumar, and Semetko 2016).
The four events we study here bear elements of communi- Related Work
cation and outreach both for and against the government’s
position, while three events are related to responses on spe- In their formative study of political polarization, Conover
cific government initiatives (Article 370, CAA, Farm Bills). et al. (2011) explain the homophily on social media through
The COVID-19 related discourse, while marked by a few partisan user behavior in the US. The authors particularly
key events, is an ongoing issue with longer term sustained find the retweet network of their sample to be highly po-
social media communication. larized, with limited interaction between the two camps. A
body of work also suggests that influencers exacerbate po-
larization on social media by endorsing and spreading parti-
san content. In an experiment with Democrats and Republi-
cans, Becker, Porter, and Centola (2019) demonstrated that
polarization does not spread in egalitarian networks where
all individuals have equal influence throughout the network.
They argue that partisan bias is amplified on social media
platforms like Twitter because these networks are organised
around few key influencers (Centola 2020) and the ability
to obtain central positions could allow extremists to gener-
ate and sustain belief polarization in social media networks
(Becker, Porter, and Centola 2019). Additionally, Garibay
et al. (2019) found through their simulations that polariza-
tion in social media helps influencers gain more influence.
Moreover, studying the elections in the US and Spain,
Figure 1: Influencers’ Timeline of Retweeting BJP (Top) Barberá and Rivero (2015) find that ideologically motivated
and INC (Bottom) Politicians. Twitter users create disproportionately more content than
moderate ones. In India, studies of Modi’s political cam-
Figure 1 shows us a weekly timeline (January 2019 - paigning have particularly noted the use of celebrities on
March 2021) of influencers retweeting politicians, with the Twitter for image building (Pal 2015). By interacting with
peaks corresponding to key events in India. We see that key influencers, Modi was able to successfully attach his
the four events studied here all see peaks, and that in gen- brand with that of leading entertainers and sportspersons.
eral, influencer tweets on both sides tend to spike around Neyazi, Kumar, and Semetko (2016) find Twitter networks
the same events with the exception of a few events, such as in India to be highly polarizing, especially in BJP’s political
”Howdy Trump” when the ruling party mobilized more in- campaigning. This polarization also affects traditional com-
fluencers, while influencers on the opposition side mostly ig- munication platforms due to the increasingly hybrid nature
nored it. Table 1 offers some insight into influencers retweet- of media systems with digital platforms acting in conjunc-
tion with traditional media in politicians’ campaigns (Chad-
Retweet Range wick 2017).
Party Total
[1,50] (50,100] (100,150] (150,max] In their study of the Brazillian presidential election that
BJP 3265 214 99 272 (max-10053) 3850 brought far-right politician Jair Bolsonaro to power, Re-
INC 3144 185 62 151 (max-54301) 3542
cuero, Soares, and Gruzd (2020) find a strong connection
Table 1: Distribution of Influencers Retweeting BJP and between polarization and disinformation spread on Twitter.
INC Politicians. This degree of polarization also has parallels in offline de-
velopments of the region. Here, the authors observed that the
pro-Bolsonaro camp was asymmetrically polarized and note
ing politicians. We see that most individual influencers have that the escalation of ‘hyperpartisanship’ also led to the pe-
generally engaged with politicians in small measure (1-50 ripheralization of traditional media. In such scenarios, par-
times). The major difference we see is among the influ- tisan media houses, that defy traditional notions of balance
encers who have engaged with politicians’ content more and detachment from the contested subject, gain prominence
than 150 times. We see a larger number of influencers en- (Faris et al. 2017). The degree of polarization also varies on
gaging with BJP – 272, compared with 151 influencers who the nature of the crises, and politics is especially distinct in
have retweeted INC politicians, though the overall maxi- its contribution (Barberá and Rivero 2015).
mum engagement is of an influencer retweeting INC’s con-
tent 54k times. This concentration of influencers on the high
end underlines the importance of a loyal set of accounts that Method
actively promote a party’s political content. We also find that We now outline our data collection and polarity score com-
the influencers who most engage with politicians tend to be putation methodology.
Data prominent members of nonprofits, affiliates of religious or
We use a publicly available dataset (Panda et al. 2020) of quasi-religious organizations who present themselves as ac-
Indian politicians and their twitter handles. The dataset also tivists, or any who self-identify with the label. The Business
has party annotations for each of the politicians, which we category includes accounts of commercial establishments,
use to obtain the accounts of 14,094 BJP politicians and their founders, and key employees. The Entertainment seg-
12,341 INC politicians. We then collect tweets 1 from these ment refers to people engaged in films, television, music,
26,435 accounts between January 2019 - March 2021. and fashion. Fan Accounts have to do with profiles that
To build the set of influencers we iteratively collect the distinctly avow to support public figures or organizations,
friends of the filtered politicians. From the resultant list, and includes unofficial accounts dedicated to actors, politi-
we remove all the politicians, non-Indian global figures cians and political parties. Journalist refers to individuals
etc., such that we are left with Twitter accounts that are who work in the reporting and production of news, features
highly followed by Indian politicians. Our process being and commentary. The Law & Policy category is for lawyers,
built ground up from the friends of politicians means that judges, and public policy practitioners. Media Houses are
our definition of influencer biases towards those accounts platforms, both digital and conventional that produce news,
that are of interest to politicians - ie for instance, if our seed entertainment or other content, and thus this category also
set had started with sportspersons, it is likely we would have includes newspapers, websites, and TV channels. Like Fan
ended up with several highly followed sports management Accounts, the Platform Celebrity category captures a digital
companies or commentators in our set, or likewise for other phenomenon of accounts whose popularity stems from their
domains. In essence, a journalist with 3000 followers, of online presence and content creation, rather than the offline
which 200 followers are politicians would be included in work of the individuals involved. The Social Work category
our set, as opposed to say a PR agent with 300k followers, refers to the non-profit sector and includes accounts of its
but with few politicians following them. employees. Sports captures both athletes, as well as the ac-
The iterative process helps mitigate this bias, and our counts of teams, tournaments, and governing bodies. Writer
team’s contextual knowledge of India is used to ensure that is a category for published authors of fiction and non-fiction,
known public figures are included in the sample as far as it does not include columnists who are employed by news
possible. Thus, we are confident that the vast majority of ’A platforms.
List’ public figures such as film stars, sportspersons, busi-
nesspersons etc. are included in our sample, but as Table 2 Tweet Classification
shows, we capture a relatively high number of journalists In order to capture event specific tweets, we use a Word2Vec
and media houses, since their work is of obvious importance (Mikolov et al. 2013) bag of words based technique to clas-
to politicians. We categorized every influencer account into sify the collected tweets (Vijayaraghavan, Vosoughi, and
only one category, whichever they are primarily known by, Roy 2016). We first define a set of high precision keywords
thus a major sportsperson with business interests is nonethe- that are indicative of the event. We then filter tweets that
less categorized under sports. contain at least one of the keywords, and consequently train
a Word2Vec model on the filtered tweets. After obtaining a
Category Count vector representation of all the words, we expand each of the
Academia 172 keywords from the seed set, by computing the cosine simi-
Activist 81 larity of pairs of words for all words in the vocabulary. We
Business 208 then iteratively add words to the seed set according to the
Entertainment 1251 cosine similarity based criteria defined in Vijayaraghavan,
Fan Account 36 Vosoughi, and Roy (2016). Some of the keywords identified
Journalist 3551 by the Word2Vec model are showed in Table 3.
Law & Policy 141
Media House 550 Keywords & Hashtags
Platform Celebrity 126 Article 370 #jammuandkashmir, #article370, 35a, abro-
Social Work 63 gation, kashmiri
Sports 348 CAA/NRC #caanrc, #citizenshipamendmentact,
#caanrcnpr, #anticaa, #caanrcprotest
Writer 99
COVID-19 covid2019, #covidpandemic, #coronapan-
Total 6626 demic, #coronacrisis
Farmers’ Protest #farmersprotests, #farmersagitation, #del-
Table 2: Influencers by Category hichalo, #farmersdelhiprotest
Table 3: Top 5 Keywords/Hashtags by Event.
The final set of influencers is divided into a 12 cate-
gories. Academia refers to accounts that largely work with
higher education institutes and conduct research. The Ac- In order to obtain a rough estimate of the performance of
tivist category includes grassroots organizers, leaders or our classification pipeline, we randomly select a represen-
tative sample (n = 100) and manually annotate whether the
1
Using Tweepy - https://www.tweepy.org/ tweet is related to the particular event or not. We report pre-
cision values of 87.0%, 98.0%, 100% and 97% for Article Content Polarity We adopt the cultural axis generation
370, CAA/NRC, COVID-19 and Farmer’s Protests respec- procedure described by Waller and Anderson (2020). We
tively. Overall, we obtain a combined precision of 95.6%. compute a partisan axis in the high dimensional user embed-
While these precision values are noisy estimates of the true ding space, by taking the difference of vector embeddings of
precision values, it provides a sanity check for the classifi- the top n (=10) INC and BJP politicians, sorted by retweet
cation methodology. polarity scores. Briefly, let x̃ and ỹ be the high dimension
embedding vectors representing INC and BJP politicians re-
User Embeddings spectively. For any user u, with a high dimension embedding
vector ũ, the content polarity score c(u) is defined as the co-
Given a set of accounts and their tweets for an event, we
sine similarity of the user with the partisan axis:
obtain user embeddings based on aggregate tweet embed-
dings (Rashed et al. 2020). We use Google’s Universal Sen- ũ · (x̃ − ỹ)
tence Encoder (USE) (Cer et al. 2018) to obtain high di- c(u) = ∈ (−1, 1) (2)
kũk kx̃ − ỹk
mensional vector representations of the tweets by a user for
a given event, and average out all the tweet embeddings to where a score close to 1 indicates a user’s stance on the par-
represent a single user. The user embeddings are projected ticular issue is highly pro-INC, whereas a score close to -1
to a two dimensional space using UMAP (McInnes, Healy, signifies that the user’s stance on the issue is highly pro-BJP.
and Melville 2020) and clustered using hierarchical density
based clustering (HDB-SCAN) (McInnes and Healy 2017). Experiments
According to Rashed et al. (2020), this method results in We capture event specific tweets based on the classifica-
clusters indicative of users’ stance on the specific event. tion pipeline discussed previously. The number of tweets for
User Polarity Topic Users Tweets
Article 370 14,050 346,197
We then use the clusters to compute polarity scores of users. CAA/NRC 13,744 527,702
Let V be the set of all users who have tweeted about a partic- COVID-19 20,871 3,069,851
ular event. We partition V into two groups, X and Y based Farmers’ Protest 12,224 269,742
on the clusters returned by HDB-SCAN. We partition V in
a way such that partitions X and Y predominantly consist Table 4: Users and Tweets by Event
of pro-INC users and pro-BJP users respectively.
each event, including the number of users who have tweeted
Retweet Polarity We construct an undirected retweet
about that particular event are enumerated in Table 4.
graph G with users in V as vertices. An edge eij exists be-
tween users vi and vj ∈ V, if vi and vj have at least two Stance Detection With Polarized Clusters
retweets between them, i.e. if either vi has retweeted vj at
least twice, or vi has retweeted vj at least once and vice To compute the stance of user for a given event, we use the
versa or vj has retweeted vi at least twice, ∀i, j = 1...|V|. user’s tweets related to that event to obtain user embeddings.
We choose 2 retweets as our threshold based on the experi- We do this for all the users who have tweeted about the
ments by Garimella et al. (2018) to avoid unreliable results. event, and then cluster their embeddings using HDB-SCAN.
We then use the Random Walk Controversy (RWC) score As shown in Figure 2, the users are clustered according to
(Garimella et al. 2018) to define a retweet based user po- their stance. We see two clusters across all the events, one
larity score. for each of the stances (pro-INC and pro-BJP). We see that
Let X∗ and Y∗ be the set of top k (=100) high degree INC and BJP politicians lie in separate clusters across all the
vertices, i.e. the top k users with the most number of retweet events, which reaffirms the polarizing nature of the events.
edges, from partitions X and Y respectively. Let luX as de- Moreover, we observe that majority of the influencers (INF)
fined in Garimella et al. (2018) be the expected hitting time also lie in either of the two clusters, confirming our hypoth-
for a vertex u ∈ V, i.e. the expected number of steps for esis that the influencers engage with these issues in a par-
a random walk starting with vertex u and ending in a high tisan manner. We observe that for all of the events except
degree vertex in X∗ . Therefore p(u)X is defined as the frac- COVID-19, majority of the influencers are in the pro-INC
tion of vertices v ∈ V for which the expected hitting time is cluster, whereas for COVID-19 majority of the influencers
more than that of vertex u, i.e. luX < lvX . Therefore a score of are in the pro-BJP cluster.
r(u)X ≈ 1 indicates that the vertex u is close to a high de- To understand this pattern further, we manually studied
gree vertex in X∗ . Subsequently, r(u)Y is defined similarly, the individual tweets. The overall clustering of influencers
leading to retweet polarity score r(u) being defined as: towards the INC is explained by the engagement of jour-
nalistic content by the opposition party, and the reliance of
r(u) = r(u)X − r(u)Y ∈ (−1, 1) (1) the ruling party on relatively more influential, but less nu-
merous platform celebrities, or mainstream celebrities. Thus
where a score close to 1 indicates that a user is more likely while the overall number of influencers seem to cluster much
to be retweeted by another user with a pro-INC stance, while higher on the INC side, this finding needs to be seen along-
a score close to -1 indicates that the user is more likely to be side the retweet polarity scores (Figure 8) which help un-
retweeted by a user with a pro-BJP stance. derstand the size and scope of influencer engagement. Weobserved also that the relatively larger clustering on the INC Quantifying Polarization Using Clusters
side is a result its politicians’ tendency to engage with influ- We further use the clusters to quantify two aspects of polar-
encers on account of their criticism of the BJP – this does not ization – retweet polarity which captures which “side” of the
mean those individuals are politically aligned with the INC, issue is more likely to endorses the influencer, and content
simply that they are critical of the ruling party. The BJP’s polarity which signifies the extent to which the influencers’
engagements and the resultant clusters, on the other hand, stance resembles that of either parties on a particular issue.
are closest to accounts who advocate for the party. To compute retweet polarity scores, we first construct the
These patterns speak to the larger state of politics in India, retweet graph for all the events in Figure 4. Each node repre-
where political opposition on a national level is weak, and sents users who are either politicians or influencers and the
civil society members and journalists, rather than the INC, edges represent retweets between them. The graph is color
have emerged as popular critics of the government. Our sec- partitioned as INC in blue, BJP in orange and influencers in
tion on categorical analysis further explains the distinctive- pink. The graph is visualised using the Force Atlas 2 layout
ness of the COVID-19 clustering. in Gephi (Bastian, Heymann, and Jacomy 2009), where we
set gravity rules for node attraction using their edge strength
to further clarify the clusters. The size of the nodes indi-
cates their degree centrality. Overall, similar to the clusters
above, we observe that for all events with the exception of
COVID-19, most influencers engage with INC and a small
fraction of them engage with BJP. Whereas, in the case of
COVID-19, maximum engagement of influencers is with the
incumbent party, BJP. While top leaders from either parties
including Members of Parliament and Members of Legisla-
(a) Article 370 (b) CAA/NRC tive Assemblies have high degree centrality in all events, IT
cell presidents wield significant influence as well. Amongst
the influencers, we see journalists, fan accounts and media
houses comprising some of the larger nodes.
Based on the clusters and retweet graphs from above, we
calculate retweet and content polarity scores for all the influ-
encers for a particular event. The distributions of the polarity
scores in Figure 5 show that a number of influencers highly
side with the narratives of a certain party for each topic. For
COVID19 the median scores for influencers is closer to the
(c) COVID-19 (d) Farmers’ Protests median of BJP politicians, while for Article 370, CAA/NRC
and Farmers’ Protest it is closer to the median of INC politi-
Figure 2: Clusters by Event. cians.
Figure 3 shows the wordclouds, based on prominence Analysis of Polarity Scores
scores (Rashed et al. 2020), for each of the stances and helps
underscore the differences. On Article 370 and CAA/NRC, In this section we use the computed polarity scores to gain
the INC side focuses on the act being “undemocratic” while insights into the partisan engagement of influencers across
the pro-BJP tweets celebrate them as “historic”, having cit- all four events. We first explore whether the polarity scores
izen support, and highlighting Hindu or refugee-related is- display meaningful patterns with respect to the number of
sues. In essence, we see here shades of populism, where followers and median retweet rates for an influencer. We
one side attempts to speak of the issue in legalistic terms, then show how the retweet and content polarity scores vary
while another focuses on the partisan appeal of the act. On according to the area of expertise of the influencers across
COVID-19, a less divisive issue (at the time), INC sympa- all the events.
thizing influencers engage on “unemployment” and “fail-
ure”, while on the side of the ruling party, the engagements Aggregate Trends
are more utilitarian, focusing on the “fight” against the Coro- Here we use the absolute values of the polarity scores be-
navirus. With the Farm Bills, we see a different form of pop- cause we are interested in how the magnitude of the scores
ulist engagement, in which the INC side uses terms such as vary with attributes like number of followers etc. We assign
“draconian”, and “anti-farmer” while the BJP side highlights an aggregate polarity score to an influencer by taking the
farmers as heroic using brand terms like “aatmanirbhar” median of the absolute polarity scores across all events.
(self-reliant), “hardworking” etc. Unlike with CAA/NRC By Followers Count To analyse the polarity scores by
and 370, where there is a clear nativist appeal in creating the number of followers, we categorize the influencers by
an “other” in those that oppose the act, the discourse with their number of followers. We categorize an influencer as
the farm bill is a bit different, since the ruling party would Very Low if the number of followers lies in the first quar-
rather not demonize farmers, and must toe a careful line in tile of the followers count distribution across all influencers,
countering opposition online. Low if it lies in the second quartile, Medium if it lies in the(a) Article 370 (b) CAA/NRC
(c) COVID-19 (d) Farmers’ Protests
Figure 3: WordClouds of BJP (Right, in Orange) and INC (Left, in Green) Clusters by Event
third quartile and High in the fourth quartile. We then plot We also plot the median of content polarity scores for each
the median retweet polarity scores for each of the categories. category, where we observe that Very Low and Low influ-
Overall, we see that the retweet polarity score increases with encers have the highest content polarity. A Welsch’s t-test
the number of followers. However, we see a sharp decrease between Low and Very Low gives a p-value > 0.1 indicating
in the polarity scores between Medium influencers and High that the two categories are not statistically different, whereas
influencers. A Welsch’s t-test between the scores of these t-tests between the scores of Low and Medium and Low and
two categories results in a p-value of 0.0684 which is a sta- High give p-values of < 0.0001 indicating that the scores for
tistically significant difference with a threshold of 0.1. The Very Low and Low are statistically different from Medium
p-values for scores between Very Low, Medium and Low, and High.
Medium are 0.0000 and 0.0081 respectively indicating that Overall we observe that there is a twilight zone for in-
the distribution of retweet polarity scores for Medium influ- fluencer polarization with respect to number of followers,
encers is statistically different from all the other categories. wherein influencers with a certain range of followers are
significantly more polarizing than other groups. In terms of
retweet polarity, which quantifies how likely an influencer’s
tweet is to be retweeted by either side, influencers with
Medium followers have the highest polarity scores. Whereas
in terms of content polarity, which signifies how extreme an
influencer’s stance is, those with Very Low and Low follow-
ers seem to have the highest scores.
Moreover, we find that from 20 of the influencers with
the highest number of followers - mostly entertainers and
sportspersons - only two were retweeted by the INC. This
(a) Article 370 (b) CAA/NRC confirms suggestions in past work that the ruling party has
a much stronger grip on the social media output of A-list
celebrities (Lalani, Kommiya Mothilal, and Pal 2019).
By Retweet Count We fit a regression model with retweet
polarity and content polarity as the independent variables
and the logarithm of median retweet count as the dependent
variable. Regression results are displayed in Table 5. Over-
all, we observe that the median retweet rate is higher for
influencers with higher retweet polarity scores, potentially
implying that polarized influencers are retweeted more in
(c) COVID-19 (d) Farmers’ Protests comparison to less polarizing ones. To explore this further,
we filter out influencers whose retweet polarity scores lie in
Figure 4: Retweet Graph by Events. the fourth quartile. We then calculate the median retweet rate(a) (b) (c) (d)
(e) (f) (g) (h)
Figure 5: Distribution of Polarity Scores by Event. Subfigures (a) - (d) depict retweet polarity scores and (e) - (h) depict
content polarity scores for Article 370, CAA/NRC, COVID-19 and Farmers’ Protests respectively
higher retweets as compared to other tweets.
log rt count R2 pvalue
rt polarity 2.1450 (0.157) 0.196 0.1
Table 5: Regression Analysis. Log of Median Retweet
Count (log rt count) is the dependent variable. The regres-
sion coefficient along with the standard error, R2 values and
p-values are displayed. Statistically significant results are
highlighted in bold.
Figure 6: Average Retweet Polarity Scores By Followers
Category. We observe that influencers with Medium number
of followers have the highest median retweet polarity score. Categorical Analysis
Retweet Polarity For each of the four events, we show in
Figure 8 how the retweet polarity scores vary with the area
of expertise of influencers. A close examination of the top
20 polarized accounts towards either side in each category
reveals that, barring COVID-19, the INC side overwhelm-
ingly retweets journalists on the three political issues. For
instance, in the activity after the abrogation of 370, 18 of the
top 20 accounts with polarity towards the party are journal-
ists. The BJP’s most proximate accounts have varied profiles
in comparison, including platform celebrities, businesses,
fan accounts, sports, and journalists. Unlike other issues in
which voices on the ground of journalists, academics etc
Figure 7: Average Content Polarity Scores By Followers are key to the political discourse, on the 370 issue, there
Category. Influencers with Very Low and Low number of is no representation of Muslim Kashmiris from the Valley
followers have the highest content polarity and those with on-ground among top accounts on either side. From the top
High number of followers have the lowest content polarity. pro-INC accounts, only one is from Kashmir. Among the
pro-BJP ones, the discussions forwarded include three Kash-
miri Pandits, a Hindu community whose exodus from the
for their polarizing tweets and median retweet rate of the rest valley in the early 1990s has been cited as a justification
of their tweets. We observe that 84% of the influencers have for the need to ‘integrate’ the region within the nation-state.
a higher median retweet rate for tweets related to polarizing The absence of on-ground accounts can partly be attributed
events as compared to their other tweets, potentially imply- to the shutdown of internet services and jailing of Kashmiri
ing that polarized tweets are more likely to be rewarded with politicians, but nonetheless points to the lack of representa-tive voices in the discourse. low scores, the top accounts in terms of content polarity, are Except journalists and activists, other categories of influ- journalists and academics critical of the move. That the high encers are less likely to take stances critical of the govern- score on the government’s side is driven by fan accounts and ment. We speculate that certain professions are more hesi- platform celebrities calls to question the role of hyperparti- tant to be critical and are thus disproportionately retweeted san voices, who may not have proven credentials of on the by the BJP side. Here, de-politicized fields such as sports, ground knowledge of the topic. businesses and entertainment have also shown tilts towards CAA/NRC protests score high on content polarity in com- the ruling government. The entertainment category polar- parison. Of the top 20 polarizing accounts, an overwhelm- izes towards INC only during the CAA protests, which is ing proportion has to do with journalism, including reporters in step with the on-ground developments in the movement and media houses, largely those critical of the government. when film industry figures spoke out against the Modi gov- We see the presence of Indian public service broadcasters ernment’s for the first time in significant numbers. (e.g @DDNational, @DD Bharati, @PBNS India), which The lower ratio of journalists in the BJP’s top polarity in principle, are meant remain autonomous and neutral. Con- scorers is an interesting trend that has connections to the na- tent from these accounts have relayed the official line and are ture of Modi’s political campaign. As Chakravartty and Roy thus engaged significantly by BJP politicians. (2015) note, in Modi’s populist rhetoric, liberal media has COVID-19 has had the most mixed reaction between the remained as a prominent antagonistic signifier, ever since categories in terms of the content, possibly because on the the press covered his role in the 2002 Gujarat riots where whole it is a less polarizing event. The relatively polarized Muslims died in disproportionate numbers. This aversion accounts here are often explicitly aligned with one of an- also translates to the lack of press conferences and inter- other party. views given by the leader, and as our study shows, perme- On the farmers’ protests, the top 20 accounts are di- ates to the ruling party’s social media strategy after coming vided between supporters and opponents of the government to power. Of the top media-related accounts the party cen- with a cross-category representation. Here too, government- tres, many are linked to partisan media houses, especially aligned accounts, in the form of public broadcasters and in- online portals that have gained prominence as vocal defen- dustrial consortia are prominent. Somewhat exceptionally, dants of right-wing ideology. Our finding is relevant with the businesses, who protesters see as the real beneficiaries of the Recuero, Soares, and Gruzd (2020) study that observed the policy, have a high polarity score on this issue. Our analysis distancing of hyperpartisan Twitter clusters in Brazil from of tweets from the highest-scoring accounts in this category mainstream media. With INC, we observe that the top ac- shows how business houses, while avoiding explicit politi- counts are not as much supportive of the party as they are cal keywords, have used a mediated, corporate vocabulary to critical of the ruling government. convey the benefits of the farm bills, which intersects with The dissimilar polarity scores on the COVID-19 crisis a lot of the official language on the issue. Sports-persons also indicate a unique set of responses it set off among Twit- have also only had higher polarity scores on this topic. En- ter influencers. We attribute the shift in polarity scores to- tertainers have scored mildly on all issues except for the wards the BJP, to the change in the type of the crisis, as pan- CAA. Note that we use the absolute values of content polar- demic foremost raised concerns about public health and pre- ity scores for this categorical analysis, as the median signed paredness, which is a bipartisan issue. Here it is likely, that scores for each category did not yield interpretable results in rather than being critical actors that invite accusations of be- the experiments. ing dispiriting, influencers were engaged in raising morale as part of government’s campaigning and disseminating front- Discussion line information and resources. We also see journalists and In this work, we present an important, non-western case of media houses disproportionately amplified by the BJP, indi- partisan engagement in a polarized political environment. cating a trend of relative cooperation with the government in While our findings are immediately of relevance in under- the face of a global health emergency. standing the political economy of social media in India, they Content Polarity A category-wide content polarity anal- are likewise relevant to approaching the role of influencers ysis in Figure 9 for the four events produces distinct results. of various kinds in political engagement. Of the four issues, we see that Article 370 draws some of Our findings show that the differences in clustering the most uniform scores between categories, that are lower around topics are themselves insightful in understanding the on content polarity. Kashmir as a territory is considered larger patterns in the discourse. Focusing our examination central to India’s narratives of national security and territo- around key events, helps understand the periods of elevated rial claims. Historically, conversations about the region are activity, i.e. those in which political parties may resort to en- sensitive and regulated, attracting legal prosecution (Bose gaging the various outreach resources at their behest, includ- 2009). This context seems relevant in understanding these ing celebrities. In this work, we have shown that the ruling scores. The overall lower scores can imply multiple strands party has unique advantages over the opposition in some as- at play. While political voices opposed to the national gov- pects, but that the opposition gains tailwind from the work ernment’s move were censored due to the communications of journalists and commentators, who on the whole, tend to shutdown, or deliberately ignored by the mainstream par- dissent with the government. We also find that there is sig- ties as taboo, what we see here also raises questions about nificant difference between the four events, and the kinds of a chilling effect of self-censorship. Within these relatively influencers that can engage in each of those.
(a) Article 370 (b) CAA/NRC
(c) COVID-19 (d) Farmers’ Protests
Figure 8: Median Retweet Polarity Scores by Influencer Category. Length of stem indicates retweet polarity score and
bubble size indicates median retweet count.
(a) Article 370 (b) CAA/NRC
(c) COVID-19 (d) Farmers’ Protests
Figure 9: Median Content Polarity Scores by Influencer Category.Length of stem indicates absolute content polarity score
and bubble size indicates median retweet count.These findings lay the ground for future work. There is a the 2019 on Computer Supported Cooperative Work and So-
need to study various kinds of influencers, their relationship cial Computing, 267–271.
with the government, and what that does to their online be- Lewis, R. 2020. “This is what the news Won’t show you”:
havior. While we studied the intersection of influencer and YouTube creators and the reactionary politics of micro-
politician behavior during crises, it is equally important to celebrity. Television & New Media 21(2): 201–217.
extend such work into its typical daily patterns.
McInnes, L.; and Healy, J. 2017. Accelerated hierarchi-
References cal density based clustering. In 2017 IEEE International
Conference on Data Mining Workshops (ICDMW), 33–42.
Barberá, P.; and Rivero, G. 2015. Understanding the politi- IEEE.
cal representativeness of Twitter users. Social Science Com-
puter Review 33(6): 712–729. McInnes, L.; Healy, J.; and Melville, J. 2020. UMAP: Uni-
form Manifold Approximation and Projection for Dimen-
Bastian, M.; Heymann, S.; and Jacomy, M. 2009. Gephi: An sion Reduction.
Open Source Software for Exploring and Manipulating Net-
works. URL http://www.aaai.org/ocs/index.php/ICWSM/ Menon, N. 2020. Hindu Rashtra and Bollywood: A New
09/paper/view/154. Front in the Battle for Cultural Hegemony. South Asia Mul-
tidisciplinary Academic Journal (24/25).
Becker, J.; Porter, E.; and Centola, D. 2019. The wisdom
of partisan crowds. Proceedings of the National Academy of Mikolov, T.; Chen, K.; Corrado, G.; and Dean, J. 2013. Ef-
Sciences 116(22): 10717–10722. ficient estimation of word representations in vector space.
arXiv preprint arXiv:1301.3781 .
Bose, S. 2009. Kashmir: Roots of conflict, paths to peace.
Harvard University Press. Neyazi, T. A.; Kumar, A.; and Semetko, H. A. 2016. Cam-
paigns, digital media, and mobilization in India. The Inter-
Centola, D. 2020. Why Social Media Makes national Journal of Press/Politics 21(3): 398–416.
Us More Polarized and How to Fix It. URL
https://www.scientificamerican.com/article/why-social- Pal, J. 2015. Banalities turned viral: Narendra Modi and the
media-makes-us-more-polarized-and-how-to-fix-it/. political tweet. Television & New Media 16(4): 378–387.
Cer, D.; Yang, Y.; Kong, S.-y.; Hua, N.; Limtiaco, N.; John, Panda, A.; Gonawela, A.; Acharyya, S.; Mishra, D.; Moha-
R. S.; Constant, N.; Guajardo-Céspedes, M.; Yuan, S.; Tar, patra, M.; Chandrasekaran, R.; and Pal, J. 2020. NivaDuck -
C.; et al. 2018. Universal sentence encoder. arXiv preprint A Scalable Pipeline to Build a Database of Political Twitter
arXiv:1803.11175 . Handles for India and the United States. In International
Conference on Social Media and Society, SMSociety’20,
Chadwick, A. 2017. The hybrid media system: Politics and 200–209. New York, NY, USA: Association for Computing
power. Oxford University Press. Machinery. ISBN 9781450376884. doi:10.1145/3400806.
Chakravartty, P.; and Roy, S. 2015. Mr. Modi goes to Delhi: 3400830. URL https://doi.org/10.1145/3400806.3400830.
Mediated populism and the 2014 Indian elections. Television Rashed, A.; Kutlu, M.; Darwish, K.; Elsayed, T.; and
& New Media 16(4): 311–322. Bayrak, C. 2020. Embeddings-Based Clustering for Target
Conover, M.; Ratkiewicz, J.; Francisco, M.; Gonçalves, B.; Specific Stances: The Case of a Polarized Turkey. arXiv
Menczer, F.; and Flammini, A. 2011. Political polarization preprint arXiv:2005.09649 .
on twitter. In Proceedings of the International AAAI Con- Recuero, R.; Soares, F. B.; and Gruzd, A. 2020. Hyperparti-
ference on Web and Social Media, volume 5. sanship, disinformation and political conversations on Twit-
Demaine, L. J. 2009. Navigating policy by the stars: The ter: The Brazilian presidential election of 2018. In Proceed-
influence of celebrity entertainers on federal lawmaking. JL ings of the international AAAI conference on Web and social
& Pol. 25: 83. media, volume 14, 569–578.
Faris, R.; Roberts, H.; Etling, B.; Bourassa, N.; Zuckerman, Singh, R. 2020. The Politics Of Offense: The Phenomena Of
E.; and Benkler, Y. 2017. Partisanship, propaganda, and Censorship In The Contemporary Discourse Of Literature
disinformation: Online media and the 2016 US presidential And Media. Literary Herald 6(3): 145–153.
election. Berkman Klein Center Research Publication 6. Sinha, S. 2017. Fragile hegemony: Modi, social media and
Garibay, I.; Mantzaris, A. V.; Rajabi, A.; and Taylor, C. E. competitive electoral populism in India. International Jour-
2019. Polarization in social media assists influencers to be- nal of Communication 11(2017): 4158–4180.
come more influential: analysis and two inoculation strate- Vijayaraghavan, P.; Vosoughi, S.; and Roy, D. 2016. Au-
gies. Scientific reports 9(1): 1–9. tomatic detection and categorization of election-related
Garimella, K.; Morales, G. D. F.; Gionis, A.; and Math- tweets. In Proceedings of the International AAAI Confer-
ioudakis, M. 2018. Quantifying controversy on social me- ence on Web and Social Media, volume 10.
dia. ACM Transactions on Social Computing 1: 1–27. Waller, I.; and Anderson, A. 2020. Community embeddings
Lalani, F. M.; Kommiya Mothilal, R.; and Pal, J. 2019. The reveal large-scale cultural organization of online platforms.
Appeal of Influencers to the Social Media Outreach of In- arXiv preprint arXiv:2010.00590 .
dian Politicians. In Conference Companion Publication ofYou can also read
