=Paper=
{{Paper
|id=Vol-2481/paper21
|storemode=property
|title=Is “manovra” Really “del popolo”? Linguistic Insights into Twitter Reactions to the Annual Italian Budget Law
|pdfUrl=https://ceur-ws.org/Vol-2481/paper21.pdf
|volume=Vol-2481
|authors=Claudia Roberta Combei
|dblpUrl=https://dblp.org/rec/conf/clic-it/Combei19
}}
==Is “manovra” Really “del popolo”? Linguistic Insights into Twitter Reactions to the Annual Italian Budget Law==
https://ceur-ws.org/Vol-2481/paper21.pdf
Is “manovra” really “del popolo”? Linguistic Insights into Twitter Re-
actions to the Annual Italian Budget Law1
Claudia Roberta Combei
University of Bologna
claudia.combei2@unibo.it
The impact of a social media post may be huge,
Abstract and unlike other prior forms of communication, it
can easily cross borders in just a few seconds. In
English. Relying on linguistic cues ob- fact, social media make things happen faster than
tained by means of structural topic model- ever before. For instance, Facebook and Twitter
ling as well as descriptive lexical anal- were crucial in allowing the Arab uprisings or the
yses, this study contributes to the general Romanian anti-corruption protests to happen
understanding of the Twitter users’ re- more efficiently and on a larger scale.
sponse to the annual Italian budget law ap-
proved at the end of December 2018. 2 Tweets and politics
Some topics contained in the dataset of
Besides their essential role in information dissem-
tweets are procedural or generic, but be-
ination, networking, and people mobilization, so-
sides those, it often emerges that Twitter
cial media are also important indicators and pre-
users expressed their concern with respect
dictors of their users’ opinions, sentiments and at-
to the provisions of this law. Supportive
titudes. In fact, various studies have explored peo-
attitudes seem to be less frequent. This pa-
ple’s reactions towards social, economic, and po-
per also advocates that findings from in-
litical issues, by analysing social media posts (e.g.
ductive studies on Twitter data should be
Burnap et al., 2014; Gaspar et al., 2016; Nesi et
interpreted with caution, since the nature
al., 2018), especially tweets, since they are easily
of tweets might not be adequate for draw-
retrievable by means of APIs.
ing far-reaching generalisations.
With over 6,000 tweets posted every second,
corresponding to roughly 350,000 per minute, 500
1 Introduction
million per day, and around 200 billion per year,
In the last decade, Internet has revolutionized hu- Twitter has become one of the main tools of com-
man communication and interaction. And among munication worldwide (Internet Live Stats, 2019).
all forms of digitally-mediated communication, The number of tweets written daily seems to be
social media stand out as one of the most effec- correlated to things happening in the real world,
tive. As Boulianne (2017) points out, the effects and, as a matter of fact, it was shown that im-
of social media depend on their nature of use (e.g. portant events generate high number of tweets (cf.
source of information; one-to-one/one-to- Hughes and Palen, 2009), something that is gen-
many/many-to-many communication; networking erally reflected also on the Twitter “trends”.
and relationship-building; expression of opinions; Based on Hootsuite’s (2019) report, each month,
etc.). in Italy there are almost 2.5 million active users2
Nowadays, potentially everyone with a com- of Twitter, a datum that confirms the popularity of
puter or a mobile device having access to the in- this network among various layers of Italian audi-
ternet can write and share contents which may be ence.
viewed and debated immediately by other people.
1
Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0
International (CC BY 4.0).
2
Users that write or share at least one tweet every month are defined “active”.
This means that Twitter may represent an easily By means of structural topic modelling (cf.
exploitable opportunity for politicians in their at- Roberts et al., 2014) and descriptive analyses (i.e.
tempt to reinforce communication with potential terminology extraction of multi-keywords and
voters in what might be defined as a permanent word sketches), we are interested in grasping the
digitally-mediated electoral campaign. Addition- Twitter users’ attitudes towards the budget law in
ally, it has been suggested that Twitter could be a significant moment for the first populist Govern-
used to model and predict public opinion and be- ment in the eurozone, namely the coalition formed
haviour regarding political events, such as elec- by Lega and M5S.
toral campaigns (e.g. Coletto et al., 2015; This topic is worth studying since the two par-
Kalampokis et al., 2017). In fact, Ott (2017: 59) ties displayed differences in economic, fiscal, in-
claims that Twitter may be the ideal tool for the frastructural, and social policies both in the elec-
afore-mentioned purposes since, it “privileges toral campaign for the 2018 general elections as
discourse that is simple, impulsive, and uncivil.” well as during the first months of government. For
While indeed tweets have been widely used to instance, Lega supported the flat taxation on in-
analyse public opinion and political discussions in comes, while M5S the citizen's basic income
all its forms, several methodological considera- (“reddito di cittadinanza” in Italian). However,
tions are dutiful. First of all, Twitter users do not these measures, although slightly modified, as
represent an optimal sample for public opinion or well as the amendment to the 2011 pensions re-
voting population, especially due to their higher form (“quota 100” in Italian) were included in the
than average level of education and political so- coalition agreement and subsequently in the draft
phistication, as well as a generally younger age for the annual budget law. The bill also contained
(cf. Gayo-Avello, 2013; Barberá et al., 2015). As various other economic and fiscal provisions (e.g.
a matter of fact, we believe it is more accurate to taxes on digital services; new VAT rates; reducing
define Twitter users as a potential share of elec- military expenses and the Italian contribution to
torate. Secondly, the language of tweets is charac- United Nations; new labour measures; environ-
terised by succinctness and sometimes informal- mental incentives; etc.)3.
ity, colloquialism, irony, and susceptibility to ru- We believe that the textual material contained
mour, all of which are aspects that render the re- in tweets may be promising in providing hints on
sults of large-scale analyses hard to interpret and how Twitter users – a fraction of the Italian voters
generalise. – reacted to the provisions of the budget law. Lin-
guistic insights into tweets might be able to guide
3 Aims and motivations us in understanding whether the so-called
“manovra del popolo” was perceived by Twitter
Acknowledging all the limitations mentioned
user as representing indeed the people’s interest.
above, this inductive exploratory study aims to
contribute to the growing body of literature exam- 4 Data
ining Twitter and its increasingly prominent role
in online communication by studying its applica- Although in the Western world there are three
tion in the context of political discourse. In partic- mainstream social media networks (i.e. Facebook,
ular, the linguistic approach presented here is Instagram, and Twitter), in this paper we analyse
providing insights into tweets regarding the dis- Twitter posts, primarily as a consequence of data
cussion and the approval of the annual Italian availability. Indeed, unlike other tools for social
budget law (in Italian “legge finanziaria” and/or media, Twitter APIs for R (R Core Team, 2018)
“legge di bilancio”). This law was also often la- allow scholars to collect large quantities of tweets
belled as “the manoeuvre” (in Italian “la and their related metadata in a rather effortless
manovra”) and “the people’s manoeuvre” (in Ital- way.
ian “la manovra del popolo”) by its proponents – Using the rtweet package (Kearney, 2019) for
in particular Movimento 5 Stelle (abbreviated R and Twitter’s developer account, we collected a
M5S) –, mainly due to some of its populist provi- dataset of 167,259 Twitter posts, for a total of 6.5
sions (e.g. the citizen's basic income and pension). million tokens, consisting in tweets and retweets
3
The full text of the annual Italian budget law (Legge 12-2018 - Suppl. Ordinario n. 62) and it is available
30 dicembre 2018, n. 145 – Bilancio di previsione dello online at this webpage: https://www.gazzettauffi-
Stato per l'anno finanziario 2019 e bilancio plurien- ciale.it/atto/stampa/serie_generale/originario (ac-
nale per il triennio 2019-2021) was published on the cessed on the 1st of June 2019).
Official Gazette of the Italian Republic (GU n.302 31-
related to the Italian budget law. Moreover, we ex- 4.1 Pre-processing
tracted 88 metadata describing the tweet (i.e. char-
Since the tweets and their metadata would have
acter length, device used, number of retweets,
been used for lexical analyses and structural topic
etc.) and the user (i.e. username, location, gender,
modelling5, we performed several pre-processing
etc.). In order to capture the most important
steps: defining a “stop words” list for Italian con-
phases of the Twitter discussion about the annual
sisting of roughly 1,000 lexically empty or unin-
budget law and considering the one-week rate
formative words (i.e. prepositions, conjunctions,
limit for tweets extraction imposed by the Stand-
auxiliary verbs, etc.); uniformizing, normalising
ard Search API4, the data were collected weekly
and cleaning the texts with various corpus pro-
from the 27th of November 2018 through the 8th of
cessing functions available on the R packages
January 2019, for a total of 43 consecutive days.
quanteda (Benoit et al., 2018), tm (Feinerer,
The hashtags used as keywords in the queries rep-
Hornik, and Meyer, 2008), and qdapRegex
resented all the names given to the budget bill by
(Rinker, 2017). Hashtags at the beginning and in-
Italian political actors, the press, and the public
side the tweet sentences were kept and decom-
opinion: “#leggedibilancio”, “#leggefinanziaria”,
posed into words (i.e. from “#trasportipubblici”
“#manovra”, “#manovradibilancio”, “#manov-
to “trasporti pubblici”), while those after the final
raeconomica”, “#manovradelpopolo”, and
point were removed, since most of the times they
“#manovrafinanziaria”. This guaranteed a large
represented one of the keywords used for extract-
coverage of Twitter users and tweet typologies.
ing tweets. Numbers, punctuation, sequences
Some of the afore-mentioned hashtags (e.g.
made up of a single character, and excessive white
“#manovra”, “#manovradelpopolo”) were also
spaces were removed as well. In order to further
trending at the end of December.
use temporal metadata as a covariate for the topi-
To avoid duplicates, we discarded all retweets
cal prevalence, the “created_at” metadatum was
and all posts that contained quotes of other tweets.
divided it into date and hour.
The removal process was obtained by filtering the
dataset, thus selecting only tweets whose values 5 Analyses and results
for “is_retweet” and “is_quote” corresponded to
“FALSE”. Duplicates other than retweets and As a result of the ever-growing interest and avail-
quotes were removed with R’s base functions du- ability of text data – often unstructured –, various
plicated – which identified duplicated tweets – statistical and machine-assisted approaches for
and unique – which extracted unique tweets. Since the analysis of textual material have been pro-
the aim of this study is to uncover the reactions of posed. In this paper we are employing the Struc-
the Italian voters active on Twitter, we removed tural Topic Model (STM) – a generative model of
the tweets written by political actors. To do so, we word counts – (cf. Roberts et al., 2014) in R to
defined a list containing the Twitter usernames of discover topics from tweets on the annual Italian
the members of the Italian Parliament, as well as budget bill and to estimate their relationship to
those of the official national and local party pro- temporal metadata.
files; this list was used to automatically filter and Similarly to Latent Dirichlet Allocation (cf.
remove tweets published by the unwanted pro- Blei, Ng, and Jordan, 2003) and Correlated Topic
files. We decided to keep tweets from news agen- Model (cf. Blei and Lafferty, 2007), in the STM
cies, online newspapers, and television channels, approach, a topic represents a mixture over words
since they could represent vectors of information where each word has a probability of pertaining to
exchange regarding the topic analysed in this a topic, whilst a document is a mixture over top-
study. The final dataset contained 20,891 tweets. ics, therefore a specific document can consist of
Tokens 701,986 various topics. The sum of the topic proportions
across topics for a specific document as well as
Words 414,803
the sum of word probabilities for a given topic
Types 75,485 both qual to 1. The main innovation of STM is the
Lemmas 31,947 possibility to model topical prevalence and topical
Table 1: Dataset statistics. content6 as a function of metadata. Here we are
4 5
A description of the Standard Search API for Twitter Considering the scope of this paper and the analyses
is available at this webpage: https://developer.twit- proposed, emoticons and emojis were left out.
6
ter.com/en/docs/tweets/search/api-reference/get- The topical prevalence shows the frequency with
search-tweets.html (accessed on the 1st of June 2019). which a specific topic is discussed, while the topical
using the date covariate to explain topical preva- Government and the oppositions on the provisions
lence over time. of the law.
After having calculated the estimated effects of
5.1 Topics the temporal covariate on topical prevalence, a
After having employed the STM’s searchK func- plot displaying this variation was created. Figure
tion to perform several tests, such as held-out like- 2 in Appendix shows how the afore-mentioned
lihood and residual analysis, the ideal number of topics varied over the 43 days considered. Topics
topics seemed to be between 10 and 14. Addition- are ordered as a function of their expected propor-
ally, STM gave the possibility to set the type of tions.
initialization, so here the spectral one was chosen, Firstly, there emerged that the variation was not
since previous studies had proven its stability and particularly strong, except for some topics. For in-
consistence (cf. Roberts, Stewart, and Tingley, stance, Topic 9 had a peak at the end of Decem-
2016). All results presented in this paragraph are ber/the beginning of January, suggesting that
based on a K of 10. The date of the tweet was used Twitter users might have written tweets of con-
as a prevalence covariate; as a word profile we cern soon after the approval of the annual Italian
opted for the highest probability. We did not use budget law. On the other hand, Topic 6, which
the stemming function on STM since it did not contained mostly tweets of support towards the
perform well on Italian. measures of the budget bill seemed to be prevalent
Figure 1 in Appendix shows the topics related primarily at the end of November and in mid-De-
to the annual Italian budget law as they emerged cember. The procedural topic was generally prev-
from the analysis of tweets. Each topic was further alent at the end of December, a timeframe corre-
classified into one category (i.e. EU & Confi- sponding to the vote and approval of the law. The
dence, Main Measures, Criticism & Concern, two topics summarising the negotiations with the
Government vs. Opposition, Procedures – Ge- EU, the confidence, and the possible infringement
neric, Support). This classification was based on procedure were pervasive during the entire period
the correlations obtained from a hierarchical clus- considered, with some peaks in early- and mid-
tering representation performed with the plot December. Topic 4 that regarded the disagree-
function of the stmCorrViz package (Coppola et ment between the Government and the opposition
al., 2016), on the review of the most characteris- was constant over time, and so were the topics de-
ing words, and on the examination of the most ex- lineating the main measures of the law.
emplar documents, namely the tweets that had the
5.2 Descriptive lexical analyses
highest proportion of words associated with the
topic. We were also interested in performing descriptive
Although we do not claim to model public lexical analyses on tweets. First of all, with the
opinion from tweets, interestingly, the topics terminology extraction tool on Sketch Engine
managed to echo various issues regarding the (Kilgarriff et al., 2014) we obtained multi-key-
budget law. Judging by the expected topic propor- words – able to convey more insights than single
tions, one could order the most prevalent topics as words on the issues examined – that appear more
follows: Topics 9, 8, and 3 (sum of topic propor- frequently in our dataset than in the reference cor-
tions: 0.29) reflect disapproval and doubts to- pus (i.e. Italian Web 2016 – itTenTen16, cf. Jaku-
wards the provisions of the budget law; Topics 1 bíček et al., 2013, for TenTen corpora). If we ex-
and 7 (sum of topic proportions: 0.22) describe the clude the hashtags used as keywords for tweets
difficult negotiation with the European Union extraction, these are the 30 most representative
(EU) and the threat of an infringement procedure; syntagmas in our dataset:
Topics 10 and 2 (sum of topic proportions: 0.19) Translation into
Syntagma
depict the main measures contained in the budget English
bill; Topic 6 (topic proportion: 0.13) illustrates the reddito di cittadi- the citizen’s basic
support to the budget bill and to the Government; nanza income
Topic 5 (topic proportion: 0.11) refers to the pro- procedura di infra- infringement pro-
cedures regarding the discussion, the vote, and the zione cedure
approval of the budget law; and Topic 4 (topic clausole di salva-
proportion: 0.06) reveals the conflict between the safeguard clauses
guardia
content represents the words used to discuss about that
topic (cf. Roberts et al., 2014: 1068).
voto di fiducia confidence vote Generally, three different scenarios are distin-
blocco assunzioni hiring freeze guishable.
professioni sanita- health professions First of all, there were several neutral verbs,
rie senza titolo without a degree nouns, and modifiers associated to the budget law,
flat tax flat tax most of which regarding its procedural aspects.
commissione bilan- The most frequent (i.e. frequency ≥ 10.81 per mil-
budget committee lion) are listed below:
cio
gilet azzurri blue vests Translation into
Word/Syntagma
taglio pensioni pension cuts English
scatoletta di tonno tuna can scrivere write
previous govern- cambiare change
governi precedenti modificare modify
ments
pensioni minime minimum pensions discutere discuss
scatola chiusa black box approvare approve
nuove tasse new taxes contenere contain
promesse elettorali campaign promises prevedere consist
fasce deboli vulnerable citizens varare launch
deficit strutturale structural deficit votare vote
technical arrange- passare pass
accordo tecnico riscrivere rewrite
ment
braccio di ferro trial of strength promulgare promulgate
appalti senza gara no-bid contracts gialloverde yellow-green
assurdità totale total nonsense economica economic
terrorismo media- finanziaria financial
media terrorism populista populist
tico
auto inquinanti polluting cars discussione discussion
più tasse more taxes commissione commission
sovereignist gov- bilancio budget
governo sovranista Table 3: Neutral associations.
ernment
manovra contro il manoeuvre against Next, some positive evaluations of the budget
popolo the people law emerged. The most frequent (i.e. frequency ≥
false promesse false promises 10.81 per million) are listed below:
IVA sui tartufi VAT for truffles Translation into
Word/Syntagma
popolo italiano Italian people English
Table 2: The most representative syntagmas in favorire (l’innova- favour (innova-
the dataset. zione) tion)
It is clear that various multi-word expressions grande big
referred to procedural aspects, such as those re- buona good
flecting the vote and the approval of the budget bella beautiful
law (e.g. “confidence vote”), while others were significativa significant
used to list its measures, especially fiscal and eco- del popolo of the people
nomic policies (e.g. “the citizen’s basic income”, del cambiamento of the change
“flat tax”, etc.). Nevertheless, various syntagmas per i cittadini for the citizens
seemed to express doubts with respect to the pro- per la crescita for the growth
visions of this law. In fact, often, the words chosen Table 4: Positive associations.
by many Twitter users to express their criticism Nonetheless, several word associations seemed
were rather strong (e.g. “total nonsense”, “black to suggest negative reactions to the budget law.
box”, “sovereignist government”, etc.). The most frequent (i.e. frequency ≥ 10.81 per mil-
These concerns and rather negative reactions to lion) are shown below:
the budget bill were reflected also in the word Translation into
sketches (i.e. visual representations of colloca- Word/Syntagma
English
tions and word combinations obtained on Sketch recessiva recessive
Engine) for the words “manovra” and “legge”. piena di errori full of errors
dannosa dangerous for drawing steady generalizations, even if the
cattiva bad prospects they offer for content and discourse
iniqua unfair analysis are indeed significant. Further research
scellerata wicked on this topic might include the investigation of
sbagliata wrong Twitter user’s reactions by means of sentiment
snaturata wretched analysis.
taroccata false
vuota empty References
assurda absurd Pablo Barberá, John T. Jost, Jonathan Nagler, Joshua
folle deranged A. Tucker, and Richard Bonneau. 2015. Tweeting
truffa fraud from Left to Right. Psychological Science,
contro il popolo against the people 26(10):1531–1542.
del popolino of the masses Kenneth Benoit, Kohei Watanabe, Haiyan Wang, Paul
del cappio of the noose Nulty, Adam Obeng, Stefan Müller, and Akitaka
da lacrime tearful Matsuo. 2018. quanteda: An R package for the
scontro dispute quantitative analysis of textual data. Journal of
Open Source Software, 3(30):774.
protesta protest
vergogna shame David M. Blei and John D. Lafferty. 2007. A correlated
bocciatura failure topic model of Science. The Annals of Applied Sta-
tistics, 1(1):17–35.
della povertà of the poverty
dell’assistenzi- David M. Blei, Andrew Y. Ng, and Michael I. Jordan.
of welfarism 2003. Latent Dirichlet Allocation. Journal of Ma-
alismo
buio dark chine Learning Research, 3(3):993–1022.
diminuire diminish Shelley Boulianne. 2017. Revolution in the making?
tagliare cut Social media effects across the globe. Information,
Table 5: Criticism associations. Communication & Society, 22(1):39–54.
Finally, using the tm’s findAssocs function, we Pete Burnap, Matthew L. Williams, Luke Sloan, Omer
calculated the associations of the lemma Rana, William Housley, Adam Edwards, Vincent
“manovra” in the term-document matrix; some of Knight, Rob Procter, and Alex Voss. 2014. Tweet-
the afore-mentioned criticism words (e.g. “ab- ing the terror: modelling the social media reaction to
surd”, “recessive”, “bad”) had a correlation higher the Woolwich terrorist attack. Social Network Anal-
than 0.03, suggesting a rather frequent co-occur- ysis and Mining, 4(206):1-14.
rence. Mauro Coletto, Claudio Lucchese, Salvatore Orlando,
and Raffaele Perego. 2015. Italian Information Re-
6 Conclusions trieval Workshop - IIR 2015. In Proceedings of the
6th Italian Information Retrieval Workshop, Ca-
This paper explored the Twitter users’ reactions to gliari. CEUR Workshop Proceedings.
the annual Italian budget bill. STM outputs and
descriptive lexical analyses showed that tweets Antonio Coppola, Margaret E. Roberts, Brandon M.
Stewart, and Dustin Tingley. 2016. stmCorrViz: A
concerned various aspects associated to the object
Tool for Structural Topic Model Visualizations. R
of this study. Apart from talking about procedural package version 1.3. Retrieved from https://cran.r-
and generic issues, users expressed their doubts project.org/web/packages/stmCorrViz/index.html/
and disapproval with respect to the measures of (accessed on the 1st of June 2019).
the budget law. Generally, tweets supporting this
Ingo Feinerer, Kurt Hornik, and David Meyer. 2008.
law were less frequent. The findings of this study,
Text Mining Infrastructure in R. Journal of Statisti-
although preliminary, might be seen as indicators cal Software, 25(5):1-54.
of what subsequently turned out to be a failure for
the first Conte government. Still, as reiterated Rui Gaspar, Cláudia Pedro, Panos Panagiotopoulos,
throughout the paper, the results might not reflect and Beate Seibt. 2016. Beyond positive or negative:
Qualitative sentiment analysis of social media reac-
the real attitudes of the Italian voting population,
tions to unexpected stressful events. Computers in
since Twitter users tend to be younger and to have Human Behavior, 56:179–191.
an above the average level of education and polit-
ical sophistication (cf. Barberá et al., 2015). Daniel Gayo-Avello. 2013. A Meta-Analysis of State-
Moreover, tweets, by nature, might not be suitable of-the-Art Electoral Prediction from Twitter
Data. Social Science Computer Review, 31(6):649– Margaret E. Roberts, Brandon M. Stewart, Dustin Tin-
679. gley, Christopher Lucas, Jetson Leder-Luis, Shana
Kushner Gadarian, Bethany Albertson, and David
Hootsuite Media Inc. 2019. Digital in 2019. Retrieved
G. Rand. 2014. Structural Topic Models for Open-
from https://hootsuite.com/it/risorse/digital-in-
Ended Survey Responses. American Journal of Po-
2019-italy (accessed on the 1st of June 2019).
litical Science, 58(4):1064–1082.
Internet Live Stats. 2019. Twitter Usage Statistics. Re-
Margaret E. Roberts, Brandon M. Stewart, and Dustin
trieved from https://www.inter-
Tingley. 2016. Navigating the Local Modes of Big
netlivestats.com/twitter-statistics/ (accessed on the
Data: The Case of Topic Models. In R. Michael Al-
1st of June 2019).
varez (editor), Computational Social Science: Dis-
Miloš Jakubíček, Adam Kilgarriff, Vojtěch Kovář, covery and Prediction (Analytical Methods for So-
Pavel Rychly, and Vít Suchomel. 2013. The TenTen cial Research), 51–97. Cambridge University
Corpus Family. In 7th International Corpus Lin- Press., Cambridge.
guistics Conference CL 2013. Lancaster, 125-127.
Evangelos Kalampokis, Areti Karamanou, Efthimios
Tambouris, and Konstantinos Tarabanis. 2017. On
Predicting Election Results using Twitter and
Linked Open Data: The Case of the UK 2010 Elec-
tion. Journal of Universal Computer Science,
23(3):280–303.
Adam Kilgarriff, Vít Baisa, Jan Bušta, Miloš Jaku-
bíček, Vojtěch Kovář, Jan Michelfeit, Pavel Rychlý,
and Vít Suchomel. 2014. The Sketch Engine: ten
years on. Lexicography, 1(1):7–36.
Michael W. Kearney. 2019. rtweet: Collecting Twitter
Data. R package version 0.6.9 Retrieved
from https://cran.r-project.org/package=rtweet (ac-
cessed on the 1st of June 2019).
Amanda Lee Hughes and Leysia Palen. 2009. Twitter
adoption and use in mass convergence and emer-
gency events. International Journal of Emergency
Management, 6(3/4):248–260.
Paolo Nesi, Gianni Pantaleo, Irene Paoli, and Imad
Zaza. 2018. Assessing the reTweet proneness of
tweets: predictive models for retweeting. Multime-
dia Tools and Applications, 77(20):26371–26396.
Brian L. Ott. 2017. The age of Twitter: Donald J.
Trump and the politics of debasement. Critical
Studies in Media Communication, 34(1):59–68.
R Core Team. 2018. R: A language and environment
for statistical computing. R Foundation for Statisti-
cal Computing, Vienna, Austria. Retrieved from
http://www.R-project.org/ (accessed on the 1st of
June 2019).
Tyler W. Rinker. 2017. qdapRegex: Regular Expres-
sion Removal, Extraction, and Replacement Tools.
R package version 0.7.2. University at Buffalo. Buf-
falo, New York. Retrieved from
http://github.com/trinker/qdapRegex/ (accessed on
the 1st of June 2019).
Appendix
Figure 1: Topics and word probabilities.
Figure 2: Variation of topic proportions over time.