Share and Shout: Proto-Slogans in Online Political Communities

This paper proposes a methodology for investigating populism on social media by analyzing the emergence of proto-slogans, deﬁned as nominal utterances (NUs) typical of a political community on social media. We extracted more than 700.000 comments from the public Facebook pages of two Italian populist parties’ leaders (Matteo Salvini and Luigi Di Maio) during the week preceding the 2019 European elections (i.e., from May 20 to May 26, 2019). These comments have been automatically clustered and manually annotated to ﬁnd proto-slogans created by the parties’ supporters. Our manual annotation consists of four layers, namely: Nominal Utterances (NUs), a syntactic device widely used for slogans; Slogans for NUs with a slogan function; Top-down/Bottom-up , to recognize the slogans produced by the politicians and those produced by supporters; Proto-slogans , for NUs devoid of speciﬁc political content that nonetheless express partisanship and support for the leaders.


Introduction
Social media have increasingly become arenas of mainstream political discourse. Platforms like Facebook and Twitter offer politicians venues to express their views, aggregate supporters and critics, and reinforce identities (Stieglitz & Dang-Xuan, 2012;Stier, Bleier, Lietz, & Strohmaier, 2018).
The vast amount of comments on political topics daily produced by users can be monitored and analyzed using Natural Language Processing (NLP) tools to focus on relevant societal issues such as hate speech and fake news. However, apart from long comments that express more complex opinions, most comments on social media are characterized by the synthetic expression of a point of view. Analyzing this type of content is challenging: due to their brevity, a topic-based analysis of users' comments performs poorly. Nonetheless, short comments have a pragmatic function in online debates, and often through the use of nominal utterances (NUs) (Comandini & Patti, 2019;Comandini, Speranza, & Magnini, 2018), help to build a common view among supporters of the same party/politician. NUs, intended as syntactic declarative constructions built around a nonverbal head, can be part of a shared vocabulary used to express the in-group sense of cohesion and belonging on political pages and fora. For example, the NUs Italia agli Italiani (Italy to Italians) and Porti chiusi (Closed harbors) uniquely characterize one of the political communities investigated in this paper, i.e. Lega Nord (LN, Northern League). Several of these recurrent NUs are slogans carefully created by the politicians' communication staff and used by supporters to reinforce the sense of belonging to a community. However, political slogans can also emerge from supporters' interactions on social media such as Facebook. They can become a trademark of a political community on other social media, such as Twitter. We define this process as proto-slogan generation.
Proto-slogans are semi-fixed linguistic expressions realized by NUs; they express a generic stance -positive or negative -toward a target. They emerge in online environments, in communities of people sharing the same perspectives or points of view. We find that proto-slogans are a communicative device exploited by populist supporters, a stylistic feature usable in the future to detect emerging populist attitudes.
In this paper, we study online political communities, extracting comments from the public Facebook pages of two populist Italian party leaders, Matteo Salvini for the Lega Nord (LN, Northern League) and Luigi Di Maio for the Movimento 5 Stelle (M5S, Five Stars Movement), during the week preceding the 2019 European elections (i.e., from May 20 to May 26, 2019). At that time, these two leaders were both covering the position of deputy prime minister in the so-called yellow-green government. Their parties were gaining consensus, with the LN winning the European elections. However, both parties have lost consensus at the time of writing, and their leaders have changed their communication. These changes are mainly due to the new roles these leaders are now covering (Salvini being part of the ruling majority with no position in the government and Di Maio being Foreign Minister). Besides this difference, the data analyzed still represent a valuable tool for gaining insights into the communication strategies of populist leaders and parties.
To investigate the linguistic behavior of these communities, we propose a semisupervised methodology that combines K -means clustering and manual annotation for the identification of proto-slogans. Additionally, we compare slogans extracted from Facebook with slogans retrieved on Twitter in different periods to distinguish between attested and emerging slogans. This comparison validates what has been extracted from Facebook's public pages on Twitter, where linguistic choices can be crucial for identifying communities if there are no other metadata available (such as the information that a user follows a politician). from social media have been extracted and pre-processed. The results emerging from the manual annotation of automatically clustered instances are discussed. In Section 4, we provide some preliminary comparisons between Facebook and Twitter data, before concluding in Section 5.

Related Work
Populist discourses have been analyzed from different perspectives, collecting them in corpora that ease qualitative and quantitative analyses. However, it is still unclear if populist discourses can be linguistically identified -independently from the historical moment, the latitude, and the political orientation. Moreover, social media has had a disruptive effect on the propagation of populist discourses, affecting some of its features in a way still under scrutiny. This section reports on relevant literature about populism on social media and how language is shaped by communication in online communities in order to frame the reception of populist discourses in online contexts. In particular, slogans are presented as stylistic devices related to the emergence of a shared attitude among users.
In this work, we will adopt a broad definition of populism as a discourse based on the juxtaposition of two homogeneous and antagonistic groups: "the good people" (the in-group) VS "the bad elite/the foreigners" (the out-group) (Mudde, 2004;Rooduijn & Akkerman, 2017).
Charismatic leaders are particularly relevant for populist parties, and on Facebook they are often more popular than the official party's page (Bobba, 2019). Thus, to study populist rhetoric, it is preferable to focus on the rhetoric of political party leaders, analyzing how supporters react to it.
Populist leaders often adopt an emotional and straightforward communication style to be more persuasive and trigger a more emphatic response on social media (Oliver & Rahn, 2016). Indeed, it has been shown that emotionalized-style messages produced by Matteo Salvini on Facebook are more popular than his more neutral messages (Bobba, 2019).
Using these emotional messages and the direct connection with the public provided by social networks, populist leaders can forge close ties with their supporters, appearing more approachable (Jacobs & Spierings, 2016). Therefore, populist leaders can transform their Facebook pages into sheltered spaces for their fans, creating echo chambers in which aggressive tones can be cultivated (Ernst, Esser, Blassnig, & Engesser, 2018;Engesser, Fawzy, & Larsson, 2017).
Together with the sense of belonging to the in-group due to the general resentment toward the out-group (Hameleers, Reinemann, Schmuck, & Fawzi, 2019), this perceived intimacy with the leader creates a strong sense of being part of a homogeneous community, supportive of their leaders. In this way, populist leader's supporters may experience inter-group emotions, with each member experiencing emotions and taking action on behalf of the group (Smith & Mackie, 2008).
Previous computational linguistics studies of populism are scarce. Recently, (Huguet Cabot, Abadi, Fischer, & Shutova, 2021) present a crowdsourcing annotated dataset for populist attitudes that collects comments about news on Reddit that mention a set of social groups (i.e., immigrants and Muslims), manually classifying attitudes toward them as supportive, critical, or discriminatory. In detecting the overall stance of comments, their analysis does not target exclusively populistic content and how populist attitudes are expressed. Instead, our work starts with the assumption that comments on the chosen politicians' Facebook public pages are mainly supportive and sympathetic to the populist rhetoric. Thus, they constitute the ideal starting point for a stylistic investigation of the reception of populist discourse online.

The language of online communities
Computational analyses of language used in online communities revealed that talking in a particular way on social media reinforces our networks and sense of belonging (McCulloch, 2019). For example, the use of written slang on Twitter depends on the number of times people see the new word and if a member of their network uses it or not (Eisenstein, O'Connor, Smith, & Xing, 2014). The central members of the network introduce lexical innovations that are successfully adopted by other members if there is a subset of adopters with strong ties (Tredici & Fernández, 2018). Creating a shared vocabulary is both a prerequisite and a consequence of being part of a cohesive community, even online.
With a data-oriented analysis, Khalid and Srinivasan (2020) show that different communities have peculiar styles, and the stylistic choices of users are a good predictor of group membership, more than the topics discussed. The adaptations at the stylistic level contribute to being well-received by a community. In (Tran & Ostendorf, 2016) the reception of content posted by users with positive feedback is investigated through a hybrid n-grams and topic models to characterize the style and the topic of language in Reddit online communities. Stylistic features have discriminatory power for distinguishing between communities: the style is a better indicator of community identity than the topic. The authors found a positive correlation between the community reception of a contribution and the style similarity to that community. On the contrary, this does not hold for topic similarity.
In this paper, we argue that nominal utterances are a stylistic device that characterizes the reception of populist discourse online, being the pragmatic choice that supports the creation of a shared vocabulary among supporters.

Slogans as Linguistic Devices
Slogans are usually short, expressive, and assertive utterances, easy to memorize and spread (Amălăncei, Cîrtiţă-Buzoianu, & Daba-Buzioianu, 2015). They are defined linguistically by their pragmatic function: expressing an idea memorably and economically. They can have a broad range of syntactic forms and can be characterized by their use of figures of speech and rhetorical devices, such as metaphors ("Imagination at Work" from General Electric implies the metaphor that General Electric is imagination), parallel constructions ("Melts in your mouth, not in your hands" from M&M's) or alliteration ("Don't dream it. Drive it" from Jaguar) (Alnajjar & Toivonen, 2020).
The slogans that have received attention in previous works are those used in advertising. Political slogans are less studied, although they generally follow the same advertising rules and have the same goal: influencing people's behaviors (Ferrier, 2014). Furthermore, political slogans usually convey a strongly supportive or condemning message towards a person or a political program/action, because voters are mainly influenced not by their conscious opinion on a politician's program, but by their feeling about a candidate or a party (Westen, 2007)).
The procedure primarily used in studying slogans is top-down, concentrating on pre-existing slogans professionally crafted by politicians or companies. On the contrary, a bottom-up approach is much more complicated, because it would require recognizing slogans in day-to-day communication focusing on linguistic features.
Top-down slogans have a pragmatic function: they are created to persuade others. On the other hand, bottom-up slogans, emerging as linguistic devices shared by like-minded people, have a different function: they are used to structure and enhance the cohesion of online communities. In this paper, we present a methodology to detect bottom-up slogans that, if widely adopted, can shed light on the emerging attitudes of political supporters online.

Methodology
This section presents a definition of proto-slogan and reports on how data have been extracted from Facebook and pre-processed before automatically clustering them. In addition, the details of the manual annotation performed on clustered data are documented.

Definition of proto-slogan
Regarding the general political rhetoric, it is possible to differentiate two different kinds of slogans: top-down and bottom-up. Top-down political slogans are produced by the politician's communication team and generally convey a complex message in a carefully crafted short, sharp form, such as porti chiusi (closed harbors); on the other hand, bottom-up political slogans are produced by the political electorate and generally convey a less complex message, as in avanti tutta (full steam ahead).
However, analyzing bottom-up political slogans, it would appear that some of the linguistic devices used by the political electorate, such as forza Salvini (go Salvini), are even simpler than the others, conveying an even less complex message. These very simple linguistic devices do not appear to be real slogans, because they convey only the user's stance (positive or negative) regarding a target, which is always explicitly mentioned.
Some of the linguistic devices used by the political electorate are slogan-like constructions that enhance cohesion inside the group. However, they are also simpler than real political slogans produced spontaneously by the politician's supporters, which still convey a relatively complex message.
For example, these slogan-like constructions usually appear as concise messages supporting (example 1) or denigrating (example 2) a politician or a group of people.

Giornalisti venduti! [Corrupted journalists!]
Short slogan-like bottom-up constructions that convey a basic message of support/denigration will be called proto-slogans, assuming that they are an embryonic form of a real slogan, since they convey a positive or negative stance, but not the more complex messages typical of slogans. More specifically, proto-slogans are classified as a subset of bottom-up slogans, as illustrated in Figure 1. In this paper, we elaborate on the notion of proto-slogan as a specific device to build cohesion in online communities, proposing a methodology to identify them in social media.
Even if peculiar syntactic structures do not characterize slogans, slogans and especially proto-slogans are often realized syntactically as nominal utterances (Comandini et al., 2018), also known as fragments without an overt antecedent (Merchant, 2005). NUs are linguistic constructions without a verb in a finite form in their syntactic nucleus and are very common in informal spoken English (Merchant, 2005) (example 3) and Italian (Cresti, 1998)  NUs convey their content in a way that is expressive and informative but also very economic (Ferrari, 2011a(Ferrari, , 2011b. It has been found that NUs are often used as a device to convey hate speech, as shown in the POP-HS-IT corpus, by frequently taking the form of a verbless, hateful slogan (Comandini & Patti, 2019), such as TOLLERANZA ZERO (ZERO TOLLERANCE) and rimpatriare subito tutti gli immigrati irregolari (immediately repatriate all irregular immigrants). Indeed, being without a verb in a finite form, NUs do not convey information about time, person, or aspect, creating messages similar to always valid maxims, mottoes, and, more importantly, actual slogans (Benveniste, 1990). Thus, not all NUs are slogans, but slogans are often NUs.
Therefore, many slogan-like constructions are NUs. Slogan-like NUs can be separated into two groups, as illustrated in Figure 1: top-down slogan-like NUs, produced by the politician's communication team, and bottom-up slogan-like NUs, produced spontaneously by the politician's supporters. Bottom-up slogan-like NUs also have an additional subset: proto-slogans, which convey a positive or negative stance towards a target.

Data Extraction and Preprocessing
This paper focuses on the online audience of Matteo Salvini and Luigi Di Maio, the two leaders of widely recognized populist parties, LN and M5S. In the selected period for the data collection, between May 20 and May 26, 2019, covering the last week of the political campaign before the 2019 European elections, their communication was primarily conveyed through social media. We used Netvizz (Rieder, 2013), a tool that crawls data from Facebook, 1 to extract posts and comments from the Facebook public pages of the two politicians. We have excluded all posts written by the leaders and the replies to comments by other users, focusing our analysis on direct comments  to the posts. Table 1 reports an overview of the extracted messages in terms of the average length of the posts published by the politicians (in tokens), the number of direct comments, and the average length of the comments. As Table 1 shows, the two groups have similarities and differences. In both cases, we observe that the average length of the users' comments tends to be shorter than that of the posts published by the politicians. At the same time, it seems that the two groups of users (and ideally, the level of the interactions with their leaders) tend to differ, with users on Salvini's page producing shorter messages than those on Di Maio's. Since short comments are often verbless, we focus on nominal utterances (NUs) as syntactic declarative constructions built around a nonverbal head, framing them as the minimal unit of meaning in online communication. Therefore, we developed a preprocessing procedure to find out all NUs contained in the comments. Each comment has been preprocessed according to the following steps: • its content has been sentence-splitted with NLTK (Bird, Klein, & Loper, 2009); • its content has been PoS-tagged with TreeTagger (Schmid, 1995); • finally, sentences that contain a verb in the finite form have been filtered out to include in the final dataset only the potential nominal utterances; sentences containing proper nouns other than Matteo, Salvini, Luigi, and Di Maio have been filtered out to exclude comments mentioning Facebook users.
The dimensions of the dataset before and after the preprocessing steps are reported in Table 2. The preprocessed data, more than 240k comments, have been the input for the clustering algorithm based on semantic similarity.

Aggregating Data Through Clustering
The amount of data after preprocessing is such that a manual exploration is not feasible (see Table 2 for details). We thus decided to aggregate the pre-processed messages using clustering and perform manual annotation on aggregated data. Our approach is based on K -means clustering.
Such an approach has advantages and disadvantages for our task and, most importantly, our data. Results from K -means are easy to interpret and can be refined by manual inspection. At the same time, we are aware that K -means is not the best solution in an exploratory task, such as ours, where the number of clusters is not known, and it can hardly be assumed a priori. In this case, using known estimate methods such as the Elbow curve does not represent a solution.
We have addressed this issue by empirically validating the clusters of different sizes by using a sample of the data of 40k messages from Salvini's comments. First, each message representing an eligible NU has been converted as a 300-dimensions vector using FastText (Bojanowski, Grave, Joulin, & Mikolov, 2016). We then computed the pairwise cosine similarity scores between vectorized messages. The result is a N by N matrix of similarity scores.
Similarity scores below 0.6 were excluded and replaced with zeros to reduce noise in the data. 2 The matrix has been used as input to the K -means algorithm. 3 We experimented with generating three groups of clusters of different sizes: 100, 150, and 200. Although none of them would correspond to an ideal amount of clusters for the aggregation of users' messages, their sizes allow for an easy and quick manual exploration of the data, ensuring a fine-grained level of analysis. We plotted their centroids and observed their distributions for each group of clusters. Quite interestingly, we could not find distinguishing differences or remarkable patterns. We finally selected 150 clusters as an appropriate level of aggregation to be subsequently manually annotated. Finally, we clustered the comments from Salvini's page daily, creating eight sets of comments, while aggregating those for Di Maio in three blocks. These differences are due to the number of messages available for the two politicians.

Manual annotation
Since centroids have been obtained by means of semantic similarity scores, focusing on them is a way to avoid annotating all the comments (a task that is not feasible) or annotating a not representative sample. The list of centroids (1,650 in total) obtained from K -means clustering has been manually annotated by two annotators with four annotation layers. These annotation layers are performed sequentially, and each of them is essential to understand the frequencies of NUs with different functions in each community. The agreement between annotators is calculated after discussion on divergent choices.
The first layer identifies NUs, which can be annotated following Comandini and Patti (2019) guidelines with a good agreement (0.96 in terms of Cohen's Kappa). We considered hashtags formed by two or more words as a single noun for this task, even if they contained a verb in a finite form. Most of these verbal hashtags are not used as VPs, but as nominal elements, linking the post to an "existing collective practice" (Zappavigna, 2015). The clause is excluded from the annotation when a NU has a coordinate clause with a verb in a finite form. Verbs in a non-finite form (infinitive, gerund, and participle) can be included in a NU, as they do not convey informations about Tense, Aspect and Mood.
The The second annotation layer recognized particular NUs with a slogan-like form, with a binary value (yes-no). As noticed in Section 2.3, an utterance is a slogan because of its purpose. Labeling an utterance produced by an anonymous user as a slogan is not a trivial or straightforward task, even if it is pretty simple to recognize political slogans created by politicians. Inter-annotator agreement for this level is 0.65 in terms of Cohen's Kappa, showing that recognizing slogans involves some form of subjective interpretation. Below we report examples of slogans in our dataset:

<NU> Forza Salvini </NU> [Go Salvini]
The third layer has been applied only to those items previously annotated as slogans by both annotators, distinguishing between top-down and bottom-up slogans. Top-down slogans are created by the political leader or party, while fans spontaneously produce bottom-up slogans. Annotators reached a better agreement on this distinction (0.74 Cohen's Kappa). One example for each category is reported below:

<NU> Forza capitano </NU> [Go captain] [bottom-up]
As illustrated by example 12, bottom-up slogan-like NUs tend to be semantically close to encouragements and cheers that characterize sports competitions. They generally do not convey complex meanings but endorse the leader's message; they are phatic expressions with a clear social function (Jacobson, 1960).
As Table 3 illustrates, these NUs are predominant in the annotated dataset. Not surprisingly, the set of top-down slogans annotated is smaller than the set of bottom-up Salvini  926  22  204  Di Maio  369  5  57   Table 3: Eligible NUs after preprocessing slogans: politicians' staff produce few slogans to communicate the politician's message.

FB page NUs Top down slogans Bottom up slogans
On the other hand, supporters use a broader set of NUs. The fourth level of annotation explicitly targets proto-slogans, with an inter-annotator agreement of 0.63: several slogan-like NUs (in alto i cuori (lift up your hearts), sempre e per sempre (forever and ever)) are not proto-slogans because they are hapax in the list of centroids and lack of specific content. We recognize as proto-slogans the following NUs:

• <NU> via i ladroni </NU> [away the robbers]
• <NU> m5s tutta la vita </NU> [m5s for the rest of my life] In Table 4 we report how many NUs have been labeled as proto-slogans. Bottomup NUs are proto-slogans when they express a positive or negative stance towards a discourse target (always explicitly mentioned), whose identity is common knowledge for the electoral base.

Source
Bottom-Up NUs Proto-slogans Salvini 196 102 Di Maio 57 25 Comparing the bottom-up slogans and proto-slogans produced by the users to those produced by the politicians, it is clear that Salvini uses both these kinds of slogans very frequently, while Di Maio generally uses only top-down slogans. Salvini often uses bottom-up slogans such as avanti tutta (full steam ahead), which appears three times a week and it is also frequently used by Salvini's followers in the comments, often preceded by a proto-slogan such as forza Matteo (go Matteo). However, the slogans most used by Salvini, appearing at least once a day, are two proto-slogans, both with a positive stance towards Italy or Italians: prima l'Italia (Italy first) and prima gli Italiani (Italians first). These proto-slogans are not used in the comments by Salvini's followers, unlike top-down slogans such as porti chiusi (closed harbors). Thus, prima l'Italia/gli Italiani, while it conveys a political stance and it is used by a political leader, does not act as a top-down slogan.
Therefore, we may suppose that these proto-slogans act like a turn in an ongoing dialogue between Salvini and his followers, both of them expressing their support to each other through proto-slogans: Salvini expresses a positive stance towards his followers, who in return express their support to him through proto-slogans such as forza Matteo (go Matteo).
Furthermore, Salvini refers to his followers as "Italians" using a very common populist strategy that identifies populist voters with "the people" and, in this case, with the Italian population as a whole. In this way, Salvini identifies his electorate with the Italian population, giving the impression of a much larger voter base and giving his electorate the perception that they are the real Italians, while their opponents are not.

Facebook and Twitter Data Comparison
Slogan-like NUs are specific linguistic items for a political community, when supporters use them. However, they display different frequency patterns over time, i.e., they emerge as more frequent in a specific period. Therefore, the relationship between the frequencies of bottom-up slogans on social media and proto-slogans need a more complex investigation based on more data.
We propose a qualitative classification of slogan-like NUs complementary to protoslogans' characterization. In order to investigate this aspect, after extracting and annotating nominal utterances from Facebook public pages, the list of NUs was searched on Twitter with the help of GetOldTweets3 Python library in three different oneweek time spans across 3 years (2019,2018,2017). 4 The aim of this analysis is the identification of three types of slogan-like NUs: • Generic slogan-like NUs: nominal utterances whose content does not directly concern populism or are specifically related to the leaders. They can not be proto-slogans; • Attested slogan-like NUs: specific to populist messages concerning Di Maio and Salvini, some attested slogan-like NUs are frequently used, but their presence varies through different periods. They tend to be proto-slogans, especially if they are bottom-up; • Episodic slogan-like NUs: these NUs are linked to a specific event or period. However, they could still emerge as attested NUs if their use continues beyond a specific period. More data are needed to decide if they are proto-slogans or not.   Table5) and with longer lifespans, primarily if they are not referred to a specific event, being instead relevant in a more general way. Thus, top-down slogans usually are attested slogan-like NUs or episodic slogan-like NUs.
A top-down episodic slogan made for the European election like #domenicavotolega (#sundayIvoteLega) is well-attested several months later, probably because it is still relevant for the next Italian Regional elections, planned on a Sunday too. Similarly, the generic, encouraging hashtag #iostoconsalvini (#Istaywithsalvini), an attested slogan, has been productive in every period considered. On the contrary, the more specific and episodic #26maggiovotolega (#26mayIvoteLega) is significantly less used after the European elections. Twitter displays some of LN's and M5S's main leitmotifs: the slogan-like NUs porti chiusi (closed harbors) and tutti a casa (everybody home). In 2018, porti chiusi had been used often to answer Matteo Salvini's tweets, while in 2019 appeared more frequently in free-standing tweets. Porti chiusi is an example of an attested slogan-like NU that is distinctive for a political community but can also be used to address this community, criticizing its members. Bottom-up slogan-like NUs are generally present on Twitter, but they show some peculiar differences from those on Facebook. Firstly, particularly familiar generic slogan-like encouragements like forza matteo (go matteo), very frequent on Facebook, are rare on Twitter, and they never appear in answers to Matteo Salvini's tweets. The less informal forza salvini (go salvini), avanti capitano (come on captain) and forza capitano (go captain) are far more frequent. Still, while on Facebook, they were placed inside the private echo chamber of the leader's page. They do not appear in answers to Matteo Salvini's tweets on Twitter, but they are characteristic of independent tweets. Most of the bottom-up generic slogan-like NUs, like noi tutti con te (all of us with you), are not attested on Twitter, but there are a few notable exceptions, such as avanti tutta (full steam ahead), sempre avanti (always forward) or vergogna (shame).
However, this investigation is still preliminary since it has not been possible to ensure that tweets with bottom-up generic slogan-like NUs, such as forza capitano (go captain), are unquestionably referred to LN. If the user explicitly mentions the politician, disambiguation is possible. Otherwise, the tweet could be used to support a football team.

Conclusions and Future Work
Political communication on social media can be investigated with real data available on Twitter and Facebook public pages. This paper introduces the concept of proto-slogan as an economical device used to build and reinforce the in-group sense of belonging in online political communities. We introduced a methodology for identifying NUs peculiar to a political community on social media. These NUs extracted from centroids, derived from the Facebook public page of Matteo Salvini and Luigi Di Maio, are often slogan-like.
The political party or leader creates top-down slogans, and they are generally more linked to the party's program. Instead, the supporters produce bottom-up slogans, which we define as proto-slogans, and they are usually less specific and more linked to informal encouragements.
Recognizing these slogan-like NUs makes it possible to recognize supporters of a specific populist political party, even when their messages are not otherwise contextually linked to it. Even if less specific, bottom-up slogan-like NUs are still recognizable on Twitter. They can uncover political support without explicit political content. However, refining the automatic recognition of NUs is still necessary since informal computermediated communication typically shows a substandard variety of Italian. For example, some verbs in the finite form may appear inside a NU, since they have a non-standard spelling.
Our analysis represents the first step toward identifying stylometric patterns in the populist electorate's informal writing on social media. We aim to characterize political affiliation in language even when explicit political themes are not mentioned. It would be advisable to remind that this kind of author profiling could have some ethical issues, but the final goal would not be monitoring opinions expressed on the web.
Instead, we believe that public and open research on these topics would be helpful to show and make transparent for everyone what commercial systems -that often do not share their approaches with the scientific and the civil communities -can do with publicly available data.