Sample

Five campaigns were selected, related to ideologically-motivated political hate speech propagated via Spanish-language Twitter. These matched four criteria selected by using the Delphi technique, which relies on reaching consensus in successive rounds of questionnaires and structured debates, in this instance of a panel drawn from various universities, involving ten academics and experts on communication, social networks, education, and sociology in the Spanish and Latin American sphere addressed by the research. They had to be:

• Representative cases, directed against governments, or public figures within them, that had a progressive or social ideology, considered left-wing.

• Campaigns whose motivation and objective were clearly political, their intention being to cause off-line mobilizations in society against these governments.

• Mass campaigns, defined as having more than 100,000 tweets related to a given hashtag, which would distinguish them from traditional political campaigns on a smaller scale.

• Campaigns making extensive use of false accounts, with an eight-digit reference. This is because, when large numbers of accounts are created, Twitter by default assigns them a user identity ending in eight figures, needing human intervention to change it. Likewise, campaigns using many recently created new accounts, in existence for under one year (Luque et al., 2021).

The five instances chosen were the following. An idiomatic English version of the title of each is given in square brackets, together with a brief explanation.

• #SáncheVeteYa (Spain) Note spelling error, Sánchez with missing Z. [Sánchez Must Go – refers to Pedro Sánchez, leader of the PSOE (Socialist party) and Spanish Prime Minister]

• #IglesiasVeteYa (Spain). [Iglesias Must Go – refers to Pablo Iglesias, leader of the left-wing Podemos party and Deputy Prime Minister]

• #AndrésNoMientrasOtraVez (Ecuador) Note spelling error, Mientras for Mientas [lie], with added R. [Andrés Stop Lying – refers to Andrés Arauz, presidential candidate of the centre-left UNES coalition]

• #GolpeDeEstadoK (Argentina). [K’s Takeover – refers to Cristina Fernández de Kirchner, Vice-President from the left-wing TODOS coalition]

• #FraundeEnMesa (Peru) Note spelling error, Fraunde for Fraude [fraud], with added N. [Electoral Fraud – refers to allegations about the election of Pedro Castillo as President by a narrow margin].

Apart from the criteria noted, a further reason to select these campaigns was that they took place between May 2020 and June 2021, coinciding with the Covid pandemic and the ensuing enhanced use of social networks (Cervantes & Chaparro, 2021). Three of them were specifically picked out because of their spelling errors, which were replicated identically by at least 20,000 accounts, this being an indicator of the use of bots or automated accounts (Puyosa, 2017), as employed in deliberate political campaigns. After all, it is out of the question that 30,000 individuals would make the same spelling mistake at the same time (Calvo et al., 2019; Persily, 2017; Vargo et al., 2017). In all, 1,442,781 tweets and retweets (RTs) were collected directly through the Twitter Application Programming Interface over different periods of time.

Research tools

T-Hoarder was used to download the tweets linked to the selected hashtags, using the open-code T-hoarder_kit method, which fulfils requirements for objectivity and transparency (Congosto, 2017). Thereafter, T-Hoarder and its algorithm were utilized to classify tweets and their relational metadata, on the basis of activity and impact (Villodre et al., 2021), along thematic lines using three orientations: time, space and relevance (Congosto, 2017). This first phase provided the foundations for using the Gephi program, an open-source interactive tool for visualizing and studying large network graphs and complex systems, looking for patterns and trends (Cherven, 2013). Gephi was employed to analyse the frequency and impact of messages, so as to undertake graphic modelling according to total RTs received, with spatial ordering using the ForceAtlas2 algorithm.

A selection of the most representative data obtained with T-Hoarder was then put through the MAXQDA Analytics Pro program (Release 20.4.1) to achieve a more qualitative approach. A first filtering of the messages picked out those containing terms repeated more than 200 times. Focus was then concentrated on terms related to hate speech (scorn, rejection, humiliation, harassment, disparagement or expressions of hatred towards individuals because they belong to a given group). The “Word Combinations” tool of MAXQDA was employed to determine which words linked to the most frequently repeated terms appeared or contributed to highlighting this hate speech. Finally, the “Keyword-in-Context” tool served to relate the selected terms and the words connected to them in their actual context, allowing localized analysis of the production of the discourse in question. The schematic for this procedure may be consulted through the hyperlink: https://doi.org/10.6084/m9.figshare.16649467.v4.

Results of scrutiny with T-Hoarder and Gephi

Five graphs produced with Gephi are presented below, one for each of the five campaigns studied. To create and shape each diagram, the following elements were taken into consideration: (a) the number of nodes, that is Twitter accounts looked at; (b) the number of edges, in other words interactions between accounts; and (c) modularity, in that the percentage of interactions is used to establish communities, understood as more cohesive subsections of the graph, or groups of nodes more strongly interconnected one with another, which are shown in the same colour in the graphs.

The spatial ordering of these nodes was achieved by means of the ForceAtlas2 algorithm, which takes into consideration a range of functions in performing this spatial distribution. Two of particular note were: (a) closeness, which sees a node as nearer to another the higher the number of interactions there are between them, and (b) intermediation, putting a node between two others when it has interacted with them, so that, for example, if a node A has interactions with two nodes, B and C, it is placed on the graph in the space between them. Figure 1 shows two large clusters. The right-hand grouping has many more nodes, edges and communities, and a greater degree of concentration, implying that there is a considerable percentage of interactions between the accounts. In contrast, the left-hand grouping is much more diffuse and is located at a distance from the right-hand community because of its eccentricity, in other words its lack of interactions with this other cluster, as its percentage of interaction is much lower. Because of its diffuseness, the left-hand cluster is of quite some spatial extent, even though the number of accounts in it is much smaller. This indicates that there is no great co-ordination among them, unlike in the right-hand cluster.

The distance between the two communities clearly demonstrates polarization, as the left and the right have virtually no interactions. In the right-hand cluster a contrast may be observed between the upper half, comprising the communities coloured yellow, orange, blue, and greenish blue, where there are no nodes with any significant degree of retweeting, and the centre and lower half, where there are concentrations of accounts with a major number of RTs. The accounts seeming largest in size can thus be interpreted as an opinion matrix, the structure for political communication creating opinions and receiving many RTs ¹ , as distinguished from the dissemination matrix, which is the set of accounts spreading messages, often false, automated or trolling centres, their names not appearing on the graph, signifying that they send RTs but do not receive them. Finally, an appendix to the cluster may be observed to the lower right, separated by reason of its limited interactions. This represents a community of Venezuelan accounts not interacting with the others, and goes to show the international reverberation of this campaign.

Of the visually largest accounts, the matrix of opinion in the right-hand cluster, several stand out, particularly that of the tweeter with greatest impact in this campaign, @juanfraescudero, Juan Francisco Escudero, a person closes to the far-right political party Vox. Similarly, there are Hermann Tertsch (a Member of the European Parliament for Vox, and an associate of the Floridablanca Network, a foundation belonging to the Atlas Network), Macarena Olona of Vox, Luis del Pino (a member of the Hazte Oír association and writer for the Libertad Digital online newspaper, both organizations linked to right-wing, extreme right-wing and ultra-Catholic sectors) or the anonymous tweetstar using Willy Tolerdo as an account name.

Figure 2 looks at the campaign launched by the losing candidate for the presidency of Peru, Keiko Fujimori, during one of her press conferences. It is possible to observe two large clusters, clearly polarized. On the left there is a single community complaining about the campaign, or mocking it, because there was no proof of any electoral fraud. On the other side, two main communities can be seen, with the accounts receiving most retweets lying at the bottom of the chart, the most prominent owned by a presenter from Willax TV, Diego Acuña (@diegoacuoficial). This is at one end of the graph, since many accounts interact only with this account, plus a few accounts in its community. Willax TV spread fake news about supposed electoral fraud in Peru (Quesada and Fowks, 2021). During the campaign, it made the claim that Pablo Iglesias had travelled from Spain to Peru to help Pedro Castillo. This hoax was also put around by right-leaning Spanish mass media, like esRadio, and led to demonstrations in front of the hotel where Iglesias was alleged to be staying (Newtral, 2021).

Figure 3 shows a single cluster without polarization, in which the accounts receiving most retweets are in a central position, indicating that these come from various communities, on the basis of the centrality of intermediation. This is the opposite case from the previous instance. Although there are various different communities, as indicated by the colours, they form one group and the same action front for spreading messages favourable to the man who, at that time, was a candidate for the presidency of Ecuador, Guillermo Lasso (@lassoguillermo), and against Andrés Arauz, the rival candidate. One prominent account is that of Carlos Andrés Vera (@polificcion), a member of the Ecuador Libre foundation, part of the Atlas Network, headed by Guillermo Lasso himself. Other noteworthy accounts belonged to Fernando Villavicencio, the director of the Periodismo de Investigación news service that, together with Carlo Andrés Vera, spread videos, later shown to be doctored, of the candidate Andrés Arauz during the electoral campaign that went viral.

Figure 4 is the graph covering the campaign relating to a claimed “Kirchnerist” putsch, the adjective being based on the married surname of the Vice-President of Argentina, Cristina Fernández [de Kirchner]. It is feasible to identify a large cluster made up of various communities and an outlier community polarized against it.

The accounts with the most RTs are fairly central, receiving interactions from the range of communities forming a common front against the Argentine government. Almost all the more prominent accounts are anonymous, like @2023macri or @chauoperetak, or they are of digital media spreading fake news and hate messages, like @laderechamedios or the account of the digital journalist Eduardo Prestofelippo “El Presto” (@elpresto2ok), who was arrested for online threats to kill the Vice-President. All of these are in the media ambit of the foundations, supported by Atlas Network, promoting Javier Milei and José-Luis Espert, right-wing libertarian political candidates, in Argentina, and politicians and groupings like Bolsonaro, Trump and Vox elsewhere.

Figure 5 shows a single cluster of several communities with scarcely any polarization, where the vast majority of accounts that stand out are on the right-hand edge and quite close to one another. The campaign was directed against the man who, at that moment, was Deputy Prime Minister in the Spanish Government. This was a campaign by right-leaning parties against the Government, but focused principally on the most radical part of the coalition, rather similarly to what happened in the Argentinian case. In this attack, certain anonymous accounts stand out, for instance @frayjosepho, @nicobolivariano or @guajesalvaje, as do the accounts of the far-right Vox and right-wing Partido Popular parties, such as the latter’s online communications director Ismael Quesada, as well as @pablocast13, a member of the same party’s youth movement. Likewise, there were journalists or media of the same tendency, such as Carlos Cuesta Arce (who has worked for the online OkDiario) or the Caso Aislado news website, both known to have published false news.

Results of scrutiny with MAXQDA

The next sections lay out the results of an analysis of the tweets selected on the basis of the frequency of words most often repeated (used more than 200 times). A visual representation was produced in the shape of a word cloud, as a function of frequency of use (Ballestar et al., 2020). Word clouds are a widely-used method because of their efficiency in summing up large quantities of data and giving an impression of the ideological views that lie behind a textual discourse. Finally, the results from the keywords in context study are presented.