Mining misinformation discourse on social media within the ‘ideological square’

Introduction

The advent of social media and social networks (both here described as social media; e.g. Twitter, Whatsapp, and Youtube) made it easier for users across the world to participate in publishing online content. Such content can be produced in several forms, such as writing (posts, tweets), speech (audio, videos), and images. Unlike printed media, social media content is not subject to editorial control in that users can be anonymous and posts do not need to be checked against credibility before publishing. Consequently, a large volume of information published and exchanged on social media, particularly news stories, may be inaccurate, incomplete, or worse, entirely counterfeit. Vosoughi et al. (2018) published one of the most influential studies that investigated about 126,000 news stories over 10 years on Twitter and found that fake news stories and false information spread much faster than real stories on every subject. Naturally, the immense flow of news exchanged every day between individuals makes the evaluation process difficult.

Driven by a natural desire to learn of and understand important events, many people are eager to keep up with the latest news, whether in general or on specific topics. However, given the ongoing deluge of news published and exchanged on social media daily, many users fall victim to misinformation. A host of misinformation is also circulated by social media users and goes unnoticed on social media, exposing other users to inauthentic content and risk. One explanation for users’ vulnerability on social media is that some of them may lack digital information literacy skills. Lazer et al. (2018) point out that the credibility of information is often unquestioned by individuals if it is not in violation of their preconceptions. As a result, their understanding, decisions, and behaviors may be affected. There are also discursive and linguistic features that may disguise manipulated information for ideological purposes. Such features may be available in their interaction with fake news as comments and replies. The present study seeks to explore how Arab social media users respond to fake news and the discourse structure of their responses utilizing van Dijk’s concept of the ‘ideological square’ (van Dijk, 2011). The ideological square emphasizes positive self-presentation and negative other-description. Critical discourse analysis is useful in learning more about the ideological underpinnings of fake news on social media while developing public awareness regarding credibility issues.

Misinformation: Definition and motivation

Fake news is currently a buzzword used to refer to the phenomenon of false or inaccurate news. In this paper, fake news is considered an instance of misinformation and is interchangeably used with it. In line with this, several attempts have been made to explicate the phenomenon. Wardle (2017) identified two concepts that have been broadly used to describe false news: misinformation, in which news is exchanged unintentionally, and disinformation, in which news is deliberately created to mislead people. She developed a typology of mis/disinformation and emphasized measuring the intent to deceive. However, there are several types of false news, each exhibiting subtle differences. Tandoc et al. (2018) conducted a review of academic studies published between 2003 and 2017 that used the term ‘fake news’ and identified six related terms often operationalized to mean fake news: satire, parody, fabrication, manipulation, propaganda, and advertising.

Of course, this typology is not clear-cut, and an overlap remains between all these categories of misinformation. Using this typology, Tandoc et al. (2018: 17) introduced a continuum that places the six terms/categories based on levels of facticity and immediate intention. Other researchers (see Lazer et al., 2018) have defined false news as ‘fake news’ and view it at the intersection of misinformation and traditional news media.

The spread of misinformation is not new to media. However, what is new is the active participation by non-journalists in publishing news stories (Wall, 2015) and the competition among users to post events on social-media sites (Hermida, 2011; Jewitt, 2009). This is further facilitated by the free access to and high interaction of users on social-media sites. Moreover, news stories on social media are generally short and have engaging titles. Examples of misinformation are Obama’s signing an executive order to investigate the 2016 election results and, at the time of Covid-19, false claims spread in India over social media stating that vegetarians were not affected by this disease (Chakrabarti, 2020).

There are several social, political, and technological factors that contribute to the spread of misinformation. For instance, Kavanagh and Rich (2018) point out that disagreements regarding facts and the indistinct delineation between opinions and facts have given rise to a state of ‘Truth Decay’. The creation and propagation of misinformation on social media have been principally informed by financial and ideological concerns (see, for instance, Allcott and Gentzkow, 2017; Au et al., 2021; Figueira and Oliveira, 2017; Lazer et al., 2018; Tandoc et al., 2018).

Misinformation and reader/user response

One of the consequences of the spread of fake news on social media is that readers’ responses have become ideologically polarized (Bakshy et al., 2015). According to Spohr (2017), social media users are ideologically polarized and may not be open to other views which explains their selective reading and sharing of fake news. Vicario et al. (2019) investigated fake news in Italian Facebook pages and found that users’ polarization and confirmation bias can be used to predict fake news topics with 77% accuracy. Guess et al. (2019) indicated that users’ age and partisanship are salient characteristics for the dissemination of fake news. Au et al. (2021) identified a three-stage mechanism of how online misinformation results in ideological polarization starting with creating online misinformation, spreading online misinformation, and polarizing the society. They argued that the spread of online misinformation is subject to urgency and cognitive bias which, in turn, spark a controversial debate polarizing the society. Using visual and textual features, Singh and Sharma (2022) developed a multi-modal framework to mitigate the spread of fake news and polarizing users with a prediction accuracy of over 81%.

Countering misinformation

Efforts to respond to misinformation on social media have recommended training or educating people to approach misinformation more critically and developing computational tools to control its spread. Information literacy (also, digital information literacy, media information literacy) has evolved as a valuable resource to equip people with lifelong learning skills that allow them to evaluate information or news stories in circulation. Information literacy (IL) encourages responsible citizenship in the information age and emphasizes critically evaluating information that is created, used, and shared in order to make an informed judgment (American Library Association, 2015; Chartered Institute of Library Information Professionals, 2018; UNESCO, 2005). Since the internet has accelerated the spread of misinformation on social media, several computational solutions, such as algorithms, have been proposed to help curb this diffusion. These algorithms often focus on the content, the diffusion dynamics, or a combination of both areas (Castillo et al., 2011; Figueira and Oliveira, 2017; Ratkiewicz et al., 2011; Vosoughi et al., 2017). Furthermore, concerned companies and platforms (e.g. Facebook, Twitter, Google) through which misinformation is published and exchanged, have announced plans to develop technical solutions to detect misinformation (see Crowell, 2017; Gingras, 2016; Mosseri, 2016). It is worth mentioning here that certain fact-checking websites such as politifact.com and snopse.com dedicate considerable effort to verifying false news stories. However, computational solutions that aim at the detection or identification of misinformation on social media face challenges in terms of semantic understanding, variation, and multimodality (Cao et al., 2018).

Misinformation and critical discourse studies

Misinformation is often triggered by important surrounding social, political, and cultural events. Such events are used to draw readers’ attention and are constructed as discourse, including audio/visual aids. As mentioned previously, misinformation can be motivated by political, religious, or other ideologies. That is, misinformation may be driven by promoting or resisting certain ideologies. Critical discourse analyses are concerned with the representations of power relations and the ideological meanings of events and sociopolitical or sociocultural issues (see, for instance, Fairclough, 1995; Gee, 2004; UNESCO, 2005). Critical discourse studies (CDS) are believed to be useful in studying the language of misinformation as a social practice to uncover ideologically related meanings. Critical discourse analysis has been adopted as both a theory and a method employed to examine language-use patterns associated with a wide range of social problems. Several examples of recent research that has applied CDA to study false news include (Creech, 2020; Green, 2021; Igwebuike and Chimuanya, 2021; Pasquim et al., 2020). These studies have examined how fake news was exploited to serve the predetermined goals of corporate and political parties. In the case of social media, fake news or misinformation can be seen as a form of manipulated discourse to draw readers’ or users’ attention to specific issues or gain their support for or against other issues (i.e. social, political, economic). Motivated by preconceptions, social media users may respond to fake news by expressing ideologically biased views and beliefs as replies or comments to show their support or opposition. CDS helps to illuminate the underlying ideological discourse structure of the responses through the discursive strategies employed by users as social actors who share sociocultural knowledge about public events and the social world.

Methods

The analytical framework of this study includes two elements: using a topic-modeling algorithm to generate implicit themes in misinformation followed by a critical analysis of misinformation discourse based on the sociocognitive approach (van Dijk, 2014; Van, Dijk, 2015; van Dijk, 2016). Topic modeling is a statistical method utilized to extract a set of probable terms that characterize a corpus or a collection of documents. It aims to uncover unknown patterns in large collections of texts to enable a determination of their underlying semantic meanings (Blei and Lafferty, 2009; Steyvers and Griffiths, 2007). As unsupervised machine-learning algorithms, topic models explore the distribution of topics over a corpus to return a set of terms divided into clusters. The method has been widely used in social science (see Grimmer and Stewart, 2013; Ramage et al., 2009; Valdez et al., 2018) to study a range of issues in media and sociopolitics.

Several approaches have been developed for topic modeling including latent Dirichlet allocation, latent semantic analysis, and non-negative matrix factorization (NMF) in which the focus is on identifying the frequency and co-occurrence of words in a corpus. This research applied NMF (Lee and Seung, 1999), factorizing input data into two matrices corresponding to words and the documents that contain them. The frequency of occurrence is calculated for each term, excluding elements that have negative values. Topic models have been recently used to study discourses from various perspectives including politics, literary studies, psycholinguistics, and media (see for instance Baturo and Mikhaylov, 2013; Griffiths et al., 2007; Mimno, 2012; Törnberg and Törnberg, 2016a, 2016b; Vaara et al., 2019). In Arabic, there are only a few studies that have focused on social and political issues on social media using topic models. For instance, Hamdi (2021) examined the construction of the concept of extremism among Arab Twitter users and pointed out that the extremism concept is under ideological control to serve users’ goals and interests. Jamal et al. (2015) investigated Twitter users’ attitudes toward the United States and Iran during 2012–2013 and found that Arabic discourses about the United States and Iran were negative due to intervention in national affairs. While these studies focused on analyzing users’ content to suggest their attitudes and views toward specific topics, the present study explores their ideological bias in response to basically false content.

The sociocognitive approach (SCA) to discourse places cognition at the interface between discourse and society and argues for the absence of a direct link between discourse and society (van Dijk, 2016). The cognitive component of SCA addresses both the mind and the categories of social cognition, such as knowledge, attitudes, and the ideology of a social group. SCA examines the discourse structure as influenced by the cognitive interface. The discourse structure can be divided into substructures that include phonological, semantic, syntactic, and rhetorical bases. The social component primarily addresses the power of dominant groups and the resistance of dominated groups (Van, Dijk, 2015). The present study focused principally on semantic structure, including the macrostructures and microstructures of misinformation. The macrostructural analysis addresses topics and themes that form the global meanings of text, while the microstructural analysis concerns the local meanings that connect propositions. The discursive strategies through which misinformation is published on social media and made to affect users’ perceptions are explicated. The selection of the two analytical tools is due to the need to reveal the latent topics underlying misinformation before departing to critically analyze the discourse structure of misinformation using SCA.

A major component of the SCA is the concept of ideological square (van Dijk, 1998, 2011) which features emphasis of positive self-descriptions and negative other-descriptions. Also, the concept involves mitigating one’s own negative properties and diminishing the positive elements of others. The resulting discourse structure is ideologically polarized between the positive and negative properties of ingroup (us) and outgroup (them). Some strategies that can be employed to fulfill the ideological square include the distribution of agency, blame transfer, and hyperboles.

Data collection and processing

Data were obtained from Alzanin and Azmi (2019), who developed a system to detect rumors in Arabic tweets. They collected topics of rumors from the Anti Rumors Commission (http://www.norumors.net/) and Ar-Riyadh daily (http://www.alriyadh.com/). The data comprised approximately 271,000 tweets, including rumors and non-rumors. It is important to note that Alzanin and Azmi (2019) defined rumors as statements that lack a verifiable source, and they considered only tweets with news information in their proposed system. The present study used only rumor tweets based on a subset of approximately 5800 tweets including replies collected from various Twitter hashtags initiated mainly by Twitter users and not official news outlets.

These tweets, whether news or conversations, were thought to imply ideologically loaded meanings concerning social, cultural, and political issues that were interesting to Twitter users as social actors and therefore appropriate for critical discourse analysis. Naturally, the topics identified were trending on Twitter during the time the data were collected. They serve as a starting point for the qualitative analysis of overall meanings and the discursive construction of misinformation.

The data were first cleaned by deleting ad tweets and highly frequent hashtags to exclude them from forming irrelevant topics. Typical preprocessing was then applied by using Natural Language Processing Tools (NLTK; Bird et al., 2009), Pandas, and Sklearn. The preprocessing included: tokenization, stop words removal based on Alrefaie (2017) and stemming by using ARLSTem (Abainia et al., 2017). These tasks aimed to reduce the size of the corpus in preparation for extracting latent meanings. The term frequency-inversed document frequency was applied to determine the distribution of terms across documents and their weighted values. Next, the NMF algorithm was applied using codes adapted from Pedregosa et al. (2011) and the number of topics was set to four, since the dataset was relatively small. The researcher translated example tweets into English for analytical purposes.

Analysis

After running the algorithm, four topics were returned, each with 10 terms. Figure 1 illustrates the distribution of terms across the topics. It can be seen from Figure 1 that the distribution of weight over terms seems normal in Topics 3–4, unlike 1–2. Several terms appeared in various topics. However, the four topics were affected by the frequency of some news and the semantic meanings of the terms. This is normal due to users’ interactions with particular news stories that were trending during the period of data collection.

OPEN IN VIEWER

Figure 1.

Topics and term weight distribution.

Thus, to improve the quality of topic selection and enrich the analysis, all of the topics were considered. This technique should help diversify the semantic structure for the subsequent qualitative analysis. The overall meaning of the four topics is motivated by users’ religious reactions toward the events and issues underlying the tweets. In the following, three pairs of tweets corresponding to three different elements of fake news are analyzed. Every tweet approaches the concerned fake news from a different ideological perspective. The schematic organization of Twitter as a genre of discourse is a hashtag that prefaces a topic or theme in which users are able to publish and cross-reference content.

Discussion

It is no surprise that the consequences of fake news have led to political polarization, decreasing trust in public institutions, and undermining democracy (Allen et al., 2020). The political polarization, as a consequence, was reflected in this study as ideologically polarized responses by users. These responses to fake news project Twitter users as two groups that express support or criticism toward the content. This polarity translates into ideologically loaded opinions and views underlying the responses using various discursive strategies as described above. The polarized construction of users’ responses corresponds to ingroup members attempting to construe positive self-presentation against outgroup negative presentation. The ideologically polarized discourse in interacting with fake news reinforces van Dijk’s concept of the ideological square.

Twitter users interacted with the three fake news elements without evaluating the credibility of each news story. Being under ideological control, they support or challenge the subject matter of the event to promote their groups’ positive schema against others’ negative schema. The discursive structure of their tweets also reproduces aspects of social and cognitive constructions. Within the SCA, the cognitive structure mediates the relationship between the social structure and the discourse structure through the mental representations of the participants. The following section further discusses the results in light of the SCA.

The aforementioned instances of fake news require Twitter users to activate mental models in the form of past experiences with similar events. These experiences include information (e.g. regarding events, settings, and participants). The occurrence of such events triggers mental models that translate into specific understandings and interpretations. The resulting discourses, whether supportive or otherwise, are influenced by previously existing mental models. For instance, the Twitter users in [5] and [6] seem to have already held opposing views regarding music in restaurants and cafes in addition to the social groups pushing for or against the alleged ban. When the ban issue was raised on social media, Twitter users responded in accordance with their views, signaling negative attitudes. These attitudes are controlled by ideologies. Twitter users may be seen to express primarily personal views. However, they are also members of specific social groups who share attitudes and ideologies as part of their social cognition. Discourses function dynamically to construe new ideologies or confirm existing ones. Thus, individuals are thought to unconsciously reproduce these ideologies in their discourses when relevant mental models are primed.

Identity and goals are two major categories defining social groups, which can be seen in [3] and [4]. In [3] Anan, being an elite military member is assumed to identify with the military and is considered unlikely to ally with another group against Sisi, who was also an elite military member. This event may be perceived as violating an institutional core and weakening ingroup members’ sense of belonging. Anticipating the negative effect on members’ perception of the group identity, the Twitter user in [3] tried to isolate Anan and project him as a dissident. Thus, his discourse seeks to maintain the group identity and power of military members, though he does not explicitly reveal whether he is a military member himself. A change in institutional identity may also suggest a change in goals and, here, Anan’s goals are thought to agree with those of an opposing or, rather, a harmful group. The user in [4] embraces the supposed change in Anan’s goals positively and considers it an attempt to serve his group’s interests and power. He encourages Anan to advance his new goals and asserts their legitimacy, which is likely to meet his group’s goals.

The instances of fake news in [1]–[6] demonstrate the opposing views of Twitter users in interacting with critical events on social media without considering their credibility. These events may not be seen as being informed by ideologies. However, ideologies ‘may very well be limited to a few basic principles’ (van Dijk, 1995: 140), which is the case in [1]–[6]. Within the SCA, ideologies are not to be viewed as positive or negative but rather as a shared social belief system that is evaluative in essence. Twitter users are also social members and have a shared belief system that defines their identities, relations, and interests. They judge world events based on their ideological system in that they maintain support for their positive schema against others’ negative schema, which is characteristic of the ideological square.

The spread of misinformation on social media is also a means to initiate ideological struggle between outgroup and ingroup members. For instance, the user in [1] addresses principally ingroup members to correct their mistakes and save the majority from forthcoming punishments. Misinformation can be exploited by certain individuals and groups to distract the public from important sociopolitical issues or incite hatred based on ideological backgrounds. In the data, many cases of fake news were often accompanied by a host of advertisements under the tweets or hashtags. These advertisements are motivated by marketing ideology, as pointed out by Lazer et al. (2018), Allcott and Gentzkow (2017), and Tandoc et al. (2018) among others. Indeed, it can be argued that a considerable amount of misinformation on social media aims at recruiting users to serve sociopolitical, religious, and marketing ideologies.

In social science, efforts to address fake news are going on in terms of the spread or diffusion, influence on social media users and the public discourse, and raising awareness of its vulnerability. Recognizing the risk and complexity of fake news, a group of 16 social scientists from several disciplines and institutions called for redesigning the information ecosystem for the 21st century (Lazer et al., 2018). Further, they proposed conducting interdisciplinary research to learn more about the phenomenon and develop solutions to curb its effects. Such proposals are consistent with Watts’s (2017) call for social science to be more solution-oriented. CDS can help to understand the public discourses and views toward various issues and events including false information. The discursive strategies along with the linguistic aspects may be utilized to discern false information language and, accordingly, develop advanced online detection solutions. In addition, the ideological square could help shed more light on the underlying ideological background of fake news propagandists through their language use since ‘one cannot always simply “read off” the underlying ideologies of a discourse’ (van Dijk, 2011: 387).

Conclusion

Social media have been misused by several parties (e.g. mainstream media, individuals, and unknown organizations) to propagate fake news. The dissemination of fake news is in many cases deliberate and motivated by ideological purposes. Two of the main reasons to publish fake news are to promote ideological struggle and exploit social media as a channel for marketing products or services (see, for instance, Allcott and Gentzkow, 2017; Au et al., 2021; Figueira and Oliveira, 2017; Lazer et al., 2018; Tandoc et al., 2018). This study contributes to the growing research on fake news discourse and ideological polarization by describing the discursive structure and strategies from a sociocognitive perspective. Unfortunately, some Twitter users interact with fake news stories without evaluating their credibility. Their discourses often feature an emphasis on positive self-presentations and negative descriptions of others, contrasting the adverse properties of readers against the writers’ positive ones. This is entirely characteristic of van Dijk’s concept of the ideological square. Discursive strategies, such as the disclaimer of blame transfer, denomination, hyperbole, and negative comparison have been employed to serve various functions of ideological polarization, self-identification, goal-description, and violations of norms and values. There were two main limitations in this study: the topics and issues were trending at the time of data collection and are thus by no means typical to the Arab world. Also, the responses belong to users from a few Arab countries and, thus, aren’t representative of the Arab world.