Emotional community and concerted action: on the emotional mobilization mechanism of disinformation in the Anti-extradition Law amendment movement in Hong Kong

3.1 Data collection

One main feature of the incident is its international reach, with protesters using a wide range of social media platforms such as Facebook, YouTube, Twitter, Telegram, Instagram, and LIHKG for the purpose of social mobilization. Gaining support from international organizations and foreign dignitaries became an aim for the protesters and a tool for them to prove the legitimacy of their movement and deepen their social mobilization. Preparatory research revealed that Twitter was the main public platform for international propaganda and social mobilization in Hong Kong Anti-extradition Law Protests worldwide. Therefore, we chose the data of Twitter as the object of our study. However, it is necessary to note that multi-platform coordination is a typical strategy for disinformation dissemination. Fact-checking used in this paper targets not only Twitter, but the same disinformation jointly disseminated by multiple platforms. Therefore, the selection of Twitter data is not intended to underestimate the influence of other platforms. The specific steps are as follows.

First, taking “‘香港’ OR ‘Hong Kong’ OR ‘HK’” as the keyword, we used the crawler to obtain the original tweets with more than one retweets, likes, and comments respectively on Twitter between January 1, 2019 and January 31, 2020 as data sample A (a total of 387,010 tweets and 102,907 accounts). Then, we retrieved fact-checked information about Hong Kong through Google Fact Check Explore (a collaborative project between Google and Duke University, the largest fact-checking tool in the world), and obtained a total of 76 pieces of fact-checked information excluding those unrelated to Hong Kong Anti-extradition Law Protests, to obtain the verified disinformation event set E. After that, based on the fact-checked information in the disinformation event set E, we traced the content of the original disinformation text, and defined keywords and Twitter search statements (including traditional Chinese, simplified Chinese and English) for each disinformation. A time series autocorrelation analysis was conducted on data sample A, and the results showed that the autocorrelation impact of topics at a certain point was about “the following 7 days.” Therefore, for data completeness, this study retrieved information posted 3 days before and 7 days after the release time of the original disinformation, and obtained the disinformation-related tweet data sample R (a total of 78,908 tweets).

Second, keyword data of the disinformation event set E was used to test the similarity between the tweet data sample A and the disinformation-related sample R based on word embedding, so that each tweet data (from R + A) corresponds to a disinformation event number (from E), and samples with low similarity scores were removed.

Third, the top 30 % of the aggregated disinformation events with the highest similarity scores (excluding samples with low content relevance) were selected as a larger disinformation-related data set RA (92,262 items) with an average similarity of 0.74. The data sample RA was the main research object (as shown in Figure 1).

Figure 1:

Data collection process.

3.2 Content analysis

According to the disinformation event set E (n = 76), disinformation in the set is characterized by cross-platform dissemination and high engagement. For example, a head CT scan of a person claimed to have been beaten and injured by the police was widely spread on multiple social platforms. The event set also contains a large number of events that are directly related to the protests and have a sustained impact, such as the “831 Prince Edward Station Incident,” which was used by protesters as an excuse to organize several large-scale rallies. The content subjects of the disinformation were manually coded into seven categories, and the emotional polarity they triggered were divided into positive and negative. In order to classify the offline movements during the Hong Kong Anti-extradition Law Protests period, this study collected information from the Wikipedia list of Hong Kong Anti-extradition Law Protests, the South China Morning Post (Robles 2020), and other media as the main sources. The timeline of the movements from June 1 to December 31, 2019, with the time, location, and form of each movement, was used to classify the daily concerted actions into four types (as shown in Table 1).

In addition, the study further screened the event samples for theme coding and emotion analysis of disinformation from the protesters’ camp based on the content subject and emotional polarity of the disinformation event set (as shown in Table 1). Two coders independently coded the above items, and Cronbach’s alpha coder reliability is shown in Table 2.

All variables had high reliability due to the distinct features. Finally, based on word embedding similarity calculation between the disinformation event set E and the Twitter sample, the above items also became attributes of the Twitter sample RA.

3.3 Emotion analysis

To study the differences in emotional types evoked by disinformation, this study adopted a text emotion analysis method based on an emotion dictionary. The NRC Emotion Lexicon, created by the National Research Council Canada, is an authoritative emotion lexicon commonly used by the international academic community. It provides eight basic emotions: anger, fear, anticipation, trust, surprise, sadness, joy, and disgust, as well as two emotional polarities: negative and positive (Mohammad and Turney 2010, 2013). The NRC Emotion Lexicon is mainly in English, but a Chinese translation is also available, which was used for emotion analysis of Chinese texts, with manual adjustment for individual translation deviations to improve the accuracy of the lexicon. The above processed NRC Emotion Lexicon was used to calculate the word frequency of each emotion type in the English and Chinese texts of the data sample RA, and the proportion of the word frequency of each emotion type in each text was taken as its score.

3.4 Time series analysis

In this study, time series analysis was adopted to explore the mechanism of emotional mobilization of disinformation. For the correlation analysis, the Spearman correlation coefficient between the time series of the variables was calculated as the basis of correlation. For the causality analysis, the nonlinear causal network analysis method PCMCI (Runge et al. 2019a, 2019b) was mainly adopted to calculate the causal relationships within n-period lags of multiple time series with each other (n is the maximum number of lags) and construct a two-way causal network to explain the complex interactions among the variables.

In terms of data processing methods for the samples, this study mainly calculates two types of time series data: (1) time series data of communication effects, obtained by mean aggregation of communication effect indicators (retweets, comments, likes) for each tweet data in each day and hour; and (2) time series data of emotion change, obtained by mean aggregation of different emotion types’ scores of each tweet in each day and hour.

In addition, considering the rigor and efficiency of data analysis, some methods and parameters need to be adjusted: (1) Narrow the time granularity. For example, the smallest time granularity for protest action is “day,” while that for emotion and communication effect is “hour”; (2) Select the maximum lag time. The maximum number of lags is 7 when the data is in “days” and 24 when it is in “hours”; (3) Residualization. Before causality analysis, it is necessary to calculate residual series of time series data (eliminate the influence of cycle and trend based on the X-11 model and STL model), to avoid the problem of “false regression” and improve the explanative power of causality (Cai 2009; Wang 2008); and (4) The significance indexes are taken as less than 0.05.

4.1 Polarized camps: analysis of the content subjects of disinformation during the Anti-extradition Law Protests in Hong Kong

The content subjects of disinformation include: Hong Kong protesters (PROTESTER, 47.37 %), Hong Kong police (HKP, 23.68 %), the central government of China (GOC, 11.84 %), HK Pro-establishment camp (PEC, 9.21 %), Hong Kong Special Administrative Region Government (HKSAR, 6.58 %), Chinese-funded enterprise (CE, 1.32 %). In general, protesters and Hong Kong police are the major content subjects of disinformation, showing a clear trend of polarization. It is not surprising that protesters are one of the subjects of disinformation, considering the subjects of any social movement will disseminate various messages to construct their images. In terms of the specific demands of the protest against the extradition bill campaign, the protesters mainly targeted at the Hong Kong Special Administrative Region Government and the central government of China. However, in the disinformation, Hong Kong police, rather than the above-mentioned two governments, become the main target of attacks. The police-versus-civilians not only creates a stronger visual effect of confrontation, but also provides more sufficient narrative tension in terms of connotation.

4.2 Constructing an emotional community: analysis of narrative theme types of disinformation during the Anti-extradition Law Protests in Hong Kong

By manually summarizing the themes of the 76 pieces of fact-checked information, we found that this disinformation can be grouped into seven types (the specific analysis process is shown in Table 1: Types of themes). In terms of the distribution of theme types, “white terror” (31.38 %) is the most prevalent theme, followed by “personal image attacking” (23.03 %), “police brutality” (22.12 %), and then “sad victim” (13.71 %), “alliances building” (7.18 %), “blame shifting” (2.44 %), and lastly “local identity” (0.14 %).

As shown in Figure 2, there are differences in the types of emotions embedded in the different theme types. For example, “white terror” and “police brutality” highlight the emotions of fear and anger. The percentage of fear in “white terror” is as high as 20.29 % and the percentage of anger is 16.91 %, while the percentage of fear in “police brutality” is the highest at 26.77 % and the percentage of anger is 18.44 %. This narrative theme, dominated by fear and anger, visualizes the “external enemy” of the protesting group, and allows the protesters to rationalize and legitimize their actions and maintain the momentum of the group’s common action.

Figure 2:

Narrative themes of disinformation and the distribution of embedded emotion types.

In comparison, the themes of “alliance building” and “local identity” both have a high percentage of positive emotions. In addition, “sad victim” also has a high percentage of trust, which accounts for 20.44 %. The positive emotions of trust, expectation and pleasure are prominent in the narrative themes to strengthen the group’s sense of identity and belonging by shaping the identity of the protesters and constructing collective memory. The theme of “sad victim” is to construct the collective memory of the protest movement, and to continue and strengthen the emotional ties of the protesters’ group through commemorative discourses and rituals. It can be seen that the protesters have gained a collective emotional experience through the widespread dissemination of this emotion and have formed a community of emotions through group representations such as retweets, comments and likes. This form of emotion-centered community is a new evolutionary trend of the “imagined community” based on social media platforms, as social media provide the possibility for the instantaneous transmission of emotions across regions, groups and cultures. It is in this sense that Michel Maffesoli has developed the concept of affectual tribes (Maffesoli and Foulkes 1988), which have become a metaphor for the emotional community that not only forms the social basis of post-modernity, but also contributes to the creation of a zeitgeist (Xu and Petit 2014). The emotional community that the protesters tried to construct is an “affectual tribe” that is linked by collective emotional experience and characterized by communication behavior on social media platforms. With the widespread dissemination of this emotion and the expansion of the community, the protesters perceived that the group is supported by the external society and that their behavior is in line with the social expectations, thus gaining a sense of empowerment and thus reshaping their selves. Therefore, these themes played an important role in enhancing the internal cohesion of the protesting groups and increasing their sense of participation.

4.3 Expanding the “imaged alliance”: a correlation between the type of emotion embedded in disinformation and the effect of online communication

For social movements, in addition to strengthening identity and forming an emotional community within the protesters, it is also necessary to find potential sympathizers and supporters in the international arena to further expand “alliances.” In the case of Anti-extradition Law Protests in Hong Kong, disinformation also plays an important role in this regard. This information is a projection of people’s political tendencies and deep-seated desires as well as a reflection of their emotional purposes (Allport and Postman 1947). Interactive behaviors such as liking, commenting and retweeting can not only quickly promote the relevant information, but also in effect form and develop an “alliance” of protesters based on such communication behaviors. Thus, the evaluation of the data on these communication behaviors becomes an important way to analyze the mobilization of protesters by setting different emotional frames for different subjects. To do so, we first analyzed the subjects in the disinformation network.

Based on the sample R of Twitter data related to disinformation (a total of 78,908 tweets), we statistically analyzed the relevant subjects involved in the dissemination of disinformation. By summing up the three types of data of retweets, comments and likes, we ranked these accounts (see Table 3: Head accounts of disinformation dissemination. In the interest of saving space, only the top 20 accounts were selected). From these influential accounts, it can be seen that the accounts involved in spreading disinformation about the Anti-extradition Law Protests in Hong Kong are relatively diverse, including Hong Kong media platforms, organizations and individuals, as well as international mainstream media, political figures, NGOs, and some social activists. According to the data, tweets published by these head accounts were also transmitted widely. This shows that during the Anti-extradition Law Protests in Hong Kong, the disinformation “captured” many international mainstream media, famous politicians and KOLs, forming an international communication matrix, and thus creating a larger communication effect and forming an international “alliance.” This internationalization of mobilization has a special significance in the Anti-extradition Law Protests in Hong Kong. In the protesters’ view, the international community’s stance and support can, on the one hand, provide proof of the movement’s legitimacy, and on the other hand, provide an opportunity to mobilize public opinion in Western countries and motivate anti-Chinese politicians to push for the introduction of relevant Hong Kong-related bills, which is the reason why the opposition attaches great importance to Twitter and actively promotes international propaganda through the platform although Twitter is not as popular as Facebook in Hong Kong.

The strategies protesters use to mobilize a wide range of international “alliances” is extremely detailed. As seen in the specific practices of Anti-extradition Law Protests, protesters not only designed different emotional frameworks for different subjects, but also released various information depending on the type of action and the stage of the movement. To this end, we compared the correlation between the communication effects (retweets, comments, and likes) and the emotional types of the disinformation targeting at the protesters, Hong Kong police, the central government of China, the pro-establishment legislators, and the Hong Kong SAR Government under two different types of actions: violent and non-violent. In order to know the dynamic communication effect, we also adopted time series correlation analysis for the combination of the above variables and retained significant inter-variate relationships (p < 0.05) for data visualization (e.g., as shown in Figure 3).

Figure 3:

The correlation results of disinformation in communication effect and sentiment. Note: only the data of the significant part (p < 0.05) are shown in the figure.

The analysis found that the relationship between the emotion type of disinformation and the communication effect was different under different conditions: (1) under the conditions of violent and non-violent actions, disinformation expressing sadness and disgust generally achieved better communication effect; (2) when targeting at the HKSAR Government, the relationship between positive and negative emotions in the disinformation and the communication effect was sharply polarized, disinformation with negative emotions achieve better communication effect and the opposite is true for disinformation with positive emotions; (3) when targeting at protesters, disinformation with positive, pleasant and surprising emotions achieve better communication effect.

The above analysis shows a correlation between the type of emotion embedded in disinformation and the communication effect, but it is insufficient to explain the interaction between the two. For this reason, we conducted a non-linear causal analysis of the same data sample based on time series (as shown in Figure 4). To sort out the causal network more clearly, we used emotion type and communication effects as independent and dependent variables, respectively, and divided the lagged period of causal influence into early, middle, and late periods (early ≤ 6, 6 < middle < 18, and late ≥ 18 with hourly time granularity) for two-way causal presentation (as shown in Figures 5 and 6).

Figure 4:

24-period lagged time series causality network between the emotion type embedded in disinformation and communication effects under different conditions. Note: the nodes of the network are various variables, the connected edges of the network indicate the existence of a causal relationship between two variables, the point of the arrow indicates the influence of the independent variable on the dependent variable, the number of lags of the connected edges indicates the existence of the causal relationship, blue indicates negative causality, red indicates positive causality.

Figure 5:

Causal relationship when emotion type is the independent variable. Note: triangular indicates positive causal influence, inverted triangle indicates negative causal influence, and the number below the image indicates the value of the causal association, and all causal correlations shown in the figure are significant (sig < 0.05).

Figure 6:

Causal relationship when communication effect is the independent variable. Note: triangular indicates positive causal influence, inverted triangle indicates negative causal influence, and the number below the image indicates the value of the causal association, and all causal correlations shown in the figure are significant (sig < 0.05).

The analysis revealed a clear difference in the order of causality between the emotion type embedded in the disinformation and its communication effect in violent and non-violent actions: (1) under the condition of violent actions, the early-stage communication effect as the independent variable mainly drove the negative emotion (mainly aversion); (2) under the condition of non-violent actions, the positive emotion (mainly expectation and trust) in the early stage promoted further transmission of disinformation.

In addition to emotional factors, do other content factors (e.g., the subject and theme contained in the content) also have an impact on the communication effect? In order to further analyze the mechanism of the effect of online communication, this study continued to include the factors of the subject and theme contained in the content to construct a causality network (as shown in Figure 7). The variables concerning subjects and themes are also time series obtained by aggregation calculations. The results show that there is a correlation between the amount of content containing individual subjects (Hong Kong police and pro-establishment legislators) and the emotions embedded in the disinformation. For example, the more content about the Hong Kong police, the more pronounced the fear embedded in the disinformation, and the opposite for pleasure. Individual subjects (blame shifting) can also have an impact on the emotions embedded in disinformation. However, neither the subjects nor the themes contained in the content have a direct effect on the communication effect itself. It is the emotion embedded in the disinformation that directly affects the communication effect. This indicates that the influence of emotional factors on online likes, comments and retweets is more significant.

Figure 7:

A time-series causal network of subjects, themes, emotions and online communication effects contained in disinformation.

It is thus clear that emotion is the direct and significant influencing factor in communication effects of disinformation, and content with emotional arousal may be more likely to proliferate. The types of emotions embedded in disinformation under different conditions have diverse relationships with the communication effects.

4.4 Concerted action: an analysis of the correlation between the type of emotion embedded in disinformation and the type of concerted action

The mechanism by which disinformation acts on social movements lies in its emotional mobilization, so what is the relationship between the different types of emotions embedded in disinformation and the forms taken by actions? In order to answer this question, an emotion analysis of disinformation in different action types is required. The daily average scores for each emotion type were calculated and Kruskal-Wallis tests were performed. The results are shown in Table 4: during the different types of actions, there were significant differences in anger, fear, and sadness in the negative emotions of disinformation (p < 0.05), and in expectation and pleasure in the positive emotions (p < 0.01). However, the overall level of significance cannot explain specifically which types of actions show emotion differences in their comparisons, so pairwise comparisons are also needed.

Looking further at the distribution of daily mean scores for each type of disinformation emotion (as shown in Figure 8) and the results of pairwise comparisons (shown in Table 4), important differences can be found: different from peaceful marches, anger is more prominent in disinformation disseminated during concentrated violent actions, and fear is more prominent in disinformation disseminated during decentralized violent actions; while compared to decentralized violent actions, the positive emotions of expectation and pleasure are more prominent in the disinformation during the peace march.

Figure 8:

Average daily distribution of each type of emotion during different types of actions.

How can we understand the distinct differences in the emotions embedded in disinformation during peaceful marches compared to decentralized or concentrated violence? The emotion types and their intensity underlie social movements from emergence to decline, and emotions not only influence the various phases of the movement but are also key to the fight (Yang and Pace 2016). Specifically, emotions not only strengthen the sense of group identity and belonging, expands sympathizers and supporters, but also may shape the group’s perception of the target of phased action (Jasper 1998). Research in political psychology have also found that anger and fear drive very different political actions (Brader and Marcus 2013): anger, which is usually attributed to a specific source from which individuals perceive themselves to be suppressed, is more likely to form unifying goals or antagonistic enemies and to increase intra-group identity and solidarity (Nepstad 2004); in contrast, however, fear, which usually points to an unknown outcome, is more of a negative emotion that lacks a sense of control. Thus, when movement participants have a clear target to blame, attention is unified and anger is directed toward hatred of the “external enemy” and the “sense of justice” that comes from attacking the “enemy.” The more likely outcome is concentrated violence. However, it was found in this study that overall fear was more prominent than anger (as shown in Figure 8), with higher levels of fear embedded in disinformation when violent actions happen simultaneously in multiple zones. This is a time when movement participants are less unified and lack control over their attack targets and are more likely to engage in decentralized violence. In contrast, positive emotions such as expectation and pleasure are more prominent during peaceful actions. Peaceful actions, i.e. marches and rallies, are used to disseminate the demands of the protesting group, focusing on the “ideals” and “values” of the group, to gain external sympathy or support, and at the same time reinforcing the sense of belonging and identity of the community through rituals. Peaceful action has a special significance for sustaining the movement itself. Goodwin and Pfaff argued from a theoretical perspective of emotion management that when collectives face fear deterrence, movement participants mobilize “motivation mechanisms” (e.g., peace rallies) to “manage fear,” and movement activists strive to build emotional solidarity and a sense of collective identity (Goodwin and Pfaff 2001).

However, it is difficult for difference statistics to reveal the direction of action among various factors. To further analyze the dynamic relationship between the emotions embedded in disinformation and offline activities, this study continued to construct a time-series causality network. In this regard, the various emotions embedded in disinformation remain aggregated by date for the scores, while the variables on violent and peaceful actions are aggregated by date for the number of occurrences of the actions. The results, as shown in Table 5, show a significant current effect between emotions embedded in disinformation and offline activities: emotions of anger and fear embedded in disinformation promote the occurrence of violent activities. Conversely, positive emotions have a negative effect on the occurrence of violent activities, which echoes the analysis above.

In conclusion, through social media platforms, disinformation, together with the narrative themes produced by various types of emotions, can construct protesters’ imagination of the community and maintain the emotional dynamics of action by establishing group identity and boundaries and strengthening the sense of identity and belonging of the group, which demonstrates the social integration role of emotions. The emotions embedded in disinformation not only influence the effect of online communication, but also influence the development of collective activities offline, which has a mobilizing effect.