Examining the escalation of hostility in social media: a comparative analysis of online incivility in China and the United States regarding the Russia–Ukraine war

2.1 Definition and categorization of online incivility

Originating from early modernity and the European Renaissance (Flores et al. 2021), incivility in the political discussion has been classified into two categories: (1) personal-level incivility and (2) public-level incivility (Muddiman 2017). Personal-level incivility, which includes behaviors such as swearing, name-calling, vulgarity, or aspersion (Coe et al. 2014; Cortina et al. 2001; Luparell 2011), highlights a tendency to avoid face-threatening acts (Muddiman 2017) and consider civility as the proper manner and tone in which interactions are carried out in everyday life (Boyd 2006). The public-level incivility, conversely, refers to behaviors involving violation of political and deliberative norms, which includes behaviors like prevent a respectful and inclusive discussion, refusing to recognize others’ views, and putting forward personal benefit in the political discussion (Muddiman 2017; Stryker et al. 2016). This differentiation is informed by Papacharissi’s (2004) distinction of incivility and impoliteness. She argues that incivility requires threats to democracy or others’ rights as opposed to merely improper communicative behaviors labeled as impoliteness. This definition is widely applied in political communication studies, defining incivility as a mode of offensive discussion that impedes the democratic ideal of deliberation or threatens democratic pluralism (Anderson et al. 2014; Rossini 2020).

This research will employ the definition of incivility from Stryker et al. (2016), which categorized it into three types: (a) utterance incivility, such as name-calling and swearing; (b) discursive incivility, which inhibits a respectful and inclusive discussion, for example by interrupting someone with a differing viewpoint; and (c) deception incivility, which focus on the outright lying, exaggeration of the fact, and providing misleading information. This definition combines both the linguistic and the political attributes of online incivility. Previous studies on platforms like X (Twitter) and Weibo have demonstrated the presence of incivility in political discussions on these platforms (Song and Wu 2018; Trifiro et al. 2021). Therefore, this study proposed the following research question.

2.2 Media attributes of online incivility

2.3 User-central metrics attributes of online incivility

Considered one of the most related affordances to explain the formation of online incivility, the anonymity of the online platform allows the users to express their opinion uninhibitedly without considering any consequences (Rossini 2020). Conceptualized by Suler (2004), the online disinhibition effect reveals the discrepancy between individual languages in cyberspace and the natural world and explains how it impacts the formation of incivility. Suler introduced six factors that could explain two types of online disinhibition effects, positive/benign and negative/toxic. The positive/benign type refers to the “unusual acts of kindness and generosity that users express online (Suler 2004)”, while the negative toxic type refers to incivilities such as rude words or harsh criticisms that the user would not use in offline communication. As one of the six major factors, dissociate anonymity describes when people separate their real-life roles and identity and their online behaviors and hide some or all their identity. Such anonymity will decrease the sense of vulnerability about self-disclosing and being free from the responsibility of any toxic expression. Other literature also considers anonymity could foster online incivility. Experimental research discovered that the perceived credibility of commenters and the persuasive quality of public discussion could be negatively influenced by incomplete identity cues and uncivil commentary (Wang 2020). However, a study on the incivility of the X (Twitter) post of three democratic primary candidates during the 2020 election demonstrates that anonymity is not related to the incivility level of the post (Trifiro et al. 2021).

Apart from anonymity, another user-level factor that will lead to the formation of online incivility is sentiment. Operationalizing online incivility as swearing language, Song and Wu (2018) studied incivility in the Weibo discussion of the Hongkong–Mainland China conflict, discovering that negative emotions with high arousal levels are among the most influential factors in predicting online incivility. Stephens and Zile (2017) also studied the connection between negative emotion, high arousal levels, and incivility. After experimenting with games with different emotional arousal levels and negativity, Stephens discovered that swearing fluency significantly increased after playing the game with high emotional arousal and high negativity.

Therefore, apart from the media frames as the primary independent variable, the research also proposes negative sentiment as the variables that would impact online incivility:

Meanwhile, the study proposes that the user-centric factor can moderate the positive relationships between negative sentiment, conflict frame and online incivility. Thus, this research proposes hypotheses H3 to H4 to explore the effect of anonymity on the relationship between negative sentiment, conflict frames, and online incivility.

Figure 1 summarizes the conceptual framework of this study.

Figure 1:

Theoretical framework.

3.1 Platform selection

This research intentionally adopts X and Weibo as the research target, as these two popular microblogging platforms, originating from leading global powers, offers valuable perspective on how the international community views and interprets the conflict.

X (formerly Twitter) is a U.S.-based microblogging platform with 229 million monthly active users globally, including 95.4 million users in the United States as of 2022. The platform plays a significant role in facilitating political discourse, serving as a public sphere where citizens engage in horizontal discussions with one another and vertical interactions with policymakers (Jaidka et al. 2019). Given that political content, such as election campaigns, is frequently disseminated and discussed across multiple major social media platforms, X/Twitter provides a representative case for studying political discourse in the U.S. social media landscape.

Sina Weibo (Weibo) is a Chinese microblogging social media platform that owns 586 million monthly active users in 2022 Q4. It features platform structure similar to the X/Twitter, allowing users to publicly post and share content including short texts, image, and videos. Weibo is also one of the first three social media platforms officially utilized by both Chinese central and regional government as the official political information outlets, making it one of the most popular and influential platforms in China. The integration of official Chinese government and media on Weibo, combined with its original user-driven discussion basis, has created a multi-layer public sphere where diverse voice coexist (Rauchfleisch and Schäfer 2015) despite the platform still follows national censorship policy. The multi-layer public sphere pattern combining its open-content design, where content is visible to all users, helps Weibo to be positioned as a microcosm of the broader Chinese public sphere. Consequently, Weibo serves as a representative case for studying political discourse on Chinese social media.

The comparability of X (Twitter) and Sina Weibo comes from their similar platform structure, user behaviors, and their similar function in amplifying public discourse and social movements. Twitter and Weibo are both private-owned microblogging platforms, allowing users to share posts with short texts, images, videos, or links. The character limits on the two platform (280 characters for X, 140 Chinese characters for Weibo originally) both encourage users to use concise language in the communication to emphasize their opinion. In terms of user behaviors, both Twitter and Weibo use hashtags and keywords/key phrases rather than the posts by friends in the more private social network sites. This approach of content organization and trends highlighting allows for the rapid aggregation of public discussion around specific topic. Furthermore, Twitter and Weibo have proven to be powerful tools for amplifying public opinion and playing central role in organizing and promoting multiple social movements (Huang and Sun 2016; Huszár et al. 2022).

Based on the formulated research questions and hypothesis, this study utilizes data obtained from both original and shared posts on X (Twitter) and SINA Weibo pertaining to the Russia–Ukraine conflict. The X (Twitter) data was collected utilizing the X (Twitter) Academic API, while the Weibo data was gathered using a Python-based web scraper. The data collection process involved capturing posts from both platforms that featured one or more keywords or hashtags related to the Russia–Ukraine conflict. Furthermore, the selected posts were limited to those published between February 27, 2022, and September 27, 2022.

To retrieve data from Twitter Academic API, the research used a query-based Python script to search in X (Twitter) historic database for all English tweets contains one or more of keywords words from the keyword list. 44,179 tweets were collected within the time range. The keyword list is built on the top 10 most popular Russian–Ukraine war hashtags and keywords identified in previous research datasets (Chen and Ferrara 2023). The specific keywords include Ukraine, Russia, Putin, Standwithukraine, Kyiv, etc. (Table 1). All the tweets were collected exclusively from the users with geolocation setting in the U.S.

The research used a Python-based web scraper to search for and download 41,806 Weibo post. Due to the missing of API permission, the searching function of the scraper was built on the integrated searching function of Weibo website instead of accessing to the Weibo historic dataset. The search keywords were compiled based on the top 10 most popular Russian–Ukraine war keywords from prior research dataset (Fung and Ji 2022). The specific keywords include 乌克兰 (Ukraine), 俄罗斯 (Russia), 俄乌 (Russia Ukraine), 基辅 (Kyiv), 普京 (Putin), etc. (Table 1). The geolocation limit was not applied in Weibo data collection.

All data collected from X (Twitter) and Weibo contains the content of the posts, the time stamps of initial posting behavior, the user engagement data including count of likes, comments, and retweets, and the account name and follower numbers of the post composer.

3.2 Measurements

3.3 Data analysis

Upon completion of data collection, measurements for each variable were computed via computer-assisted content analysis. To address RQ1, descriptive statistics was used to summarize and compare the usage of incivility on X (Twitter) and Weibo to examine whether the degree of online incivility differs between social media users in China and in the United States when discussing the Russia–Ukraine War. Hypotheses 1 through 4 were tested using Hayes’ PROCESS macro model 1 to investigate the moderation relationships between variables.

4.1 Descriptive statistics

After preprocessing the data, 44,179 tweets and 41,806 Weibo posts were used in the analysis. In response to RQ1, the descriptive analysis result of the incivility score reveals a differentiation in the level of online incivility across two platforms in the discussion of Russia–Ukraine war. There were mild fluctuations observed over the eight months. The average incivility score for X (Twitter) over the 8-month period was 0.115 (M _{TwitterIncivility} = 0.115, SD_{Twitterincivility} = 0.101), with the highest score of 0.120 in July and the lowest score of 0.111 in September. Weibo’s cumulative incivility score averaged at 0.089 (M _{Weiboincivility} = 0.089, SD_{Weiboincivility} = 0.052), with February marking the highest monthly score at 0.095 and March the lowest at 0.085 (Figure 2). A two-sample independent t-test was conducted to compare the means of X and Weibo for incivility score. The result (t = 44.843, p < 0.001) indicates that the difference between two groups is highly statistically significant.

Figure 2:

Average incivility score of X (Twitter) and Weibo post from Feb to Sep 2022.

For the utterance incivility, 4,880 X (Twitter) posts (11.1 % of 44,180) included words from the designated incivility lexicon. X (Twitter)’s utterance incivility score (M _{Twitteruttrance} = 0.006, SD_{Twitterutterance} = 0.023) peaked in July at 0.00694 and dipped to 0.00540 in September. On Weibo, 1,313 posts (3.1 % of 41,798) featured incivility lexicon terms. Weibo’s average utterance incivility score was at 0.0002 (M _{Weiboutterance} = 0.0002, SD_{Weiboutterance} = 0.002) over the same period, hitting a maximum of 0.00028 in February and a minimum of 0.00017 in June (Figure 3). However, it is also noticeable that the average length of the Weibo posts (M _Weibolength = 223.38, SD_Weibolength = 235.08) is significantly longer than the average length of X (Twitter) post (M _{Twitterlength} = 25.13, SD_{Twitterlength} = 13.33), which potentially contributes to the large discrepancy in the utterance scores of the two datasets. A two-sample independent t-test was conducted to compare the means of X and Weibo for utterance incivility score. The result (t = 48.582, p < 0.001) indicates that the difference between two groups is highly statistically significant.

Figure 3:

Average utterance incivility score of X (Twitter) and Weibo post from Feb. to Sep. 2022.

For the discursive incivility, 4,183 tweets (9.5 %) are found to have a score above 0.5, indicating a toxic content, with an average of 0.224 (M _{Twitterdiscursive} = 0.224, SD_{Twitterdiscursive} = 0.190), across eight months – highest in July (0.234) and lowest in September (0.217). Compared to X (Twitter), Weibo have fewer posts with toxicity score over 0.5, at 348 (0.3 %), with an overall toxicity average of 0.177 (M _{Weibodiscursive} = 0.177, SD_{Weibodiscursive} = 0.103), reaching its apex in February (0.189) and its bottom in March (0.169) (Figure 4). A two-sample independent t-test was conducted to compare the means of X and Weibo for discursive incivility score. The result (t = 41.794, p < 0.001) indicates that the difference between two groups is highly statistically significant.

Figure 4:

Average toxicity score of X (Twitter) and Weibo post from Feb to Sep 2022.

The research then counted the top 20 keywords of the posts with an online incivility score of over 0.3 for both X (Twitter) (N = 2,657) and Weibo (N = 97) datasets. The keyword list only contains nouns and verbs to focus on the most impactful terms that contribute to the sentiment and context of the discussions.

The word frequency analysis (Table 4) reveals similar thematic focuses and emotional intensities between the two platforms. On X (Twitter), keywords such as “Putin” (1,071 mentions), “Russia” (609 mentions), “Ukraine” (540 mentions), and “war” (273 mentions) are prevalent, highlighting the central figures and geopolitical entities involved in the conflict. On Weibo, terms like “乌克兰” (Ukraine, 260 mentions), “俄罗斯/毛子” (Russia, 263 mentions), and “俄乌战争/战争” (Russia–Ukraine War/war, 162 mentions) dominate the conversations, indicating a similar focus on the geopolitical conflict.

For the uncivil words, the frequent use of profanity on X such as “f*ck” (410 mentions) and “sh*t” (237 mentions) suggests a high level of emotional expression and anger among users. The frequent occurrence of terms on Weibo like “狗/走狗” (dog/lackey, 36 mentions) also indicates the presence of derogatory language and strong disdain on the Chinese platform.

4.2 Relationship of negative sentiment, conflict frame, and online incivility

The hypotheses H1–H4 postulated that the negative sentiment and conflict frame present in social media posts would influence the level of online incivility. These effects were projected to be moderated by anonymity and mediated by online incivility. We used the PROCESS macro model 1 to test the moderate model and evaluate the direct and indirect impacts of two independent variables, Negative Sentiment and Conflict Frame, on the online incivility.

H1 proposed that the conflict frames in a post act as a positive predictor for the online incivility. This hypothesis was corroborated in both X (Twitter) (b = 0.056, SE = 0.001, t = 60.557, p < 0.001, 95 % CI [0.054, 0.058]) and Weibo (b = 0.012, SE = 0.001, t = 19.184, p < 0.001, 95 % CI [0.011, 0.013]) datasets. The empirical findings affirm a significant positive direct association between the presence of conflict frames and the extent of incivility.

H2 posited that the negative sentiment inherent in social media posts serves as a positive predictor of online incivility. This hypothesis was substantiated in both X (Twitter) (b = 0.169, SE = 0.001, t = 153.876, p < 0.001, 95 % Confidence Interval [0.166, 0.170]) and Weibo (b = 0.031, SE = 0.001, t = 36.290, p < 0.001, 95 % Confidence Interval [0.029, 0.033]) datasets. The results reveal a statistically significant positive direct correlation between negative sentiment and the extent of incivility.

H3 proposed that anonymity moderates the positive effect of conflict frame on online incivility such that the effect is positive and stronger when the username is anonymous. This hypothesis is rejected in both X (Twitter) and Weibo dataset.

For X (Twitter) data, the H3 was rejected. The interaction between conflict frame and anonymity significantly influenced online incivility (b = 0.005, p = 0.005), which implies that the effect of conflict frame on incivility was more pronounced in non-anonymous contexts. For Weibo data, the H3 was also rejected. The interaction between negative sentiment and anonymity in the regression model was not statistically significant (b = 0.003, p = 0.112).

H4 proposed that anonymity moderates the positive effect of negative sentiment on online incivility such that the effect is positive and stronger when the username is anonymous. This hypothesis is also rejected in both X (Twitter) and Weibo dataset.

For X (Twitter) data, the H4 was rejected. The interaction between negative sentiment and anonymity significantly influenced online incivility (b = 0.010, p < 0.001). Specifically, the effect of negative sentiment on incivility was more pronounced in non-anonymous contexts. For Weibo data, the H4 was also rejected. The interaction between negative sentiment and anonymity in the regression model was not statistically significant (b = 0.001, p = 0.746). This implies that the slope of the relationship between negative sentiment and online incivility does not significantly differ based on the level of anonymity (Figures 5 and 6).

Figure 5:

Hypothesis testing results of Twitter dataset.

Figure 6:

Hypothesis testing results of Weibo dataset.