Impact of ByteDance crisis communication strategies on different social media users

The study investigates how crisis communication strategies perform across different social media platforms in the context of ByteDance’s response to the Trump administration’s 2020 sanctions against TikTok. Guided by the Social-Mediated Crisis Communication (SMCC) model, the authors note that while SMCC explains effects of information source and form on publics’ responses, it does not explicitly account for platform differences. In China, ByteDance issued crisis communications via its own platform Toutiao, yet discussions were more intense on Weibo where third-party accounts reposted or reframed those announcements. This context created both an existential threat for TikTok in the U.S. and a domestic trust crisis for ByteDance. The study’s purpose is to compare the impact of ByteDance’s crisis statements on user sentiments and discourse on Toutiao versus Weibo, combining sentiment classification (machine and deep learning) with semantic network analysis (SNA). The importance lies in extending crisis communication theory to incorporate platform heterogeneity and in offering practical guidance on platform choice and messaging for corporate crisis management in the Chinese social media environment.

The review covers: (1) SMCC—differentiating social, traditional, and offline word-of-mouth media; outlining five factors shaping social media crisis management (crisis origin, type, infrastructure, message strategy, message form) and user roles (influential creators, followers, inactives). Prior SMCC research shows source credibility (organizational vs third-party) affects public sentiment but overlooks platform-specific opinion dynamics. (2) Crisis origin and type—victim, accidental, intentional clusters with varying reputational threat (Table 1). (3) Message strategies—deny, diminish, rebuild, with secondary tactics such as attack the accuser, justification, and apology (Table 2). (4) Homogeneity vs heterogeneity of platforms—evidence suggests user demographics and platform functions (e.g., Facebook’s two-way vs Twitter’s one-way model) shape perceptions and responses; comparable cross-platform research is limited in Chinese contexts beyond Weibo. With Toutiao’s growth, competition with Weibo warrants comparative analysis. (5) Prior China-focused corporate crisis communication on Weibo centers on SCCT, attribution, image restoration, and contextual framing, but no prior work focuses on Toutiao. (6) Sentiment classification—contrasts lexicon-based vs machine learning approaches; lexicon methods struggle with target specificity, while ML can classify stance relative to the focal organization (positive/negative/neutral). (7) SNA—preferred over LDA for short social media texts due to LDA’s challenges with topic interpretability and short-text sparsity; SNA enables visualization and quantitative-qualitative interpretation of word co-occurrence networks. The review motivates combining SMCC with ML-based sentiment analysis and SNA to examine platform-specific differences in the ByteDance case.

The study follows four components: SMCC analysis, data collection, data processing with sentiment classification, and SNA. SMCC analysis: The authors qualitatively coded crisis origin, crisis type, infrastructure, message strategy, message form, and identified influential social media creators. Data collection: ByteDance’s official announcements were posted on Toutiao; Weibo discussions arose via third-party reposts. Using Octoparse, the authors crawled five announcements that generated controversy on both platforms and their comment sections. The dataset comprised 50,702 comments (Toutiao 23,541; Weibo 27,161). For Weibo, comments were taken from the top third-party account per announcement (e.g., Sina Tech, Guancha syndicate, People’s Daily, Global Times). A stratified random sample of 8,793 comments was drawn for manual labeling (Toutiao 3,992; Weibo 4,801). Manual annotation: Five PhD annotators followed a detailed guide. Three annotators independently labeled each comment’s polarity toward ByteDance (positive, neutral, negative). Disagreements were adjudicated by two additional annotators, achieving high reliability (Holsti coefficient ~0.9). Only non-controversial comments were retained. The labeled set contained 2,403 positive, 3,200 neutral, and 3,190 negative comments. Model training and evaluation: Machine learning classifiers (logistic regression, KNN, random forest, SVM, Naive Bayes) and deep learning models (LSTM, BERT, ERNIE) were trained. Hyperparameters for ML were optimized via grid search; deep learning settings included common learning rates (e.g., 2e-5 for BERT/ERNIE), sequence length 256, batch size 16, and iterations/epochs as specified. Performance was evaluated using F1-score. The best-performing model (ERNIE) was applied to the full dataset for polarity classification. SNA: From comments segmented by polarity and platform, TF-IDF extracted the top 50 words per sentiment. Word co-occurrence matrices were constructed and clustered/visualized in Gephi (Louvain algorithm). Domain experts interpreted clusters into topics. The SMCC component also coded ByteDance’s five announcements (August 2, 3, 7, 23; September 20, 2020) for response strategy (deny/diminish), and the study identified influential Weibo creators by retweet volume across categories (organization, media, journalist, netizen).

• SMCC codings: Crisis origin was external (U.S. sanctions), placing ByteDance in the victim cluster. Infrastructure was centralized via the official ByteDance Toutiao account. Message strategies across the five focal announcements included deny (three times, attack the accuser) and diminish (two times, justification). Message form relied on social media, with announcements disseminated promptly (within a day of related rumors or developments). - Influential Weibo creators (by retweets): organization: The Rule of Law in Sichuan (2,458); media: Global Times (18,013); journalist: Cattle to play the piano (31,714); netizen: Carrier of a secret (1,695). - Sentiment classification performance: Among ML classifiers, logistic regression achieved the highest F1-score (62.2%). Among deep models, ERNIE was best (F1 = 82.4%), outperforming LSTM and BERT; ERNIE was used for full-dataset inference. - Cross-platform sentiment dynamics (notably the 4th announcement, 2020-08-23): Toutiao’s positive sentiment increased markedly from 49.16% to 86.15%. Weibo’s positive sentiment dropped from 72.31% to 24.34%, with neutral rising to 33.00% and negative to 42.66%. - SNA topic patterns: • Positive comments—Toutiao exhibited five communities supporting ByteDance’s global direction, leadership (Zhang Yiming), and resilience; Weibo showed no positive-topic communities in the analyzed slice. • Neutral comments—Toutiao had two topics: (1) ByteDance remains based in Beijing; TikTok HQ relocation out of the U.S.; (2) Withdrawing from the U.S. could protect other markets. Weibo had six topics emphasizing patriotism as a baseline for corporate conduct, minimizing losses via partial divestment, pro-American vs patriotic stances, dependence on national support for globalization, debates among commenters, and the idea that U.S. administrative actions contradict free trade. • Negative comments—Toutiao had three topics: preferring exit over sale to Microsoft; skepticism about moving HQ to London; concern over potential Chinese government takeover if the U.S. operation was sold. Weibo had seven topics: fears of cascading demands by U.S. allies if selling now; Zhang Yiming being pro-American; calls (often sarcastic) to sell algorithms/data; labeling ByteDance as cowardly; profit-first management critique; personal discontent with Zhang; and claims that U.S. handling resembles unequal treaties. - Overall, platform differences were stark: Toutiao (ByteDance-owned) users were more supportive of the company’s defensive responses; Weibo (third-party reposts) users were more critical and skeptical.

Findings extend the SMCC model by incorporating platform affiliation and user-base differences as determinants of public sentiment. When the crisis is externally attributed and communications are issued on a platform belonging to the same corporate group as the focal organization, public acceptance of defensive strategies (deny/diminish) is higher. Conversely, when messages are mediated by third parties on platforms outside the corporate group, acceptance decreases and skepticism increases. Two mechanisms are proposed: (1) user attribute heterogeneity across platforms (demographics, norms, and expectations), and (2) differences in information sourcing (direct official messaging vs third-party reposts/reframings) influence credibility and acceptance. The study also observes users’ anthropomorphization of corporate traits—interpreting tough vs passive messaging as tough vs cowardly corporate personalities—which shapes affective responses. Practically, firms should adopt multi-platform strategies, avoid over-reliance on owned platforms, and tailor messages to platform-specific audiences and trust dynamics to mitigate negative spillovers and polarization.

This work pioneers a comparative crisis communication analysis between Toutiao and Weibo in the Chinese context and the first focused examination of public opinion on Toutiao. Using SMCC-guided coding, machine/deep learning-based sentiment classification (with ERNIE as the optimal model), and SNA, the study shows that platform differences strongly condition public reception of the same crisis messages. Key contributions: (1) integrating platform heterogeneity and platform–organization affiliation into SMCC to explain acceptance of defensive strategies when the crisis is externally attributed; (2) identifying anthropomorphization of corporate traits by social media users; and (3) offering actionable guidance to engage across multiple platforms and tailor strategies to user attributes and information sources. Future research should deepen analyses of platform–user attribute relationships, leverage richer weighting schemes (e.g., likes/retweets) in text networks, and test generalizability beyond the ByteDance/TikTok case.

The study focuses on a single case in a specific national context, which may limit generalizability. While TF-IDF is efficient and realistic for feature extraction, it ignores word position and engagement signals; comments with higher likes/retweets that may better represent salient public views were not differentially weighted. The work does not include experimental or survey data to directly measure user attributes or causal mechanisms behind platform differences; further exploration of platform modes and their relationship to user attributes is needed.