The retirement placement policy serves as the fundamental institutional guarantee for athletes to achieve career transitions. China sees approximately 4,000 athletes retire annually, with numbers increasing during Olympic and National Games years. Their primary employment channels include organizational placements and self-initiated career choices. As China's economic reforms deepen and social transformations progress, the conflict between retired athletes' employment placement and labor market mechanisms has become increasingly prominent (Zhu & Zou, 2021). Current research predominantly examines retired athletes' status, challenges, and re-employment factors through sports science perspectives, exploring implementation barriers of placement policies, social support systems, and income impacts. Most scholars focus on livelihood capital development, entrepreneurial realities, business environments, and support frameworks for retired athletes. Notably, research from non-sports disciplines remains scarce (Wang & Liu, 2018). Furthermore, studies rarely address public discourse and attitudes toward retired athletes' re-employment processes, with social media discussions receiving minimal attention. Existing research predominantly adopts policy and athlete-centric approaches, resulting in relatively narrow coverage of this critical field.

On November 22, 2024, Guan Chenchen, a retired national gymnast and Tokyo Olympic balance beam champion, left a comment on a TikTok video posted by former gymnast and World Cup champion Wu Liufang. She wrote, "Sis, if you want to polish your shoes, go ahead. Don't keep criticizing gymnastics [crying]. Your positive energy doesn't seem to be needed." Wu Liufang responded: "Can't eat grapes, so you complain?!" Guan Chenchen replied: "Sis, you're amazing! I'm so envious of your achievements and medals. You're truly amazing! I should learn from you!!!" Wu Liufang then responded: "You're actually more impressive [thumbs up]. When you lose weight, you'll be able to wear pretty clothes too. Keep it up! [heart gesture]." The heated exchange, captured by social media screenshots, sparked widespread attention and discussion. Wu Liufang first joined China's national gymnastics team in 2008 but retired due to an injury during the women's balance beam final at the 2012 National Gymnastics Championships and Olympic Trials. She later studied at Beijing Sport University and became a gymnastics coach at Hangzhou Zhongxiang Sports Company after graduation. During her live stream, she revealed her monthly salary was only 3,500 yuan and that her boss often delayed payments. In 2021, she taught at a school in Jiaxing under the promise of a permanent position, but it was taken by someone else. In March 2024, Wu Liufang returned to Hangzhou and launched her TikTok account, becoming a popular female streamer who posted multiple dance videos online. The conflict between Guan Chenchen's cool dance in the video and her verified identity as a retired athlete on her account page sparked her dissatisfaction, ultimately leading to public discussions about this incident on TikTok platforms. The conversation evolved from debates about Wu Liufang's career choice as an "aesthetic anchor" to broader reflections on the employment challenges faced by other retired athletes, and further to discussions about China's retirement policy and institutional safeguards for athletes. This public discourse demonstrated diverse perspectives and reflective characteristics.

Based on this, this study aims to analyze public discourse regarding retired athletes' career transitions in the Wu Liufang case on TikTok platforms. We manually selected 23 official media videos related to the Wu Liufang incident from November 22 to December 7, 2024 TikTok platforms, and collected comments beneath these videos using web crawling technology. Subsequently, we applied computational communication research methods to conduct basic characteristic analysis, sentiment analysis, and thematic analysis of public discourse in this event. The core research question is: "Public opinion on retired athletes transitioning to online streaming careers through social media." This question is explored through the following sub-questions:

·[Rq1] What are the defining characteristics of public discourse surrounding the Wu Liufang incident on TikTok?

·[Rq2] What is the predominant sentiment of public discourse regarding the Wu Liufang incident on TikTok?

·[Rq3] What stances and appeals were expressed in the public discourse surrounding the Wu Liufang incident on TikTok?

·[Rq4] How did public opinion evolve over time throughout the Wu Liufang incident on TikTok?

·[Rq5] What were the respective emotional tendencies and key discussion focuses of the public during each phase of the Wu Liufang incident on TikTok?

To address the aforementioned research questions, this study adopts a step-by-step approach. First, it analyzes the basic characteristics of the research samples in terms of time, space, quantity, and emojis. Then, sentiment analysis is conducted using the built-in dictionary in Python's SnowNLP library and three self-constructed emotion dictionaries. Next, LDA topic modeling is applied for thematic analysis. Finally, the evolution cycle of public opinion is analyzed by combining the temporal trends of sentiment analysis and topic modeling. The findings encompass four core outcomes: basic characteristics, sentiment analysis, thematic analysis, and public opinion evolution cycles.

Research on Career Transition of Retired Athletes. Currently, China's resettlement policies for retired athletes face implementation challenges including low policy effectiveness, complex execution procedures, and operational difficulties (Wang et al., 2024). The career transition strategies for retired athletes primarily fall into two categories: government-sponsored placements and independent employment outside the system. Government-sponsored placements refer to mandatory assignments by government departments to athletes with outstanding achievements. According to policy documents issued by relevant authorities, retired athletes seeking independent employment outside the system receive a one-time economic compensation (Li, 2015). In recent years, with the deepening integration of sports and education, exploring career transitions of retired athletes into roles such as physical education teachers or school coaches has become a new focus for government departments and society. Scholars have begun investigating the current status, influencing factors, challenges, and countermeasures regarding retired athletes transitioning into sports teaching positions (Yang & Li, 2023).

The career transition and quality of retired athletes are influenced by multiple factors. From macro, meso, and micro perspectives, these transitions are affected by national policies at the macro level, constrained by family economic status and social support at the meso level, and influenced by individual characteristics and athletic careers at the micro level (Zhang et al., 2022). Emotional comfort from close relationships, experience sharing, and employment information, training resources, and entrepreneurship guidance provided by professional institutions can comprehensively help athletes cope with psychological setbacks, career confusion, and skill gaps during the transition process, effectively reducing the risk of transition failure (Lavallee & Wylleman, 2000). Meanwhile, career transition competence is also a key factor enabling retired athletes to smoothly achieve career transitions. This refers to an individual's ability to take action in response to career changes and better adapt to new professions (Chang & Liu, 2024). Educational attainment shows a significant positive correlation with career transition competence among retired athletes, as education equips them with general knowledge and thinking skills beyond specialized skills, enabling them to more smoothly cross industry barriers and integrate into diverse workplace environments after retirement (Brown & Levinson, 1983). The marketization level of sports significantly impacts the employment quality of retired athletes. Athletes from high-market sports like ball games and swimming demonstrate notably higher post-retirement employment quality compared to those from heavy competitive sports like weightlifting and wrestling (Yin & Zou, 2023).

Furthermore, to better address the re-employment challenges faced by retired athletes, academic research has been conducted on specific topics with notable achievements, including studies on resettlement policies, psychological research, social security systems, employment planning and training programs, capital investment for re-employment, and continuing education initiatives (Wang & Liu, 2018). The innovation of transition support mechanisms for retired athletes in the new era faces challenges, with institutional innovations in athlete development systems and coordinated advancement of vocational and cultural education emerging as pathways for their transition under the new institutional framework. Social support remains a crucial factor influencing career capital among Chinese retired athletes, where optimized social support can enhance transition quality during their professional transitions (Gao & Liu, 2022). In this process, career planning and educational interventions serve as core strategies. Many national sports academies and professional clubs have incorporated career planning courses into athletes' regular training systems, utilizing diversified approaches like aptitude tests, career counseling services, and internship programs to help athletes anchor post-retirement career directions and accumulate professional experience in advance (Lavallee, 2005). Mental health interventions also hold significant importance, as retired athletes often face psychological imbalance during early transition periods. Cognitive behavioral therapy and psychological counseling courses have emerged to guide athletes in reshaping cognitive patterns, regulating emotional states, and strengthening psychological resilience, thereby injecting mental energy into their career transitions (Gustafsson et al., 2017).

In methodological research, qualitative studies have secured their place in the field of retired athletes' career transitions through their capacity for deep insights. The application of interview methods and case analysis allows researchers to delve into the inner world of retired athletes, meticulously depicting their psychological journeys, challenges encountered, and coping strategies during career transitions. For instance, interviews provide empirical evidence to reveal athletes' subjective experiences and psychological transformation trajectories during retirement, enriching theoretical frameworks (Coakley, 1983). Quantitative research employs standardized data collection methods like questionnaires and scales to conduct quantitative analyses of variables involved in career transitions, precisely measuring correlations between indicators such as occupational adaptability, mental health levels, and social support intensity. This approach helps validate theoretical hypotheses and identify universal patterns (Roberts et al., 2023). Mixed-method research has also emerged as a trend, where scholars skillfully combine qualitative and quantitative approaches. This method respects individual experiences while leveraging data mining to identify common patterns, yielding conclusions that balance depth and breadth to comprehensively map the career transition landscape of retired athletes.

Social Media Sentiment Analysis and Topic Modeling Methods. Social media platforms provide citizens with spaces to express opinions, share information, and organize actions, enabling broader public participation in discussions and decision-making regarding public issues. However, phenomena such as polarized discourse and emotional expression have become increasingly prevalent on these platforms (Bennett et al., 2018; Blumler et al., 1995). In this incident, public comments on official accounts on TikTok platforms sparked extensive discussions, containing numerous viewpoints and emotions. Continuing to use traditional topic analysis methods would be redundant for such emotional and biased information, leading to the emergence of sentiment analysis research that has gained popularity (Liu & Liu, 2018). Previous studies employed various natural language processing models, including neural networks, random forests, support vector machines, and BERT, to investigate topics related to emotions (e.g., depression, anxiety, anger) on social media platforms like X, Reddit, YouTube. For instance, scholars such as Verbeij have conducted research using the neural topic modeling approach, which provides new insights into how adolescents express emotions of happiness and sadness on Instagram (Verbeij et al., 2024).Current research increasingly focuses on methodological challenges in sentiment analysis, indicating that the field of emotional discourse characteristics in social media networks is evolving and maturing (Peng et al., 2024). Sentiment analysis, also known as opinion mining, refers to the use of computers to analyze emotionally charged subjective content, identify underlying emotional tendencies, and categorize attitudes (Wang & Yang, 2021). Through the sentiment analysis of text content, we can determine the positive and negative attributes of content and identify the expressions of emotion and evaluation (Wilson et al., 2005).For example, in their study, Dave et al. (2003) utilized information retrieval technology to perform feature extraction and statistical scoring on large-scale product review texts obtained from platforms such as CNet and Amazon, thereby effectively distinguishing between positive and negative reviews (Dave et al., 2003).

The vast data contained in social media has made topic modeling a crucial task in the field of text mining (Chen & Li, 2019). As a type of probabilistic model, topic modeling extracts abstract thematic information from large-scale texts or corpora. Essentially, it is a fast unsupervised machine learning algorithm that clusters similar word sets by observing the distribution patterns of words in texts or corpora (Zhang & Song, 2012). This process uncovers hidden themes within texts and reveals evolving correlations between topics. These thematic patterns reflect the focal points of public discourse, providing authorities with actionable insights for precise public opinion guidance.

Scholars frequently combine sentiment analysis with topic modeling in research, though some studies employ either method independently. For instance: Wang Yijun and Huang Shiyun (2024) employed an image sentiment analysis model to investigate the shifts in emotional expression by TikTok bloggers before and after government responses, analyzing the impact of these responses on public sentiment; Zhou Li et al. (2018) applied fine-grained sentiment analysis to study online emotional attribution and influencing factors in anti-corruption discussions; Cheng Siqi and Yu Guoming (2022) conducted a comparative analysis of Weibo users' emotional experiences with their domestic and international counterparts on YouTube through text sentiment analysis; Li Qingzhen and Tang Xi (2022) identified key themes contributing to social media fatigue using Latent Dirichlet Allocation (LDA) models; Cheng Zheng (2023) developed a hybrid BERT-topic modeling approach for detecting misinformation on social media; Shao Yuanhang (2024) analyzed hierarchical effects in online health information dissemination using BTM topic modeling. In studies that apply the two in combination, the research outcomes include those on the evolution of public opinion regarding sudden/hot events on Weibo (An Lu & Wu Lin, 2017; Zeng Li et al., 2022), the expression of mainstream online public opinion (Shen Hongzhou et al., 2023), the communication landscape and emotional distribution of global climate change issues on the international social platform Twitter (Wei Jianyi, 2024), and the representation of China-related topics by English-speaking users on Twitter (Yi Hongfa et al., 2014).

The theory of public opinion evolution cycle. In a broad sense, public opinion refers to the complex interplay of emotions, intentions, attitudes, and opinions held by the public-comprising individuals and various social groups-toward matters of concern or closely related to their interests within specific historical periods and social contexts (Gu & Zhou, 2009). Online public opinion represents an enhanced manifestation of this concept, describing the political attitudes held by citizens toward government officials regarding mediated social issues within cyberspace (Li & Zhang, 2010). During the "Wu Liufang incident", Wu Liufang's personal account was banned on the TikTok platform, with official penalties including account suspension, reinstatement, and account cleanup. The primary venue for public discourse in this case was also the TikTok cyberspace. Notably, public discussions repeatedly referenced state entities like the "Liuzhou Sports Bureau", "National Sports Bureau" and "leadership", confirming this as a genuine public opinion event.

The evolution of events typically follows a distinct life cycle pattern. Domestic and international scholars have established a well-developed framework for studying public opinion dissemination processes (An Lu & Wu, 2017). These studies categorize public sentiment into distinct phases based on event occurrence sequences and developmental cycles. Among the most renowned communication models are the three-stage model proposed by B.T. Burkholder et al. (1995), the four-stage model developed by Robert Heath based on the three-stage framework (Heath, 2004), and the crisis life cycle model introduced by S. Fink et al. from a medical perspective (Fink, 1986); Ma Jianhua and Chen An(2009) proposed four phases for emergency event evolution-latent, outbreak, spread, and recovery phases ; while Jia Yamin et al. categorized urban emergency events into four stages: initial phase, outbreak phase, decline phase, and resolution phase (Jia et al., 2015). This paper applies Jia's classification to the "Wu Liufang incident," identifying four evolutionary stages: initial, outbreak, decline, and resolution phases. Additionally, online public opinion events exhibit three primary evolution patterns: unimodal, bimodal, and multimodal. The unimodal pattern is characterized by a single peak, whereas bimodal and multimodal patterns feature two or multiple peaks respectively (Jiang, 2014).

Sentiment Analysis. Building upon the built-in dictionary of the Python SnowNLP library, this study developed a three-dimensional sentiment lexicon system of "positive support-neutral rationality-negativity opposition" to better accommodate the complexity of public opinion regarding this event. The positive support lexicon comprises 288 terms, integrating online subcultural symbols like "[666]", "[like]", "[rose]", "[chong chong chong]" and legal defense phrases such as "non-illegal", "compliant and reasonable", along with sports honor terms like "national glory", "honor" and "glory". This reflects public recognition of the "legitimacy of survival rights" regarding retired athletes're-employment. The neutral lexicon contains 61 terms focusing on institutional words like "policy" and "guarantee", as well as role references such as "athletes", "common people" and "ordinary individuals", capturing rational discussions like "improving mechanisms" and "relevant organizations". The negativity opposition lexicon includes 41 terms emphasizing moral evaluations like "vulgarity" and "baseline", along with emotional expressions such as "disgusting" and "chilling", reflecting normative expectations for "athlete image".

Furthermore, during the data analysis process, a negation processing mechanism was established by incorporating negative terms such as "no", "not", "none", "do not", and "which" to achieve polarity reversal. For instance, the combination of "no" with "illegal" transforms the negative or neutral connotation of "violating" into a positive one, while pairing "which" with "wrong" converts the negative or neutral meaning of "wrong" into positivity, thereby enhancing the ability to identify ironic expressions. Additionally, this study implemented quality control through conflict detection algorithms to eliminate cross-category duplicate words and prioritized phrase-level matching (e.g., assigning higher weight to "what's wrong with it" compared to the single character "wrong").

LDA Topic Modeling. This study employed the discontinued lexicon from Harbin Institute of Technology for comment data segmentation. Given the unique characteristics of public discourse in sports-related TikTok platforms, we manually curated the top 1000 most frequent words with their parts of speech and occurrence frequencies using Python code. Subsequently, we constructed specialized dictionaries for both "sports domain-specific vocabulary" and "TikTok domain-specific vocabulary" through manual curation. The LDA topic modeling was then conducted using Python programming.

Basic Feature Analysis. To address Research Questions 1 and 4, this study conducted a comprehensive analysis of public discourse in the Wu Liufang incident on TikTok platform, examining quantitative characteristics, temporal distribution patterns, spatial distribution features, and emoji usage. The analysis revealed that 25,829 users posted 28,697 comments from November 22 to December 7, 2024, with an average of 1,614 users leaving approximately 1,794 comments daily over 16 days. Notably, five of these days saw over 1,440 comments each (equivalent to at least one comment every minute).

Figure 1 reveals two distinct peaks in the public opinion evolution of the Wu Liufang incident, demonstrating a bimodal pattern. The first peak emerged on November 24,2024, with 7,494 comments (26.1% of total). The second peak occurred on December 1, 2024, featuring 7,251 comments (25.3%). The Tiktok comment data pertaining to the Wu Liufang incident was periodized according to the public opinion evolution cycle framework proposed by Jia Yamin et al. The six stages are defined as follows: the initial bimodal stage from November 22 to 23, the outbreak stage from November 23 to 24, the first decline stage from November 24 to 28, the second growth stage from November 28 to December 1, the second decline stage from December 1 to 3, and the resolution stage beginning on December 3 (Jia et al., 2015; An & Wu, 2017). The first peak resulted from Wu Liufang's TikTok account being banned for violating community guidelines that evening, while the second peak followed the removal of the ban. This analysis conclusively addresses Research Question 4: The incident exhibited a bimodal evolution pattern.

Figures 2 reveal that user IP locations for comments span 34 provincial-level administrative regions in China, 58 foreign countries, and 55 unknown locations. Remarkably, even the Macao, China with the fewest comments (20) still generated 20 reviews within 16 days, while overseas comments accounted for 27.3% of total interactions, demonstrating cross-border dissemination patterns. Guangdong Province led with 5,183 comments, followed by Zhejiang, Jiangsu, Shandong, and Beijing (43.0% combined), collectively constituting the communication core. Concentration analysis shows Guangdong (18.1%), Zhejiang (7.3%), and Jiangsu (7.2%) provinces contributed 32.6% of total comments through a power-law distribution, validating the "core-periphery" communication model.

The National Bureau of Statistics' newly released "2024 China Statistical Yearbook" reveals that Guangdong, Zhejiang, Jiangsu, Shandong, and Beijing collectively account for 38.6% of China's GDP and 26.7% of its population. This massive economic scale and demographic base form the material foundation for the geographical distribution patterns observed in public opinion during this event. Notably, these five provinces boast an internet penetration rate of 72%, surpassing the national average of 58%, while their mobile user activity (DUA) levels are 1.3 times higher than the national average, highlighting the geographic clustering of "digital discourse power" in this context. The inclusion of Cantonese-language comments in some user discussions has further created a dialect cultural sphere, enabling local users to TikTok through the platform's hometown recommendation algorithm to access relevant videos and boost engagement within their regions. Additionally, as Wu Liufang is currently located in Hangzhou, Zhejiang Province, the algorithm's recommendation system directs related videos to TikTok users from Zhejiang and Jiangsu provinces, which has contributed to increased comment volumes in these regions.

There is an inherent connection between the fluctuation of word frequency and social phenomena as well as information phenomena. Certain social phenomena and information phenomena will trigger specific fluctuations in word frequency (Deng, 1988). In other words, there is also an inherent link between the word frequency in Douyin comments regarding the Wu Liufang incident and the attitudes and opinions of the public toward this incident; changes in the Wu Liufang incident also lead to fluctuations in the word frequency of Douyin comments. Moreover, as a scientific research method that enables us to see through appearances to perceive the essence, the word frequency analysis method features objectivity, accuracy, systematicness, practicality and other characteristics, and it also has great application value in media research (Wu, 2008). Therefore, this study briefly analyzes the public opinion of the Wu Liufang incident on the Douyin platform through word frequency statistics. Figure 3 shows that the top 10 most frequent words are: "support" (3,231 times, accounting for 10.72% of the 50 most frequent words), "life", "cover your face", "retire", "country", "athlete", "gymnastics", "no problem", "champion" and "respond" (597 times).

The wordcloud (Fig. 4) highlights professional legitimacy-related expressions such as "what's the problem?", "no legal violations", "no theft", "no robbery", and "rights". It also reveals terms associated with career transitions, including "internet celebrity", "streamer", "live streaming", "dancing", "gymnastics", and "profession". Notably, words reflecting survival pressures and professional dignity, such as "earning money", "making a living", "surviving", "having enough to eat", "struggle", "effort", and "self-reliance", stand out prominently. Personal identities like "Wu Liufang", "Olympic champion", and "Guan Chenchen" are clearly visible, while emotional support symbols such as "roses", "applause", "heart gestures", and "smiles" are also evident. Collectively, these elements demonstrate that public discourse shows a tendency to support retired athletes' career transitions, address their livelihood realities, and uphold their professional legitimacy.

Furthermore, statistical analysis indicates that 28.46% of the comment texts contain emojis. Through manual organization of these symbols in this study, two distinct formats were identified: the first is the Chinese character format within brackets (e.g., [Like], [Tear], [Crying), and [Nine-Turn Intestine], while the second consists of Unicode emoji characters.

The analysis reveals (see Fig. 5) that the TOP10 most frequently used emojis in the commentary text are: "[Like]" (5,128 occurrences, accounting for 33.96% of all emojis), "[Facepalm]" (2,321 occurrences, 15.37%), "[666]" (318 occurrences, 2.11%), "[Look]" (318 occurrences, 2.11%), "[Rose]" (318 occurrences, 2.11%), "[Sniff]" (318 occurrences, 2.11%), "[Clap]" (318 occurrences, 2.11%), "[Pinch Nose]" (318 occurrences, 2.11%), and "[Tear] (318 occurrences, 2.11%)." Among these, emojis like "[Like]," "[Rose]," and "[Clap]" vividly convey public support for Wu Liufang. Meanwhile, "[Facepalm]", "[Tear]", "[Sniff]", "[Clap]", "[Pinch Nose]" and "[666]" primarily express public sympathy or dissatisfaction with authorities' handling of Wu Liufang's case, carrying strong ironic undertones.

Sentiment Analysis.This study employed Python's built-in SnowNLP library and three self-developed sentiment lexicons for sentiment distribution analysis. To evaluate the accuracy of SnowNLP's data analysis, we conducted stratified sampling and compared its annotated results with manual annotations. First, 168 seed data points were manually labeled covering various emotional tendencies (positive, neutral, negative) and comment types (text, emojis, long texts, etc.). Next, Python code was used to generate Excel files containing approximately 2,000 samples from six phases through stratified sampling-each phase comprising 40% positive, 30% neutral, and 30% negative sentiment data-with corresponding content, emotion labels, and category tags. Subsequently, three graduate students specializing in journalism and communication manually annotated the aggregated samples from all six phases and performed reliability testing, achieving a K = 0.912 score indicating high consistency that served as the "gold standard" for evaluating SnowNLP's accuracy. Comparative analysis revealed SnowNLP's annotated results showed only about 40% accuracy, failing to meet the study's sentiment analysis requirements. Therefore, this study developed three types of sentiment lexicons through iterative refinement, resulting in 283 positive words, 56 neutral words, and 39 negative words. The developed lexicons were then integrated with the SnowNLP library for sentiment analysis, with manual annotations serving as the gold standard. The results demonstrated that the combination of SnowNLP and the lexicons achieved an accuracy rate of 92.08% in sentiment analysis, proving its suitability for this research.

The analysis reveals (see Fig. 6) that throughout the public opinion development process, the majority of the public maintained a positive supportive sentiment (74.5%), followed by neutral rational attitudes (24.0%), with negative opposition constituting only a small fraction (1.5%). Figure 7 shows that all three sentiment categories peaked on November 24, 2024; positive support reached another peak on December 1; while negative opposition remained relatively stable throughout the event's progression with consistently low comment volumes. This occurred because Wu Liufangs' TikTok account was banned from being followed on November 24 due to violations of community guidelines, which sparked widespread public discussion. People expressed diverse opinions across the platform, leading to rapid development of public sentiment. On November 27, an article by "Zhejiang Gonggong Public Welfare Development Center" highlighting Wu Liufang's charitable contributions revealed her compassionate side, leading to increased positive support. The subsequent removal of her TikTok account on December 1 aligned with public expectations, causing positive sentiment to peak. As the event stabilized, public discussions diminished and the emotional momentum gradually subsided. From the dynamic process of opinion evolution, positive sentiment peaked during the outbreak phase, reached another peak in the second growth stage, while the subsequent decline and calm phases saw minimal comment activity with no significant emotional manifestation. This shows that the evolution of public opinion on social media platforms is closely related to the development of events, and has a high degree of instability and great fluctuations. Therefore, relevant departments need to take timely measures to guide and control the development of public opinion.

Thematic modeling. This study employs Latent Dirichlet Allocation (LDA) for thematic modeling to analyze public concerns and identify latent opinions and demands. Perplexity serves as an internal evaluation metric in LDA modeling, measuring the model's predictive performance on new texts. Lower perplexity values indicate better prediction accuracy (Blei et al., 2003). The formula for calculating perplexity is:

Consistency (coherence) serves as another evaluation metric for LDA models, assessing the coherence and interpretability of inferred topics. Higher consistency scores indicate better thematic coherence and explainability. While multiple methods exist for calculating consistency, scholars Both A and Hinneburg A have demonstrated that c_v (consistency score) aligns most closely with human interpretability. This study therefore employs c_v consistency to evaluate topic coherence (Both et al., 2015). A higher c_v score is preferable, with values above 0.4 generally considered acceptable. As shown in Figs. 8 and 9, the model achieves minimal confusion and maximum consistency when processing 8 topics. Figure 10 presents a perspective map of LDA topic modeling, revealing 8 mutually exclusive topics with no overlap. This configuration ensures optimal predictive performance and interpretability when operating with 8 topics.

This study analyzes the topics and manually merges the topics according to the TOP10 keywords of each topic obtained by topic modeling (see Table 1) and the specific situation of the event.

The first section, titled "Legitimacy and Public Support for Career Transition", includes themes 0, 3, and 6. This theme demonstrates public endorsement of Wu Liufang's post-retirement career choices, with supporters citing legal justifications ("no theft, no robbery", "no illegal acts", "no issues"). Emotional endorsements were expressed through emojis like roses and applause, emphasizing the principle that "what's not prohibited by law is permissible". Critics rejected moral coercion, with comments like "supporting legality" and "no moral coercion allowed" -as seen in the comment: "If you're not breaking the law, why impose restrictions? Just because someone retired doesn't mean they can't live normal lives."

The second theme, "Survival Pressure and Lack of State Support", encompasses Topics 2, 4, and 7. This section highlights public concerns about retired athletes' economic struggles ("living", "earning", "relying on themselves"), criticizing systemic inadequacies. It also calls for state accountability ("the nation", "work units", "government arrangements"), contrasting the treatment disparity between Olympic champions and ordinary athletes. The discourse extends to broader social equity issues through poignant remarks like: "While words may be harsh, the truth remains the same-ordinary people share her perspective, all striving to make ends meet", and "Even Olympic champions must hustle for survival; you can imagine how tough life is for common folks".

The third theme, "Moral Controversy and Image Conflict", encompasses Topics 1, 5, and parts of Topic 7. It highlights the clash between athletes' public identities as "gymnasts" and "champions" versus their personal choices as "streamers" and "live broadcasters". The discussion also addresses normative expectations surrounding the "national athlete" label, with comments like: "Gymnasts should maintain their image-how can they dress so sensually?" and "This damages national sports reputation; severe punishment must be meted out without exception". Some netizens criticized Wu Liufang's behavior as "undermining sportsmanship", while others engaged in moral condemnation through sarcastic remarks such as "facepalm", "What's wrong with that?" and "No offense meant."

The fourth aspect, "the playful expression in online communication", encompasses themes 5 and 6. Netizens deconstruct the serious controversy linking retired athlete Wu Liufang with national image through memes like "smile" and "Jiuzhuan Dachong". Ironically, comments such as "No one helps me reach my lofty aspirations; my junior sister sent me to the mountain peak" and "How much is a pound of feathers?" amplify public sentiment. These expressions reflect TikTok' s entertainment-oriented trend in public discourse, which increasingly dissolves traditional moral debates.

The "legitimate support" in Category One and the "image controversy" in Category Three stand in direct opposition, reflecting a clear divergence in public values. The themes of "survival pressure" and "absence of institutional guarantees" identified in Category Two form the foundational basis for both supportive and neutral public opinion, while also serving as a starting point for systemic criticism. Lastly, the humorous expressions prevalent on TikTok, as captured in Category Four, lower the threshold for public participation and accelerate the spread of sentiment. Although this demonstrates the platform's capacity to amplify public discourse, it may also risk obscuring core substantive issues.

Analysis of Public Opinion Evolution Cycle. This study examines themes and corresponding emotions at each developmental stage to reveal thematic correlations and emotional diffusion patterns within the same phase. Figures 11 and 12 show that during the initial stage, themes 0 and 2 showed significant prominence, with similar emotional distributions where positive and neutral sentiments dominated. Relevant authorities could therefore implement comparable emotional guidance strategies, actively respond to netizens' concerns about retired athletes' career transitions, and propose improved national safeguard measures. In the outbreak stage, positive support surged to its peak level, while neutral rationality also reached its highest point, although it remained significantly less intense than positive sentiments. Negative opposition emerged but remained relatively weak. Theme 2 became most prominent during this phase, as netizens extended their concern over Wu Liufang's post-retirement survival pressures to highlight systemic inadequacies in national resettlement support for retired athletes, expressing strong support for Wu and reflections on institutional safeguards for retired athletes. The incident of Wu Liufang's TikTok account being banned due to violating community regulations pushed three types of emotions to their peak throughout the public opinion development process, making it one of the most critical events in the outbreak stage and a pivotal breakthrough for public opinion management. During the first decline phase, positive and neutral emotions exhibited multiple peaks. Early in this stage, multiple themes developed concurrently, reflecting diversity in the discourse field. In the mid-phase, themes 1 and 2 became prominent, with positive and neutral sentiments developing while neutral sentiment remained dominant. Many netizens, starting from Wu Liufang's identity as a retired "national" athlete, contrasted this state-representing image with her highly contrasting role as a TikTok "boundary-pushing" host. They pointed out the moral controversy and image conflict embedded in this juxtaposition, reflecting their critical engagement with the issue. This comparison highlighted moral disputes and image conflicts, reflecting public contemplation. Relevant authorities could leverage this perspective to guide rational discourse. During the second growth phase, all themes exhibited significant fluctuations, particularly themes 0 and 2, where most netizens showed positive support, while other categories maintained low emotional intensity. In the second decline phase, theme 0 remained prominent with majority positive sentiment, though its intensity significantly decreased compared to active phases. The calming phase witnessed rapid theme shifts and volatility, indicating sustained public engagement. As all three emotional intensities dropped further, tensions gradually subsided. Therefore, authorities should implement targeted responses to address various public concerns and ease public anxiety.

This study takes the "Wu Liufang Incident" on the TikTok platform as a case study. Through computational communication analysis of 28,697 valid comments, it systematically addresses five research questions and reveals multidimensional characteristics of public opinion during retired athletes' career transitions. The findings demonstrate that public discourse exhibits high complexity in emotional expression, thematic focus, and evolution patterns. This complexity reflects both widespread support for individual career autonomy and critical reflections on systemic inadequacies in institutional safeguards.

The evolution pattern of public opinion. This study demonstrates that the "Wu Liufang incident" followed a bimodal evolution model (Jiang, 2014), with two peaks directly triggered by the TikTok platform's "ban" and "unban" operations. This stands in stark contrast to traditional event-driven public opinion dynamics (e.g., natural disasters, public health crises), highlighting the pivotal role of platform governance in shaping the rhythm of public sentiment (Bennett & Livingston, 2018). These findings indicate that in platform-centric societies, public opinion analysis must prioritize platform regulations and intervention measures as critical variables.

Public sentiment.Public sentiment analysis reveals a balanced mix of positive and rational attitudes, with 74.5% of respondents expressing optimism, 24.0% neutral, and only 1.5% pessimistic. This distribution is closely related to the identity of retired athletes as "winning glory for the country" and the legitimacy of their career choice to "make a living lawfully" (Lavallee & Wylleman, 2000). By constructing a domain-adapted three-dimensional sentiment lexicon, this study effectively identified high-empathy terms such as "support", "livelihood" and "retirement", as well as legitimacy defense phrases like "not illegal" and "neither stealing nor robbing". This approach reveals a public value consensus grounded in the prioritization of the right to subsistence. (Wang & Yang, 2021). Emotional responses closely mirrored event milestones-account suspensions and reinstatements, third-party public welfare disclosures-all triggering peak emotional reactions. This confirms social media's acute sensitivity to external stimuli (Liu & Liu, 2018).

Thematic Analysis. The four core issues identified through thematic modeling further refine the focal points of public discourse. "The Legitimacy of Career Transition" (Thematic 0, 3, 6) and "Survival Pressure and Lack of State Protection" (Thematic 2, 4, 7) form the main body of public opinion. While supporting individual career freedom, the public strongly criticizes systemic inadequacies, particularly deficiencies in economic compensation, institutional staffing allocation, and vocational training systems (Wang et al., 2024). "Athlete Image Conflict" (Thematic 1, 5, 7) reflects the clash between traditional sports elite narratives and market-driven professional logic. Some citizens uphold normative expectations of "national athlete" identity, while others deconstruct moral coercion through irony, demonstrating modern value differentiation (Gustafsson et al., 2017). The "Irony in Online Communication" (Thematic 5, 6) section uses symbolic expressions like "Jiuzhuan Dachang" (a traditional dish) and "How Much Is a Pound of Feather?" to lower discussion barriers and facilitate cross-circle dissemination of public opinion. However, this approach may undermine the seriousness of core issues (Peng et al., 2024).

Key Characteristics. Public discourse exhibits a "core-periphery" spatial distribution pattern (with Guangdong, Zhejiang, and Jiangsu provinces accounting for 32.6% of total comments), which is closely associated with economic scale, internet penetration rates, and algorithmic recommendation mechanisms (National Bureau of Statistics, 2024; Mo, 2024). Overseas comments constitute 27.3% of the total, spanning 58 countries, demonstrating transnational attention to the issue and expanding international comparative perspectives in retired athlete studies.

In conclusion, this study employs sentiment analysis, thematic modeling, and evolutionary cycle analysis to reveal a dual attitude in the "Wu Liufang incident" where public support for retired athletes' career transitions coexists with institutional criticism. Public discourse not only focuses on the legitimacy of individual survival rights and professional freedom, but also extends to a systematic reflection on the implementation challenges of retirement athlete resettlement policies (Zhu & Zou, 2021). This provides empirical evidence for understanding sports-related public opinion in the social media era and offers a reference point for policy optimization.

Research Findings. This study examines the "Wu Liufang incident" on TikTok platforms using computational communication analysis to analyze public sentiment patterns during retired athletes' career transitions. The results demonstrate that public attitudes toward this transition remain predominantly positive, with neutral and rational sentiments dominating emotional responses while negative emotions account for a relatively small proportion. Thematic analysis identified four key discussion themes: legitimacy of career transitions and public support, survival pressures and gaps in state welfare, ethical controversies and image conflicts, as well as satirical expressions in online discourse. These findings provide new perspectives for understanding social media sentiment dynamics and offer valuable references for policy development.

Research Contributions. Theoretically, this study fills the gap in public discourse analysis regarding career transitions of retired athletes, expanding the research perspective on retired athletes. Previous studies predominantly approached the subject from sports science and policy angles, whereas this research integrates methods from journalism and computational communication to reveal the complexity and dynamism of social media public opinion. Sentiment analysis demonstrates the coexistence of rational thinking and emotional expression in public discourse, validating the "dual-process model" (Marcus et al., 2000) in public opinion research. Thematic modeling identifies associations between "survival pressure" and "national security guarantees", providing empirical support for the "institutional transition" theory in sports sociology. Additionally, the study's exploration of image conflicts between retired athletes and popular male streamers has bridged a gap in media and identity research within the sports domain.

At the methodological level, this study enhances text analysis accuracy by constructing a three-dimensional sentiment dictionary and optimizing the Latent Dirichlet Allocation (LDA) topic model. The sentiment dictionary construction fully considers the unique characteristics of the sports field and TikTok platform, including terms like "Olympic athletes", "internet celebrities", "streamers" and "heart wishes".

From a social perspective, the findings of this study offer valuable insights for improving retirement athlete resettlement policies. While public support for career transition freedom reflects evolving societal attitudes, criticisms regarding inadequate state guarantees highlight implementation challenges. Relevant authorities should prioritize addressing the survival pressures faced by retired athletes and resolve the "supply-demand mismatch" through institutional optimization. Meanwhile, the entertainment-oriented tendencies on social media platforms may obscure core issues, necessitating targeted guidance to prevent irrational public opinion spread.

Research Limitations and Future Directions. However, this study also has certain limitations. The data sources are limited to official media comments from TikTok platforms, which may not fully represent broader public opinion. Additionally, while the constructed sentiment lexicon and topic models have undergone manual validation, there remains room for improvement. Future research could expand the data scope by incorporating multi-platform comparative analysis to more comprehensively reflect public opinion characteristics. Simultaneously, introducing more multidimensional variables may help deepen our understanding of the driving factors behind public sentiment dynamics.