Social image presentation of virtual sports in social media through the analysis of Twitter data

Abstract

Virtual sports have become an important site for shaping public value alignment through its social representation on digital platforms, particularly within social media environments. The current study examines public discourse on X(Twitter) surrounding virtual sports during the period from the 2021 olympic virtual series to the 2023 olympic esports week. Informed by framing theory, our methodology integrates computational approaches to analyze a dataset of 15,585 tweets. We employed topic modeling to identify salient thematic patterns and conducted sentiment analysis leveraging large language models to evaluate affective dimensions within public discourse. The study identifies six core thematic dimensions that characterize the social image of virtual sports. Although overall sentiment leans positive, the distribution across themes is uneven. Notably, the themes related to event-based mobilization exhibit a marked tendency toward negative sentiment. Overall, the study offers a systematic account of how the social image of virtual sports is presented in social media discourse, providing a valuable reference point for the future construction and communication of virtual sports’ public identity.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-36725-6.

Introduction

The wave of digital transformation has propelled virtual sports into the spotlight, creating an entirely new realm of competitive athletics that captivates audiences worldwide. Virtual sports harness advanced technologies, such as virtual reality (VR) and augmented reality (AR) to enable remote embodied participation and the application of actual athletic skills within digital environments^1,2.

Indeed, virtual sports increasingly leverage social media platforms to achieve greater public visibility. A recent example is the 2024 ITTF World Esports Table Tennis Championships, which used YouTube to livestream the event globally for free³, thereby projecting an image of openness and global accessibility. Existing scholarship has predominantly focused on domains such as athletic skill enhancement⁴, improvements to physical and mental well-being⁵, injury prevention and rehabilitation⁶, and pedagogical applications in physical education^7,8. However, the construction of virtual sports’ image within public discourse has been the subject of relatively limited research. These gaps not only hinder the identification of stakeholder-specific perspectives on virtual sports, but also leave unaddressed how different thematic frames and sentiment categories on social media influence the virtual sports’ social image. Therefore, examining public themes and sentimental trends in social media can not only enrich theoretical discussions about the social image of sport in digital spaces⁹, but also can hold practical relevance for stakeholders. For instance, policymakers require a clear understanding of public perceptions of virtual sports in order to develop comprehensive policies.

The research aims to fill this gap by applying framing theory to examine how the social image of virtual sports is thematically framed and sentimentally expressed in social media discourse. The overall analytical framework is illustrated in Fig. 1. Based on this framework, the study poses the following research questions:

Fig. 1.

Methodological flowchart for the study of the social image of virtual sports.

RQ1: What thematic frames are employed in social media discourses to construct the social image of virtual sports?

RQ2: How is sentiment distributed across different thematic frames of virtual sports’ social image on social media?

Results

Chronological display of tweets

Figure 2 shows the sample distribution of this study. Overall, the number of tweets has been on a downward trend. However, several notable peaks appear on June 23, 2021, November 17, 2022, March 2, 2023, and June 23, 2023. These peak dates correspond to critical milestones in the development of virtual sports, including the conclusion of the Olympic Virtual Series, the announcement of the Olympic Esports Week, the release of its official competition events, and the commencement of the Olympic Esports Week itself.

Fig. 2.

Chronological distribution of virtual sports tweets.

Topic analysis

To address Research Question 1, an unsupervised Latent Dirichlet Allocation (LDA) model was utilized for thematic analysis. The Coherence and perplexity curves were jointly used to guide the model selection (Fig. 3). Coherence increases and stabilizes with more topics, indicating clearer semantic structure, while perplexity decreases with diminishing improvement, suggesting that the selected model balances interpretability and complexity. LDA tuning combined with CV-based coherence evaluation showed that six-topic model (α = 0.13, β = 0.78) achieved a coherence score of 0.47, while a nine-topic model (α = 0.11, β = 0.83) yielded a coherence score of 0.45. Both configurations demonstrated relatively high semantic coherence. Based on the combined evaluation of coherence and perplexity scores, we used pyLDAvis to visualize the topic distribution and determine the optimal number of topics for the LDA model. Figure 4 visualizes the latent topic structure using a pyLDAvis-based representation. Each circle represents a topic, where circle size corresponds to the relative prevalence of that topic in the corpus. The spatial distance between circles reflects semantic similarity, with closer topics sharing more overlapping vocabulary. Six topics were chosen as the optimal solution for parameter tuning in this study.

Fig. 3.

Coherence and perplexity scores of the LDA model.

Fig. 4.

Comparison of intertopic distance between the nine-topic model and the six-topic model.

This study summarizes the top 20 high-frequency keywords from each of the six LDA-derived topics, along with the two representative tweets associated with each topic. Given that the LDA model operates in an unsupervised manner, two experts were invited to review the high-frequency keywords and assign a meaningful label to each topic (as shown in Table 1). In case of the topic 0 focused on the esports and traditional sports, with keywords such as esports, real, like, play, IOC, committee, and Gran Turismo. It highlighted discussions around the resemblance between virtual and physical sports, the role of international institutions like the IOC, and player participation in virtual sports events modeled on real-world athletic disciplines. Hence, the topic is summarized as Integration of Electronic Sports and Traditional Sports. The other topics were also summarized in a similar manner based on the probability distribution of keywords. We ultimately identified six topics (Table 2).

Table 1.

Keywords and topics generated by the three-topic LDA model.

Topic	Keywords	Percentage of Twitter (%)
Integration of electronic sports and traditional sports(Topic 0)	Like, world, one, see, esports, play, committee, real, international, dance, would, dont, players, know, ioc, time, gran, turismo, think	2.53
Functions and participation of virtual sports(Topic 1)	Amp, register, join, free, learn, session, open, book, program, time, link, students, available, april, conference, info, registration, health, management, centre	43.56
Embedding educational scenarios in virtual sports(Topic 2)	School, conference, cup, press, schools, south, activity, took, summer, ceremony, africa, cricket, physical, rugby, world, youtube, desktop, opening, tour, journey	11.57
Digital extension of sports(Topic 3)	Cycling, golf, tour, insanity, weeks, join, safety, awesome, car, france,1st, starts, app, taekwondo, started, link, stages, today,3000kms, tips	21.61
Event-driven social mobilization(Topic 4)	Week, event, singapore, first, part, today, get, june, take, amp, day, year, athletes, next, women, events, great, place, inaugural, chess	1.77
Technology-enabled media communication(Topic 5)	Reality, team, fortnite, finals, live, sport, experience, amp, watch, player, betting, top, future, media, use, technology, become, leading, innovation, growing	18.94

Table 2.

Selected popular tweets per topic.

Topic	Content of the tweet
Integration of electronic sports and traditional sports	Our very own Joar is live right now on the Olympic Virtual Series to talk about performance anxiety and differencessimilarities between traditional sports and esports Tune in
	Applying highend technologies including virtual and augmented reality and merging that with traditional #Japanese motifs #OlympicGames in #Tokyo starts fresh chapter that lacked the glitz amp glamour of previous editions but was apt for the current situation ravaged by pandemic
Functions and participation of virtual sports	If esports becomes an Olympic sport Which will you Participation out of those
	When her university cuts her track team a hardheaded wannabe pro gamer defies her mother and abandons their family’s Olympic shotput legacy to join a mismanaged eSports team and its cocky nerdletes
Embedding educational scenarios in virtual sports	Another great event by @AyrshireColl HND Sport A Happening on WB May 10th Virtual Walk Killie to Killie Students will be hoping to reach 3500 miles Good Luck Its ok not to be ok @krisboydcharity @MctJohn @mental_united #PassingPositivity
	Something new for our students todaya virtual PE lesson with @SouthamptonFC
Digital extension of sports	Your experience will go beyond doing sport you will choose @onepeloton to be part of a community that creates immersive experiences with real people on a virtual playground quality networking without having to travel just get on your bike and put your VR glasses on
	Its 2022 Head over to Oculus TV for free quick challenging and immersive workouts in virtual reality Practice Yoga Dance Sculpt HIIT and Stretch in gorgeous natural environments #workout #vrfitness #exercise #fitness #sport #health @oculus LINK
Event-driven social mobilization	According to @iocmedia the new Olympic Virtual Series will mobilise virtual sport esports and gaming enthusiasts all around the world in order to reach new Olympic audiences
	@iocmedia Partners Five IFs To Produce Olympic Virtual Series The Olympic Virtual Series will mobilise virtual sport esports and gaming enthusiasts all around the world in order to reach new Olympic audiences #sportsbiz #events #Olympics
Technology-enabled media communication	The @iocmedia will partner with five International Sports Federations IFs including World Sailing and game publishers to produce the Olympic Virtual Series OVS the firstever Olympiclicensed event for physical amp nonphysical virtual sports
	IOC makes major esports move with Olympic Virtual Series launch SportsPro Media

In addition, Topic 1 accounts for the largest proportion of tweets (43.56%) and shows the most concentrated topic weight and keyword distribution. Topic 3 is the second most prominent, covering 21.61% of the tweets, followed by Topic 5 (18.94%), Topic 2 (11.57%), Topic 0 (2.53%) and Topic 4 (1.77%). However, a topic-based analysis alone provides only a partial view and does not fully reflect the broader social image of virtual sports, as it overlooks how different user groups frame and emotionally respond to these topics in their comments.

Sentiment analysis

To address Research Question 2, this study used the API of the Deepseek Large Language Model to identify the sentiment polarity of tweets. Among the total of 15,585 tweets, 58% were identified as positive, 28% as neutral, and 14% as negative. Overall, the tweets were predominantly positive in tone, reflecting users’ enthusiasm for participating in virtual sports and their approval of the integration of emerging technologies with traditional sports. For example, one user wrote: “Win a gold medal from your front room. IOC launches Olympic Virtual Series. OK, how do I enter? This is my big chance to become an Olympian.” However, some tweets also expressed negative sentiments, often involving concerns about the boundaries between the real and the virtual in the context of venues and events. These negative expressions suggest that, although realistic simulations may enhance audiovisual experiences, they may also diminish the sense of physical presence for participants and spectators. As one tweet noted: “@BrianRoemmele 75–100 years from now, our sport stadiums will be as derelict as these grand theaters and opera houses of the past. With entertainment being more virtual every year, it might even happen faster than that.” In addition, some comments conveyed a neutral tone through information sharing or events reporting, focusing on issues such as the applicability of regulations. For instance, one tweet stated: “DopingSome testing more education for gamers at Olympic Esports week SINGAPORE Reuters Hot on the heels of the World AntiDoping Agency WADA announcing that esports might ultimately come under its code”.

After delineating and examining each topic, we analyzed the sentiment polarity distribution within them. The results show that positive sentiment predominates across all six topics, albeit to varying degrees. Topic 0 (“Integration of Electronic Sports and Traditional Sports”) displays the highest proportion of positive sentiment (67.3%, 95% CI [62.7%, 71.9%]), with relatively low levels of neutral (21.6%) and negative (11.1%) sentiment. Topic 5 (“Technology-Enabled Media Communication”) similarly exhibits a high positive share at 73.6% (95% CI [72.0%, 75.2%]) and minimal negative responses (2.4%). By contrast, Topic 4 (“Event-Driven Social Mobilization”) shows a much more even distribution, with positive sentiment at 33.2% (95% CI [27.6%, 38.8%]), neutral at 31.0%, and negative at 35.8% (95% CI [30.2%, 41.4%]), indicating substantial divergence of opinion. Topics 1 (“Functions and Participation of Virtual Sports”) and 2 (“Embedding Educational Scenarios in Virtual Sports”) present intermediate patterns, where neutral sentiment (28.3% and 34.4%, respectively) approaches or exceeds one quarter of the discourse, suggesting that discussions within these topics are less dominated by overtly positive or negative expressions. Topic 3 (“Digital Extension of Sports”) also shows a high positive proportion (64.1%, 95% CI [62.5%, 65.7%]), but with a notable neutral share (29.4%).

Discussion

This study examines tweets from X users concerning virtual sports to analyze the presentation of themes and sentiments, thereby providing a descriptive account of how the social image of virtual sports is represented. Compared to previous research that has primarily focused on the image of esports¹⁰, this current study offers a unique context in which virtual sports have. For the first time, received official authorization from the International Olympic Committee and have subsequently attracted widespread attention. Interestingly, the most frequently discussed topic on social media is Functions and Participation of Virtual Sports, accounting for 43%. This finding is reasonable, as virtual sports communities on social media not only provide information and interaction but also create opportunities for value co-creation between fans and clubs, which may be associated with higher levels of public participation and immersive experiences¹¹. For example, expressions such as “If esports becomes an Olympic sport, which will you Participation out of those?” (@yuga). Moreover, words such as “amp”, “join”, “free”, and “available” are often mentioned, which is consistent with an interpretation of virtual sports as an “instrumental” entity. This interpretation also aligns with the immersive and presence-enhancing features of virtual reality and augmented reality technologies^12,13. However, the topic of “Embedding Educational Scenarios in Virtual Sports” remains marginalized, accounting for only 2%. However, the topic of “Embedding Educational Scenarios in Virtual Sports” remains marginalized, accounting for only 2%. This finding differs from previous scholars’ arguments that virtual sports enhance students’ motivation to participate in physical education courses and foster deeper knowledge retention¹⁴. One possible explanation may lie in the lack of clear terminological distinction between virtual sports and esports¹⁵, which could contribute to the spillover that has led to the transfer of negative perceptions associated with esports. For instance, tweets such as “When her university cuts her track team a hardheaded wanna be pro gamer defies her mother and abandons their family’s Olympic shotput legacy to join a mismanaged esports team and its cocky nerdletes” (@AustinElliott) illustrate this phenomenon. This pattern may suggest that, on social media platforms, the social image of virtual sports is narrowly understood as a symbol of entertainment, making it difficult to be integrated into school-based physical education systems.

This study reveals the public’s sentiment towards the social image of virtual sports on social media. By further classifying the emotional tendencies of each theme, this study reveals the sentiment concerning the social image of virtual sports on social media. The findings indicate that the public generally holds a positive attitude toward virtual sports, thereby corroborating previous conclusions^16,17. However, the theme of “Event-Driven Social Mobilization” stands in sharp contrast to other themes, as it demonstrates the highest proportion of negative sentiment and considerable controversy. One possible explanation for this contrast may be related to broader concerns about platform governance and algorithm transparency reported in prior studies¹⁸. In this sense, the observed negative sentiment may reflect skepticism toward event-driven digital mobilization rather than direct evaluations of virtual sports themselves. Such skepticism has been discussed in the literature as arising in contexts where organizational decisions, often made for reasons such as data security or operational efficiency, are perceived by some audiences as commercially motivated. These perceptions may be associated with contested views of fairness and reduced levels of trust.

Method

Data collection and procedures

Under the dual impact of structural changes in society and the COVID-19 pandemic, traditional sports activities have been hindered, while virtual sports have developed rapidly, becoming a new way to meet people’s sports needs. The International Olympic Committee elevated virtual sports onto the global stage by hosting the Olympic Virtual Series in 2021 and the Olympic Esports Week in 2023, attracting participants worldwide. By combining advanced technologies with traditional athletic components, virtual sports offer immersive environments that enhance user engagement and support physical activity. Against this backdrop, virtual sports have attracted widespread attention—particularly on X, where users rapidly initiated a cascade of tweets and comments to share their perspectives and emotions. Accordingly, this study collected, processed and analyzed Twitter data as outlined in Fig. 1.

Data for this study were collected using Twitter’s official API, which allows researchers to retrieve user-generated content and associated metadata based on predefined keyword conditions. The metadata obtained included timestamps, user IDs, post titles, hashtags, as well as engagement metrics, such as likes and comments. A customized web crawler was developed to send data requests at fixed one-minute intervals. This design aimed to ensure data collection efficiency, while strictly adhering to ethical standards for web-based research. By spacing requests appropriately, the program minimized the risk of overloading Twitter’s servers and avoided interference with regular user activity on the platform. The data collection process focused on keywords closely associated with virtual sports, including “Virtual Sports,” “Olympic Virtual Series,” and “Olympic Esports.” Two graduate research assistants reviewed the initial tweet corpus to identify recurring, relevant terms and phrases. This iterative process continued until no new keywords emerged, indicating saturation (Table 1).

Major events serve as critical reference points for segmenting the temporal scope of this study. It is interesting to track public perceptions and evaluations of virtual sports following different major events. The data collection period spans from April 22, 2021, the date when the Olympic Virtual Series was officially announced, to June 25, 2023, marking the conclusion of the Olympic Esports Week. Within this timeframe, a preliminary dataset of 26,990 tweets related to virtual sports was obtained, capturing user discourse across key developmental stages of the field.

Data preprocessing

To enhance data quality and ensure the extraction of analytically meaningful textual features, a structured preprocessing workflow was applied to the raw Twitter dataset, using Python-based tools. Initial standardization procedures were conducted to maintain data integrity and prepare the dataset for subsequent analysis. Duplicate tweets generated during the crawling process were identified and removed to ensure uniqueness of observations.

For textual data, a multi-stage cleaning and normalization procedure was implemented using Python’s re library and the NLTK. Tweets were first filtered to retain only English-language content from the multilingual corpus. Word-level tokenization was performed using NLTK utilities, followed by normalization steps, including lowercasing and the removal of non-informative elements such as URLs, mentions, emojis, punctuation, and other extraneous Unicode characters via regular expression matching. Common English stop words were excluded using the standard NLTK stopword list, supplemented with additional high-frequency terms identified as semantically irrelevant during exploratory inspection. To reduce lexical sparsity and improve semantic consistency, stemming and lemmatization were applied, and tokens occurring below a minimum document frequency threshold were excluded from further analysis.

Beyond rule-based preprocessing, semantic validation was conducted using the Deepseek large language model to identify and remove irrelevant or off-topic content. This additional filtering step enhanced the relevance and analytical validity of the corpus. Following the completion of all preprocessing and validation steps, the final dataset comprised 15,585 tweets. Figure 1 provides an overview of the data collection and preprocessing pipeline.

Topic modeling

The Latent Dirichlet Allocation (LDA) model is a Bayesian, generative topic model that efficiently uncovers latent themes and underlying semantic structures within large text corpora¹⁹. It has been widely applied to social media text analysis²⁰. To accommodate the heterogeneous characteristics of our multi-source tweet dataset, we employed a symmetric prior configuration to mitigate bias in semantic representation. By equalizing the Dirichlet prior over topics and words, this approach prevents any one semantic domain from dominating the model, thus ensuring balanced interpretation across diverse discourse structures.

Guided by precedent in the literature^20,21, we used Python’s Gensim library to perform automated semantic identification. We trained LDA models with topic numbers ranging from 3 to 11, and determined the optimal number by evaluating perplexity and coherence metrics, complemented by interactive visualizations via pyLDAvis. Form a specific theme presentation by combining literature and key words.

Sentiment analysis

Leveraging the precision of generative large-language models in detecting user sentiment²², we accessed the Deepseek API to classify each cleaned tweet into positive, neutral, or negative sentiment categories. In configuring the model, we adhered to best practices from recent studies^23,24 and optimized for our context. Specifically, we set “temperature” to 0.3 to balance diversity with result stability, capped “max tokens” at 100 to focus on core sentiment cues, and left both “frequency penalty” and “presence penalty” at 0 to preserve semantic completeness. To validate classification reliability, we conducted multiple rounds of verification: 2,000 randomly sampled tweets were manually coded by two doctoral students in big-data science. The inter-coder agreement yielded a Krippendorff’s Alpha of 0.91, and the concordance between model labels and human codes reached 87.5%, demonstrating robust sentiment classification. Additionally, we implemented a ternary scoring scheme to quantify sentiment polarity for subsequent analysis. Tweets with positive sentiment were assigned scores ranging from 0.5 to 1.0, neutral sentiment was assigned a fixed score of 0.5, and negative sentiment was scored within the range of 0 to 0.5.To rigorously assess the statistical reliability of the sentiment distributions across topics, a non-parametric bootstrap analysis was employed. Specifically, 1,000 resampled datasets were generated with replacement for each topic from the original tweet corpus. Sentiment proportions were then recalculated for each iteration to derive empirical 95% confidence intervals (CIs). This procedure quantified the uncertainty inherent in the sentiment estimates, ensuring that the observed thematic differences were statistically robust rather than artifacts of sampling variance.

Limitations

From a machine learning and causal inference perspective, this study is primarily descriptive in nature. While the identified thematic prevalence and sentiment distributions are supported by the data, the analysis does not test causal mechanisms. For instance, interpretations linking negative sentiment in the “Event-Driven Social Mobilization” theme to platform governance or algorithmic transparency should be understood as tentative rather than directly validated. Furthermore, the absence of comparative control datasets limits the ability to attribute observed sentiment patterns specifically to virtual sports rather than to broader event-driven or Olympic-related communication dynamics. Finally, regarding the topic assignment uncertainty inherent in LDA, this study utilizes dominant topic assignment. Future studies could benefit from soft-clustering approaches to better capture semantic ambiguity at the document level.

Future directions

Future research may build on the present findings by incorporating comparative control datasets, such as baseline sports discourse, sports-only discussions, or non-Olympic virtual sports events, to better distinguish patterns specific to virtual sports. In addition, the application of inferential and longitudinal analytical approaches may help clarify the robustness and temporal dynamics of thematic and sentiment patterns observed in social media discourse.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization, W.D. and C.X.; methodology, C.X.; software, C.X.; validation, W.D., C.X. and S.Z.; formal analysis, X.S.; data curation, X.S.; writing—original draft preparation, W.D and S.Z.; writing—review and editing, T.M and S.Z.; visualization, X.S.; supervision, W.D.; project administration, T.M. and X.S.; funding acquisition, C.X. All authors have read and agreed to the published version of the manuscript.

Data availability

The data that support the findings of this study are not publicly available due to sensitivity, but are available from the corresponding author upon reasonable request. Data are located in controlled access data storage at School of Intelligent Sports Engineering, Shanghai University of Sport, Shanghai, China.

Declarations

Competing interests

The authors declare no competing interests.