‘Your Strength Is Inspirational’: How Naomi Osaka’s Twitter Announcement Destigmatizes Mental Health Disclosures

Abstract

On May 31, 2021, Naomi Osaka, one of the top-ranked female tennis players, and one of the highest-paid female athletes in the world, announced her withdrawal from the French Open on her social media (Twitter) account, citing mental health issues. There exists a stigma around mental health; and people suffering from mental health conditions often experience “discrimination and stigma” (World Health Organization, 2019). Such disclosures by a noted sportsperson provide an opportunity to help combat the stigma. The present study uses unsupervised machine learning and qualitative thematic analysis to analyze 11,800 English language responses to her tweet. Results indicate that Osaka’s tweet mostly garnered a lot of support and encouragement. However, there also existed some negative comments. Additionally, 40% of the negative comments were disseminated by bot-like automated accounts. Practical implications for sports communication are also discussed.

Keywords

osaka athlete mental health stigma unsupervised machine learning

The pressure athletes — amateur or professional— face to perform at elite levels looms as a vexing part of sport (e.g., Kroshus et al., 2019; Parrott et al., 2020). With increasing amounts of both money and media exposure (e.g., Parrott et al., 2020; Pedersen et al., 2020; PricewaterhouseCoopers, 2021), some athletes are struggling to balance their sporting responsibilities and their mental health. The present study explores the social media fallout from Japanese tennis star Naomi Osaka’s very public withdrawal from the 2021 French Open, one of the sport’s four major annual tournaments. During the event, Osaka, a four-time major tournament winner, announced her withdrawal via Twitter, citing that facing the press caused her great anxiety (Arnold, 2021). In her post, Osaka also said, “I would never trivialize mental health or use the term lightly. The truth is that I have suffered long bouts of depression since the US Open in 2018 and I have had a hard time coping with that” (Osaka, 2021).

Thanks to social media, athletes and other public figures can now frame their own narratives, bypassing traditional gatekeepers, such as journalists and public relations experts (Sanderson & Kassing, 2016). Thus, discussions that might have taken place in private or been reframed by mass media professionals are now occurring unfiltered in front of an interactive global audience. With the stigma that currently surrounds mental health, (Corrigan & Watson, 2002; Gwarjanski & Parrott, 2018) could these public disclosures and discussions by a top performing athlete potentially reduce stigma? Additionally, because of social media’s multidirectional communication elements, non-elites — aka the public — can also reframe the debate on societal issues, inserting overlooked perspectives, thereby challenging hegemonic ideas (Locke et al., 2018). For example, the dominant narrative around mental health blames problems on the individuals themselves (Corrigan & Bink, 2016), leaving the dominant societal structures uncritiqued (Major & Jankowski, 2020). Social media can create new spaces to redirect the attribution of blame in regard to mental health problems.

The situation with Osaka’s withdrawal from the French Open and the voluminous social media reactions provide a valuable opportunity to explore such concepts. The present study uses unsupervised machine learning to understand the fans and other social media reactions, and examine the comments through the lens of previous literature on sports, mental health, and stigma. Another fundamental question that the study aims to address is “who” is contributing to stigmatizing social media reactions. Prior research suggests that in addition to human users, social bots (automated computer programs mimicking human behavior) are increasingly becoming influential voices in the digital ecosystem (e.g., Kollanyi, 2016; Shao et al. 2018). The current study also focuses on detecting the presence and role of social bots in stimulating stigmatizing conversations and responses to Osaka’s tweets.

Stigmatization and Mental Illness

Depression and anxiety are two prominent mental health conditions resulting in a global yearly economic loss of $1 trillion (WHO, 2019). However, according to the World Health Organization (WHO), people suffering from mental health conditions often experience “discrimination and stigma” (WHO, 2019). Stigma is defined in myriad ways across literature from different disciplines including sociology (Goffman, 1963), psychology (Link & Phelan, 2001), and communication (Smith, 2007). For the purpose of the present study, we define stigma as a social construction put in place that devalues an individual or a group of individuals for possessing a characteristic that deviates from the societal norm (Smith, 2007).

According to Corrigan and Watson (2002), the stigma around mental illness expresses itself in two ways: internalized stigma and public stigma. Internalized stigma refers to the self-perceptions of members of a stigmatized group. The term “public stigma” refers to the general public’s unfavorable attitudes and perceptions about mental illness, which contribute to stereotyping, prejudice, and discrimination against people with mental health problems (Corrigan & Watson, 2002). Both public and internalized stigma have negative effects on the people suffering from mental illness. Prior research indicates that public stigma deters an individual from seeking treatment and also leads to prejudice and discrimination (Corrigan & Shapiro, 2010). Similarly, internalized stigma can lead to reduced self-esteem, and in some cases reduces the prospects of recovery (Watson et al., 2007).

According to Corrigan and Bink (2016), three main stereotypes associated with people suffering from mental illness contribute to its subsequent stigmatization. The first stereotype is that of responsibility in that people suffering from mental illness are responsible for their own condition and the attribution of blame is often directed towards the person either as offset or onset attribution (Weiner, 1995), rather than critiquing the social structure. The second stereotype is that people suffering from mental health issues are not competent enough to do their job. The third stereotype is that people suffering from mental health are often considered dangerous.

While there exists a stigma around mental health, the framing of mental health-related events can affect how information receivers respond to the message (Gwarjanski & Parrott, 2018). For example, Gwarjanski and Parrott (2018) found that online comments in response to news stories about schizophrenia were more positive and prosocial when the news stories’ frames were more supportive of people with schizophrenia. Researchers have also identified strategies that help with the destigmatization of mental illness. The three most commonly identified strategies are “protest,” “education,” and “contact” (Corrigan & Penn, 2015, p. 5). Protest strategies generally include calling out inaccurate stereotypes making it more of a “reactive strategy” (Corrigan & Penn, 2015, p. 5). Education strategies replace negative stereotypes about mental health and people suffering from mental health with accurate information that is more supportive. Finally, the contact strategy — which is deemed by previous studies as the most powerful strategy — includes interpersonal contact with people suffering from mental health issues (Corrigan & Penn, 2015). However, the important caveat for a contact strategy to be effective is the role of the opportunity to “meet” with the group, regardless of medium. Parrott et al. (2020) argue that an athlete’s disclosure about their mental health struggles and directly talking with their audience via the athlete’s social media pages can serve as a “parasocial contact.” Therefore, such disclosures (including Osaka’s) serve as an important tool to help with destigmatization efforts.

Sport, Society, and Stigma

The positioning of sport in myriad societies makes discussing mental illness very difficult (Parrott et al., 2020). In fact, society often views sport as a respite from the difficulties and pain of daily real life (Segrave, 2000; Wann et al., 2004). Through the consumption of sporting events and now social media interactions with athletes, fans often experience a parasocial relationship with athletes (Parrott et al., 2020; Sanderson & Kassing, 2016). Essentially, the fans perceive they have a bond with an athlete or celebrity — even though they never met and do not know each other — through watching them perform and by consuming their media messages. The construction of sport itself supports the idea of sport as an alternate universe when one considers that many sports can suspend or bend rules that exist in general society. For instance, some violent acts are tolerated in sports that would be shunned outside of sport (Coakley, 2021).

In fact, athletes are valorized for playing while injured (Coakley, 2021). Coakley details that athletes who do not perform while injured can be shunned by teammates and fans. Historically speaking, an athlete’s body and labor served their city-states, leaders, or patrons (Scanlon, 2006). Through the heavy commercialization of modern sport, one can argue that athletes’ bodies primarily serve capitalistic purposes (Kalman-Lamb, 2019). Thus, athletes have value only when they perform. Though not often demonstrated as a visibly discerning injury, mental illness can also be debilitating to an athlete (Parrott et al., 2020). Recognizing that athletes are regular people helps shatter the heroic myths that accompany sports figures (Segrave, 2000). And so, when athletes acknowledge their mental health struggles, it could be a space to destigmatize mental illness because of athletes’ parasocial interactions with the public (Parrott et al., 2020). However, if fans adopt more of an ownership mentality regarding athletes’ labor (Kalman-Lamb, 2019), such disclosures could be perceived negatively. Additionally, such an ownership mentality could feed the “disruptiveness” dimension of Jones et al.'s (1984) conceptualization of stigma as such disclosures and subsequent withdrawal (in the case of Osaka) can impact the relationship between the athlete and the fan. Essentially, when the fans are consuming sport, they engage with the athlete and experience a relationship. A mental health episode breaks that relationship.

Prior research in sports communication has examined the news coverage framing of athletes disclosing mental health issues. Parrott et al. (2021) analyzed media coverage of National Basketball Association star players Kevin Love and DeMar DeRozan after both announced their struggles with depression and anxiety. The study found that the news stories were generally positive as journalists frequently focused on athletes’ stories and embedded a call to action to end mental health stigma. Journalists’ “shift” in mental health coverage toward focusing on their players’ individual narratives was also found in an analysis of articles comparing DeRozan and fellow U.S. basketball player Royce White (Cassilo, 2022, p. 111). Additionally, Thompson et al.’s (2022) study of newspapers’ coverage regarding U.S. gymnast Simone Biles at the 2020 Olympics — held in 2021 — also found generally positive coverage about Biles’s mental health struggles.

Prior research has also examined social media audience reactions to other athletes who disclosed their mental health issues. Parrott et al. (2020) examined the Twitter audience’s reactions to the mental health disclosures of Love and DeRozan. The study found that the response garnered on social media was mostly positive for both athletes. Similarly, Pavelko and Wang (2021) analyzed the role of hope in the audience’s reactions to Love’s Instagram post sharing the op-ed piece written by the player wherein he details his personal struggles with mental health.

While some significant studies have looked at media framing and social media audience responses to athletes’ mental health struggle disclosures, Osaka’s withdrawal, which resulted in her not performing for fans during one of tennis’s elite events, presents a unique opportunity to study the overall destigmatization of mental health and social media via an athlete who is a non-Western citizen participating in an international tournament. Osaka’s withdrawal citing mental health issues feeds into the stereotype that people suffering from mental health are not competent enough to do their job (Corrigan & Bink, 2016). Most of the athletes previously studied either continued to play their sport or returned to the competition. Prior research on social media users’ responses has not fully explored how the public attributed the responsibility or blame for the athlete’s mental health struggles as well. According to the causal attribution model developed by Weiner (1995), when a negative event is attributed to personal responsibility, it leads to anger and subsequent discriminatory behavior due to diminished willingness to help. On the other hand, not attributing personal blame leads to pity, which leads to diminished discriminatory behaviors and a greater willingness to help. Understanding how social media users assigned blame in a negative situation where the athlete did not perform for the audience helps enhance the current literature on public testimony of athletes and its impact on the destigmatization of mental health. Therefore, the present study seeks to examine the following research question:

RQ1:

What were the themes of responses garnered by Osaka’s tweet that announced the withdrawal (RQ1a), and how were social media responses attributing blame (RQ1b)?

Bots as an Agent of Information

Bots are computer programs and software systems that run certain automated tasks in the digital space at a quick scale and pace without the control of human users (Ferrara et al., 2016). Bots can be of different types and rely on artificial intelligence and algorithms with specific codes and scripts to perform designated tasks (Howard et al., 2018). Most central to the purposes of the current study are social bots. These bots are defined as “a computer algorithm that automatically produces content and interacts with humans on social media, trying to emulate and possibly alter their behavior” (Ferrara et al., 2016, p. 96). Social media platforms like Facebook and Twitter have millions of social bots. Varol et al. (2017) “estimate that the percentage of Twitter accounts exhibiting social bot behaviors is between 9% and 15%” (p. 280). The human tasks simulated by bots range from simple activities like liking or sharing predefined content to more advanced activities like carrying one or many-sided communications in social networks (Grimme et al. 2017). Bots can perform these tasks autonomously and repetitively at a much faster speed than humans (Shao et al., 2018).

Based on the intended function and purpose, the bots can be both useful and harmful. Positive aspects of bots include “delivering news or automatically generated updates for a service” (Kollanyi, 2016, p. 4933). On the other hand, bots can be malicious in nature such that they can be used to systematically perform tasks like spreading misinformation and amplifying conspiracies (Shao et al. 2018).

These inauthentic actors proliferate vitriol and hate speech on social media platforms, drawing attention and visibility to the topic and helping it gain popularity (Albadi et al., 2019). Uyheng & Carley (2020) used machine learning and network analysis to investigate bot-driven racist hate speech in tweets about the COVID-19 pandemic in the United States and the Philippines. The study found that hostile bots influenced the Twitter conversation by controlling the flow of information and disseminating higher hate in the form of racist remarks in both the U.S. and the Philippines. Sport is not immune to negative bot activity. In a study examining bot behavior regarding National Football League athlete protests, Yan et al. (2021) found that bots on Twitter amplified “antagonistic perceptions” (p. 105).

Prior research on social media and health indicates the presence of stigmatized messages on social media. For example, Lydecker et al. (2016) found obesity-related stigma-laden tweets in their dataset. Similarly, from a mental health perspective, Reavley and Pikington (2014) found stigmatizing tweets related to depression and schizophrenia. Additionally, prior research has also documented that athletes face online abuse (Kavanagh, Jones, & Sheppard-Marks, 2016). However, previous studies have not examined the presence and role of bots in the spread of stigmatizing content potentially leading the way to discrimination when responding to an athlete’s social media posts. The current study aims to fill this gap and seeks to examine the following research question:

RQ2:

Is there any presence of social bots in the response tweets to Osaka’sTwitter post that were spreading stigma-laden content?

Methods

Data Collection

The study used the package “Tweepy” on Python, an open-source programming language to retrieve comments posted in response to Osaka’s original disclosure tweet on May 31, 2021. A total of 11,801 responses in English were analyzed. Data were collected from May 31, 2021, to July 30, 2021, to capture the maximum number of comments. The tweets from Python were transformed into an Excel format for the researcher’s ease of reading.

Unsupervised Machine Learning

The study utilized unsupervised machine learning to explore the themes of the reponses garned by the original tweet. Specifically, the study utilized Latent Dirichlet allocation (LDA; Blei et al., 2003). LDA is a probabilistic-based unsupervised machine learning method that uses Bayesian statistics and a bag of words approach to quickly and efficiently discover major themes or patterns in text (Maier et al., 2018). LDA is a popular topic modeling technique that has been gaining popularity in mass communication research (Maier et al., 2018) and has recently been used in sports communication to analyze coverage of Paralympics (e.g., Yoo et al., 2024). The present study has used the approach recommended by Maier et al., (2018) to align the methodology to suit communication research.

Dataset preprocessing

First, the researchers cleaned the dataset by removing duplicate tweets and empty blank spaces. Next, the dataset was preprocessed. This is an important step to ensure the elimination of any undesirable characters that could potentially impact the result as the LDA model uses a bag of words approach wherein the model looks for a combination of words to determine the topics. The researchers removed punctuation from all tweets. Next, non-alphanumeric characters were removed. Then, the entire text was converted into lower case. Next, all the hyperlinks were removed. Finally, tweets that contained less than 5 words were also eliminated as it can be hard for the algorithm to associate short texts with specific topics (Maier et al., 2018) and previous research have also adopted this procedure (e.g., Kwon & Park, 2022). All these steps were performed on Python using the Gensim library (Rehurek & Sojka, 2011), a popular open source library used in prior communication research (e.g. Guo et al., 2016).

Dataset preparation

Stopwords or generic words like prepositions were removed. Next, the words were converted into their root form in a process called lemmatization. For example, “rooting” would convert into the base form “root.” The dataset was then tokenized, wherein all the texts were converted into smaller pieces called “tokens.” This study employed TweetTokinizer on Natural Language Toolkit (NLTK; Loper & Bird, 2002) library on Python to perform this step. After tokenization, stopwords generated by lemmatization were removed. To improve LDA’s model performance, Maier et al. (2018) recommend relative-pruning, wherein frequently occurring words (over 90% of the time in the dataset) and infrequently occurring words (less than 5% of the time in the dataset) are eliminated. Finally, the researchers used the term frequency-inverse document frequency (TF-IDF) to weigh each token based on its dataset appearance. This determines token importance.

Topic determination

Maier et al. (2018) suggest tuning the data by running multiple models and systematically varying model parameters. This is done so human coders can interpret the topics. This study applied Roder et al.’s (2015) Cv coherence value measure to the LDA tuning run to determine the optimal number of dataset topics. Coherence measures semantic similarity between topic-related words. It uses normalized pointwise mutual information (NPMI) and cosine similarity to score co-occurring words (Roder et al., 2015). Higher coherence scores indicate topic coherence. This study used the tuning dataset on topics 1-8 labeled ‘k' and varied two hypertuning parameters (α & β) from 0.01 to 1.0. The researchers calculated and plotted Cv values for each model parameter combination to determine which topic had the highest Cv scores. After determining Cv values, the data visualization tool pyLDAvis (Sievert et al., 2014) was utilized to select the optimal number of dataset topics.

Assigning topics to tweets

After determining the topic numbers, the LDA model’s final step is to determine which tweets are associated with the topic. This is accomplished by utilizing Gensim’s feature vectors (Rehurek & Sojka, 2011). This classification assigns a probabilistic ratio to each tweet based on the word weight and the subject number. The algorithm assigns a probabilistic ratio from .001–1. For this study, for the tweet to qualify for a topic in the dataset, the probability should have been above .50.

Thematic Analysis

While the LDA can cluster tweets into topics, human interpretation is still required to understand what is being said and label the topics. Therefore, a thematic analysis was used to identify the topics. The researchers randomly selected 20% of the tweets from each topic for the purpose of the thematic analysis. Once the tweets were winnowed, the researchers conducted a thematic analysis based on the guidelines from McCracken (1988). All the authors first jointly analyzed the tweets and highlighted the tweets that initially attracted their attention The tweets were then re-examined to make sure each tweet grouping was distinct (Lindlof & Taylor, 2011; McCracken, 1988).

The researchers then worked to discern connections between groupings (Lindlof & Taylor, 2011; McCracken, 1988). As such, the groupings fell into two broad categories: positive and negative tweets. From there, the positive tweets were categorized into three themes or topics: affirmation, critiquing the structure, and destigmatization through visibility. For the negative tweets, the dismissal of mental health concerns served as the dominant theme.

Detection of Bots

In order to detect if accounts were bot-like in the dataset, this study relied upon the popular tool Botometer (Davis et al., 2016) which has been used in prior research including the Pew Research center study on social bots by Wojcik et al. (2018). The tool uses supervised machine learning and specifically uses a Random Forest classification model. The tool is trained on a variety of datasets which are then classified as feature vectors (Yang et al., 2022). When a user inputs the Twitter username, the tool fetches 200 most recent tweets sent by the user, and then evaluates several parameters and compares those parameters to the training dataset (Yang et al., 2022). Finally, based on the rules of the classifier, each account is assigned a “bot score” ranging from 0–1; a score closer to 0 indicates human activity while a score closer to 1 indicates an automated account (Yang et al., 2022). The present study utilizes the Botometer-V4 API on Python.

While there are no set gold standards for setting thresholds on the “bot score” to identify if the account is a bot. In order to reduce the risk of false positives, this study adopted the standards set by Broniatowski et al., (2018). The researchers segmented the accounts into three categories: bot scores with less than .40 were labeled non-bot accounts; bot scores between .40, and .75 were labeled uncertain, and bot scores over .75 were labeled as bots.

Results

Results from the LDA Model

RQ1 aimed to examine the themes that emerged in the response to Osaka’s withdrawal announcement tweet and how users were attributing the blame. This study used an unsupervised LDA algorithm coupled with thematic analysis to examine RQ1. Results from the tuning run LDA and subsequent Cv value calculation indicated that 5 topics with α = .11 and β = .81 had a coherence value of .43, and 6 topics with α = .11 and β = .81 had a coherence value of .42, indicating that both had high coherence value scores. Figure 1 displays the coherence values chart for all the topics for the tuning run.

Figure 1.

Cv values of the LDA tuning run.

After running the final LDA models with 5 and 6 topics, the researchers visualized the findings with pyLDAVis visualization package. Visualization results indicated some overlaps within the topics for the LDA model, which contained 6 topics. However, the model with 5 topics had an overall good inter-topic distance. Hence, for the final analysis, the researchers considered the model with 5 topics. Figure 2 displays the topic visualization for both the models (5 and 6 topics respectively). Finally, Table 1 presents the top keywords generated by the LDA model.

Figure 2.

Intertopic distance for topics 5 & 6 on PyLDAVis.

Table 1.

Keywords Generated by the LDA Model.

Topic	Keywords	Percentage of tweets, %
1	Stand, thank, strong, love, respect, naomi, courage, much, support, proud	21.15
2	Care, health, mental, love, naomi, support, protect, wish, self, strong	28.4
3	Press, medium, athlete, sport, player, talk, conference, rule, fine, force	18.69
4	Tennis, hope, play, player, watch, soon, tournament, naomi, court, let	16.18
5	Understand, anxiety, help, depression, life, world, suffer, hope, introvert, struggle	15.55

Results from Thematic Analysis

As described in the methods section, the researchers followed the McCraken (1988) data analysis framework to identify the main themes for topics generated by the LDA model. Results indicated that the labels for Topic 1 were tweets that were applauding and respecting Osaka for standing up for herself. Example tweets from this topic included: “Thank you for standing up to the stigma. Your strength is inspirational,” “Hang in there and thanks for sharing your vulnerability and strength to persevere! You are an inspiring athlete, woman, person, human. This is huge, go you!” Labels from the tweets belonging to Topic 2 were asking her to prioritize her health and well-being. Examples from this topic include: “Dear Naomi, do what you need to do, mental health care is health care, is self-care, and comes before what others feel they need from you!” “Protect your health. Tennis can wait. We love you.” Labels from tweets belonging to Topic 4 were tweets that were looking forward to seeing her back playing on the court. Tweets from this topic include: “all the best to you hope to see you in the future you are great for the game,” and “You have to take care of the young lady in the mirror. The tennis world will be there when you get back.” All these topics together formed the themes of “affirmation.”

Tweets from Topic 3 were themed around the attribution of blame. These tweets indicate that the social media users were not attributing responsibility toward the player but actually placing the blame on external environmental factors including critiquing the structure, the organizers, and the journalist for putting the player’s well-being at risk. Tweets from this topic (and themes) included “I support you I also condemn the french open and tennis organizations requiring media time you are there to play tennis not pay for the tennis org leaderships new car,” and “the tournament organizers are just interested in making money they are never truly interested in the well-being of athletes the rights of players not to speak to the media must be respected anxiety is a health concern for some athletes and this must be considered”; “I’ve long held the viewpoint that sportspeople should be allowed to deny media access & have the express right to decline interviews if not mentally fit I’ve had the misfortune to watch many an athlete put under unnecessary pressure by smug questioning or callous phrasing Osaka.”

Tweets from Topic 5 were about the destigmatization of mental health through visibility wherein, several tweets were by users disclosing their own mental health issues. Examples include: “I suffer from anxiety as well when I’m in certain situations so I totally get Naomi. Trust me, it’s like suffering in silence and nobody should have to go through that. Naomi’s presence is very good for the game and she’s a lovely person. Let’s make it work guys!”; “I praise you for bringing your mental Health struggles to the world’s attention, you truly have won another grand slam by raising the stigma that this puts on people. To be fined for being physically injured would never happen. World needs to change. I commend your actions”; “I have depression too. Thanks for telling us about your condition. I support, respect, and love you from Japan.”

The negative tweets primarily fell under the theme of dismissal of mental health concerns. Negative tweeters focused on Osaka’s responsibility to perform in the tournament. Some comments: “You’re just a spoiled athlete. I have terrible anxiety, but do my job. Get over yourself” and “ITS PART OF YOUR JOB!!!! Get professional help or quit. Compared to nurses, doctors, aged care workers, emergency staff you have it so easy Naomi. Do you think you’re the only introverted person to ever play the game. Would you be saying/doing this if you were ranked 80 #no”. The latter comment reflects a job-role tension in the tweets. The negative tweeters also emphasized that Osaka is well compensated and thus should not be criticizing elements of her job that cause added stress, which for Osaka was speaking to sports reporters: “Naomi is not the first pro tennis player or athlete for that matter with mental health issues. Being a professional tennis player means facing the media. It’s part of the job”; “Athletes make ginormous money, but they have to give interviews. Just because she threw a fit, does not change that. She has to give interviews or get a new profession. I’ve watched her interviews and she’s smiling and laughing. So I don’t get it. Spoiled?!“; and “Weak and spoiled. Do you how many people feel the stress and anxiety when going for interviews? Mostly everyone. Do they also hide and decide to make a big deal out of it. NO. You get all that fame and play in front of crowds then claim you are socially anxious. Absurd!”

Results from Bot Analysis

RQ 2 aimed to examine the evidence for the presence of bots in the Twitter conversations promoting stigma-laden responses to Osaka’s tweet. In order to examine this, we used the botscores generated by the “botometer” API on all the negative tweets that were manually removed from LDA Topics 3 and 5. Results indicated that over 40% of the users’ tweeting negative comments were bot-like accounts (i.e., their botometer scores were above .75). Additionally, another 12% were mostly accounts that were suspended by Twitter which could potentially mean they could also have been bots. About 42% of the tweets had bot scores between .40, and .75; we categorized them as uncertain accounts. About 6% of the tweets were sent by non-bot accounts whose bot scores with less than .40.

Discussion

The results from the LDA Model as well as the thematic analysis indicate that Osaka’s public disclosure of her mental health struggles elicited an extremely positive response on Twitter. The positive reaction aligns with findings from previous sports destigmatization studies (e.g., Parrott et al., 2020). The three dominant themes of the positive tweet were affirmation, critiquing the structure and attributing the blame to the external environment, and destigmatization through visibility. The tweeters respectively supported and validated Osaka’s feelings and pain, and self-disclosed their own mental health problems while thanking Osaka for providing a space for them to do so. These results support the idea that fans have parasocial interactions with athletes (e.g., Sanderson & Kassing, 2016) and that sports figures’ mental health disclosures could have prosocial outcomes (Parrott et al., 2020). Interestingly, after announcing her withdrawal, Nike, one of Osaka’s major sponsors came out to support her decision and tweeted “Our thoughts are with Naomi. We support her and recognize her courage in sharing her own mental health experience” (Ziady, 2021). Such positive support from fans, and sponsors alike helps communicate to the general public a more normative environment wherein, laypeople might be encouraged to accept and disclose their own mental health issues. It may further help with reducing the stereotypes surrounding mental health issues that were highlighted by Corrigan and Bink (2016).

The largely positive tweets lend some support to the findings from Gwarjanski and Parrott’s (2018) findings, as Osaka’s self-disclosure tweet validating the experiences of people with mental health problems might have helped facilitate the positive responses. Thus, it appears Osaka’s tweet could be viewed as akin to a contact strategy way of destigmatization (Corrigan & Penn, 2015), which yielded further contact tweets and protest tweets.

The protest response can be seen in one of the main topics generated from the LDA results that contained tweets that criticized tennis leaders, French Open organizers, and sports journalists for creating a hostile environment for Osaka. These results were also particularly encouraging as one of the stereotypes surrounding mental health issues is the attribution of blame, with people suffering from mental health being blamed for their current mental health issues by both individuals (Corrigan & Bink, 2016) and news media (Gwarjanski & Parrott, 2018; Major & Jankowski, 2020). These tweets indicate that several social media users understand that Osaka was not responsible for her mental health issues, further reducing the stereotypes and subsequent stigma around mental health. Such acceptance provides further reassurance of reduction of public stigma and it could potentially aid in reducing self-stigma for other social media users who could be suffering mental health issues. The findings lend support to the idea that social media can be a place of resistance to dominant narratives (Locke et al., 2018) and contribute to the existing literature on athletes’ personal testimony as a destigmatization strategy.

Results also indicated a small number of negative tweets, wherein, Twitter commenters downplayed Osaka’s mental health concerns and emphasized that the areas giving Osaka the most anxiety were just part of her job. There is a similarity with Parrott et al. (2020) in that the negative commenters wanted Osaka to cope with her problems. Interestingly, there was also hostility toward Osaka from commenters who dismissed her mental health concerns by intimating that non-famous people have similar problems and continue to work. Commenters perceived that Osaka’s fame and financial success should blunt any mental health problems she has. And even though Osaka withdrew from the French Open, fans were still indigent, perhaps because they were denied watching one of tennis’s top performers. Kalman-Lamb’s (2019) analysis that athletes’ labor is the main thing valued by sports consumers helps explain commenters’ frustration with Osaka.

Results from the bot analysis indicated that 40% of the tweets belonging to the negative category in the thematic analysis were posted by the bots. The majority of these negative bot tweets targeted Osaka with mean and inflammatory comments about her being unprofessional towards the fans and sponsors, abusing her privilege asking for special treatment, and calling out her hypocrisy of continuing to do media campaigns but refusing to talk to the press, etc. Certain bot messages went to the personal level of character assassination as well. This predominant presence of bots promoting negative, harsh, and offensive comments on online discourse is consistent with previous research (Albadi et al., 2019; Uyheng & Carley, 2020; Yan et al., 2021) wherein bots have relentlessly engaged in spewing vitriol and hate speech in the online networks. Again, congruent with prior research where bots facilitated and amplified negative emotions in cyberspace (Shi et al., 2020; Yan et al., 2021), it can be argued that exposure to disruptive bot content regarding mental health disclosures could distort the online discourse and adversely impact the primary goal of mental health advocacy. While it is difficult to discern if the bots are designed with the intent to stigmatize, the frequency at which bot content is disseminated, and the fact that low credibility content posted by social bots is likely to be shared and retweeted (Shao et al. 2018), this can possibly steer conversations to create a climate of stigmatizing collective opinion. While bots can amplify specific voices in online health-related discussions, this can potentially drown out and stifle relevant public voices and conversations that support the cause of mental health, which can have stigmatizing ramifications. Yan et al. (2021) concluded their study with the following: “The arrival of bots in the digital sphere of athlete protest is not to be treated as an innocuous phenomenon, but one that engages in the broader, ongoing social struggles of our time” (p. 106). The authors of this paper agree with this sentiment, especially considering Osaka’s refusal to play can be seen as a rebuke of the existing tennis structure.

Theoretical and Practical Implications

By combining insights from unsupervised topic modeling and bot detection, the current study has several theoretical and practical implications. From a theoretical perspective, the study adds to the growing body of literature on the positive effects of parasocial interactions on stigma reduction. The use of social media by elite athletes to disclose stigmatized health issues and directly talk with their fans and audiences allows for parasocial interaction strategy. It is particularly encouraging as previous case studies from celebrity disclosure have yielded positive results for help-seeking behaviors from the general public (e.g., “The Diana Effect”; Calhoun & Gold, 2020). The findings of the study also add to the growing literature on mental health and sports by highlighting the fact that social media users were not adhering to the stereotypes around mental health by placing the blame on Osaka and adopting a more protest-based strategy to help with destigmatization of mental health. These results also have practical implications in the discourse of stigma reduction. As indicated earlier, mental health stigma has deleterious effects on people suffering from mental health issues (WHO, 2019). The positive and supportive comments from Osaka’s fans and other social media users can serve as an indication that celebrities’ disclosures can act as a way to challenge mental health issues. This might be particularly useful for young athletes to care for their mental health.

Additionally, while prior research on the presence of social media bots has primarily focused on contentious issues (Himelein-Wachowiak et al., 2021) like climate change (Marlow et al., 2021), vaccine debate (Broniatowski et al., 2018), the present study theoretically widens the scope of literature by evidencing the presence of bots in propagating stigmatizing content on mental health and sport. On a practical level, as mentioned above, in the mental health context, automated accounts can sway public opinion by increasing social media users’ exposure to polarizing and negative inflammatory messages about mental health disclosures. Such exposure can hamper the destigmatization efforts; adversely impacting athletes’ willingness to disclose, and seek help and treatment for their mental health concerns. Such infiltration of social bots also has the potential of introducing stigmatizing mental health narratives to online conversations, thus defeating the purpose of celebrity disclosures that help with the destigmatization of mental health. Furthermore, given that prior research suggests that athletes face online abuse (Kavanagh Jones, & Sheppard-Marks, 2016), and social bots are deepening the trend of creating highly clustered echo chambers (Yuan et al., 2019), athletes need to be more vigilant about the presence of such social bots interacting with their social media posts and tarnishing their image. As such, athletes’ communication teams should devise strategies to detect, report, and flag bot content and activity regularly.

Limitations and Future Research

The present study is not without limitations. First, we only analyzed English language tweets, which limits the generalizability of the results. Given that Naomi Osaka is a Japanese citizen, several fans could have posted in Japanese, which the current study didn’t include. The present study relied on computational methods, and as with any research method, these methods are not without limitations. First, the LDA uses a bag of words approach to organizing topics based on frequently co-occurring words. Computers might not always recognize the finer nuances of human language. Although the present study used a large sample (20% of the total tweets) of randomly selected samples to perform a fine-grain thematic analysis, there is some probability that certain tweets might not have belonged 100% to the theme or that they could have two different themes embedded within them. Finally, the Botometer API is not a perfect tool in itself given the ever-changing nature of social media platforms. The API is trained on a training dataset, to detect similar automated behavioral patterns. Given that bots are becoming more sophisticated, the model might misclassify some accounts. The present study used a stringent measurement to classify accounts; however, false positive or false negative scores are still possible.

Future research needs to examine whether there are gender effects based on the gender identities of both the athlete and the social media commenters aid in reducing stigma. Additionally, future research should also examine the impact of professional athletes’ disclosures on young elite athletes.

In conclusion, recognizing that athletes are regular people helps dispel heroic myths about them (Segrave, 2000). Because of athletes’ parasocial interactions with the public, recognizing mental health struggles could help destigmatize mental illness (Parrott et al., 2020). This study examined user responses to a top-earning female athlete’s withdrawal citing mental health concerns. As mental health issues are a growing concern globally, the positive responses garnered by Osaka’s tweet announcing her withdrawal offer hope for reducing the stigma around mental health issues

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Sushma Kumble

Steve Bien-Aimé

Pratiti Diddi

References

Albadi

Kurdi

Mishra

(2019). Hateful people or hateful bots?: Detection and characterization of bots spreading religious hatred in Arabic social media. Proceedings of the ACM on Human-Computer Interaction, 3(CSCW), 1–25. https://doi.org/10.1145/3359163

Arnold

(2021, May 31). Naomi Osaka drops out of French open after dispute over media appearances. NPR. https://www.npr.org/2021/05/31/1001917952/naomi-osaka-drops-out-of-french-open-after-dispute-over-media-appearances

Blei

D. M.

A. Y.

Jordan

M. I.

(2003). Latent dirichlet allocation. Journal of Machine Learning Research, 3(1), 993–1022. Retrieved from https://www.jmlr.org/papers/volume3/blei03a/blei03a.pdf?ref=https://githubhelp.com

Broniatowski

D. A.

Jamison

A. M.

AlKulaib

Chen

Benton

Quinn

S. C.

Dredze

(2018). Weaponized health communication: Twitter bots and Russian trolls amplify the vaccine debate. American Journal of Public Health, 108(10), 1378–1384. https://doi.org/10.2105/ajph.2018.304567

Calhoun

A. J.

Gold

J. A.

(2020). ‘I feel like I know them’: The positive effect of celebrity self-disclosure of mental illness. Academic Psychiatry, 44(2), 237–241. https://doi.org/10.1007/s40596-020-01200-5

Cassilo

(2022). Royce white, DeMar DeRozan and media framing of mental health in the NBA. Communication & Sport, 10(1), 97–123. https://doi.org/10.1177/2167479520933548

Coakley

(2021). Sports in society: Issues and controversies (13th Edition). McGraw-Hill.

Corrigan

Bink

A. B.

(2016). The stigma of mental illness. In Friedman

H. S.

(Ed), Encyclopedia of mental health (2nd ed., pp. 230–234). Academic Press.

Corrigan

P. W.

Penn

D. L.

(2015). Lessons from social psychology on discrediting psychiatric stigma. Stigma and Health, 1(S), 20. https://doi.org/10.1037/2376-6972.1.S.2

10.

Corrigan

P. W.

Shapiro

J. R.

(2010). Measuring the impact of programs that challenge the public stigma of mental illness. Clinical Psychology Review, 30(8), 907–922. https://doi.org/10.1016/j.cpr.2010.06.004

11.

Corrigan

P. W.

Watson

A. C.

(2002). The paradox of self-stigma and mental illness. Clinical Psychology: Science and Practice, 9(1), 35–53. https://doi.org/10.1093/clipsy.9.1.35

12.

Davis

C. A.

Varol

Ferrara

Flammini

Menczer

(2016). BotOrNot: A system to evaluate social bots. In Proceedings of the 25th international conference companion on world wide web, Montréal, Canada, 11–15, May, 2016, WWW ’16 Companion, 273–274 https://doi.org/10.1145/2872518.2889302

13.

Ferrara

Varol

Davis

Menczer

Flammini

(2016). The rise of social bots. Communications of the ACM, 59(7), 96–104. https://doi.org/10.1145/2818717

14.

Goffman

(1963). Stigma; notes on the management of spoiled identity. Prentice-Hall.

15.

Grimme

Preuss

Adam

Trautmann

(2017). Social bots: Human-like by means of human control? Big Data, 5(4), 279–293. https://doi.org/10.1089/big.2017.0044

16.

Guo

Vargo

C. J.

Pan

Ding

Ishwar

(2016). Big social data analytics in journalism and mass communication: Comparing dictionary-based text analysis and unsupervised topic modeling. Journalism & Mass Communication Quarterly, 93(2), 332–359. https://doi.org/10.1177/1077699016639231

17.

Gwarjanski

A. R.

Parrott

(2018). Schizophrenia in the news: The role of news frames in shaping online reader dialogue about mental illness. Health Communication, 33(8), 954–961. https://doi.org/10.1080/10410236.2017.1323320

18.

Himelein-Wachowiak

Giorgi

Devoto

Rahman

Ungar

Schwartz

H. A.

Epstein

D. H.

Leggio

Curtis

(2021). Bots and misinformation spread on social media: A mixed scoping review with implications for COVID-19. Journal of Medical Internet Research, 23(5). https://doi.org/10.2196/26933

19.

Howard

P. N.

Kollanyi

Bradshaw

Neudert

L. M.

(2018). Social media, news and political information during the US election: Was polarizing content concentrated in swing states? arXiv preprint arXiv:1802.03573 .

20.

Jones

E. E.

Farina

Hastorf

A. H.

Markus

Mille

D. T.

Scott

R. A.

(1984). Social stigma: The psychology of marked relationships. W. H. Freeman & Co.

21.

Kalman-Lamb

(2019). Athletic labor and social reproduction. Journal of Sport and Social Issues, 43(6), 515–530. https://doi.org/10.1177/0193723519850879

22.

Kavanagh

Jones

Sheppard-Marks

(2016). Towards typologies of virtual maltreatment: Sport, digital cultures & dark leisure. Leisure Studies, 35(6), 783–796. https://doi.org/10.1080/02614367.2016.1216581

23.

Kollanyi

(2016). Where do bots come from?: An analysis of bot codes shared on GitHub. International Journal of Communication, 10, 4932–4951. https://doi.org/10.1932-8036/20160005

24.

Kroshus

Hoopes

Bernstein

Chrisman

S. P.

Rivara

F. P.

(2019). Direct verbal communication by parents, pressure related to sport achievement, and concussion safety in youth football. Journal of Health Communication, 24(10), 770–779. https://doi.org/10.1080/10810730.2019.1668512

25.

Kwon

Park

(2022). Understanding user responses to the COVID-19 pandemic on Twitter from a terror management theory perspective: Cultural differences among the US, UK and India. Computers in Human Behavior, 128, 107087. https://doi.org/10.1016/j.chb.2021.107087

26.

Lindlof

T. R.

Taylor

B. C.

(2011). Qualitative communication research methods. SAGE Publications.

27.

Link

B. G.

Phelan

J. C.

(2001). Conceptualizing stigma. Annual Review of Sociology, 27(1), 363–385. https://doi.org/10.1146/annurev.soc.27.1.363

28.

Locke

Lawthom

Lyons

(2018). Social media platforms as complex and contradictory spaces for feminisms: Visibility, opportunity, power, resistance and activism. Feminism & Psychology, 28(1), 3–10. https://doi.org/10.1177/0959353517753973

29.

Loper

Bird

(2002). NLTK: The natural language toolkit. ACL workshop on effective tools and methodologies for teaching natural language processing and computational linguistics, Philadelphia, PA.

30.

Lydecker

J. A.

Cotter

E. W.

Palmberg

A. A.

Simpson

Kwitowski

White

Mazzeo

S. E.

(2016). Does this tweet make me look fat?: A content analysis of weight stigma on twitter. Eating and Weight Disorders — Studies on Anorexia, Bulimia and Obesity, 21, 229–235. https://doi.org/10.1007/s40519-016-0272-x

31.

Maier

Waldherr

Miltner

Wiedemann

Niekler

Keinert

… Adam

(2018). Applying LDA topic modeling in communication research: Toward a valid and reliable methodology. Communication Methods and Measures, 12(2-3), 93–118. https://doi.org/10.1080/19312458.2018.1430754

32.

Major

L. H.

Jankowski

S. M.

(2020). Thematic and episodic frames in depression news. In: Health news and responsibility: How frames create blame. Peter Lang.

33.

Marlow

Miller

Roberts

J. T.

(2021). Bots and online climate discourses: Twitter discourse on President Trump’s announcement of U.S. withdrawal from the Paris Agreement. Climate Policy, 21(6), 765–777. https://doi.org/10.1080/14693062.2020.1870098

34.

McCracken

(1988). The long interview. SAGE Publications. https://doi.org/10.4135/9781412986229

35.

Osaka

[@naomiosaka]. (2021, May 31). Hey everyone [Tweet]. Twitter. https://twitter.com/naomiosaka/status/1399422304854188037/

36.

Parrott

Billings

A. C.

Buzzelli

Towery

(2021). ‘We all go through it’: Media depictions of mental illness disclosures from star athletes DeMar DeRozan and Kevin Love. Communication & Sport, 9(1), 33–54. https://doi.org/10.1177/2167479519852605

37.

Parrott

Billings

A. C.

Hakim

S. D.

Gentile

(2020). From #endthestigma to #realman: Stigma-challenging social media responses to NBA players’ mental health disclosures. Communication Reports, 33(3), 148–160. https://doi.org/10.1080/08934215.2020.1811365

38.

Pavelko

R. L.

Wang

(2021). Love and basketball: Audience response to a professional athlete’s mental health proclamation. Health Education Journal, 80(6), 635–647. https://doi.org/10.1177/00178969211006161

39.

Pedersen

P. M.

Laucella

P. C.

Kian

E. M.

Geurin

A. N.

(2020). Strategic sport communication (3rd Edition). Human Kinetics Inc.

40.

PricewaterhouseCoopers (2021). PwC 2021 sports outlook. Retrieved from https://www.pwc.com/us/en/industries/tmt/assets/pwc-2021-tmt-sports-outlook.pdf

41.

Reavley

N. J.

Pilkington

P. D.

(2014). Use of Twitter to monitor attitudes toward depression and schizophrenia: An exploratory study. PeerJ, 2, e647. https://doi.org/10.7717/peerj.647

42.

Rehurek

Sojka

(2011). Gensim—statistical semantics in python. Org. Retrieved from genism .

43.

Roder

Both

Hinneburg

(2015). Exploring the space of topic coherence measures. Proceedings of the eighth ACM international conference on Web search and data mining (pp. 399–408). https://doi.org/10.1145/2684822.2685324

44.

Sanderson

Kassing

J. W.

(2016). New media and the evolution of fan-athlete interaction. In Billings

A. C.

Hardin

(Eds), Routledge Handbook of sport and new media (pp. 247–258). Routledge.

45.

Scanlon

T. F.

(2006). Sports and media in the ancient Mediterranean. In Raney

A. A.

Bryant

(Eds), Handbook of sports and media (pp. 3-19). Lawrence Erlbaum Associates.

46.

Segrave

J. O.

(2000). Sport as escape. Journal of Sport and Social Issues, 24(1), 61–77. https://doi.org/10.1177/0193723500241005

47.

Shao

Ciampaglia

G. L.

Varol

Yang

K. C.

Flammini

Menczer

(2018). The spread of low-credibility content by social bots. Nature Communications, 9(1). https://doi.org/10.1038/s41467-018-06930-7

48.

Shi

Liu

Yang

Zhang

Wen

(2020). Social bots’ sentiment engagement in health emergencies: A topic-based analysis of the COVID-19 pandemic discussions on Twitter. International Journal of Environmental Research and Public Health, 17(22), 8701. https://doi.org/10.3390/ijerph17228701

49.

Sievert

Shirley

(2014, June). LDAvis: A method for visualizing and interpreting topics. Proceedings of the workshop on interactive language learning (pp. 63-70): Visualization, and Interfaces. https://doi.org/10.3115/v1/W14-3110

50.

Smith

R. A.

(2007). Language of the lost: An explication of stigma communication. Communication Theory, 17(4), 462–485. https://doi.org/10.1111/j.1468-2885.2007.00307.x

51.

Thompson

Carter

Lee

Alshamrani

Billings

A. C.

(2022). ‘We’re human too’:Media coverage of Simone Biles’ mental health disclosure during the 2020 Tokyo Olympics. Electronic News, 16(3), 187–201.

52.

Uyheng

Carley

K. M.

(2020). Bots and online hate during the COVID-19 pandemic: Case studies in the United States and the Philippines. Journal of Computational Social Science, 3(2), 445-468. https://doi.org/10.1007/s42001-020-00087-4

53.

Varol

Ferrara

Davis

Menczer

Flammini

(2017). Online human-bot interactions: Detection, estimation, and characterization. Proceedings of the International AAAI Conference on Web and Social Media, 11(1), 280–289. https://doi.org/10.48550/arXiv.1703.03107

54.

Wann

D. L.

Allen

Rochelle

A. R.

(2004). Using sport fandom as an escape: Searching for relief from under-stimulation and over-stimulation. International Sports Journal, 8(1), 104–113.

55.

Watson

A. C.

Corrigan

Larson

J. E.

Sells

(2007). Self-stigma in people with mental illness. Schizophrenia Bulletin, 33(6), 1312–1318. https://doi.org/10.1093/schbul/sbl076

56.

Weiner

(1995). Judgments of responsibility: A foundation for a theory of social conduct. Guilford Press.

57.

Wojcik

Messing

Smith

Rainie

Hitlin

(2018, April 9). Bots in the twittersphere. Retrieved March 25, 2022, from https://www.pewresearch.org/internet/2018/04/09/bots-in-the-twittersphere/

58.

World Health Organization (2019). The WHO special initiative for mental health (2019-2023): Universal health coverage for mental health. World Health Organization. https://apps.who.int/iris/handle/10665/310981

59.

Yan

Pegoraro

Watanabe

N. M.

(2021). Examining IRA bots in the NFL anthem protest: Political agendas and practices of digital gatekeeping. Communication & Sport, 9(1), 88–109. https://doi.org/10.1177/2167479519849114

60.

Yang

K. C.

Ferrara

Menczer

(2022). Botometer 101: Social bot practicum for computational social scientists. arXiv preprint arXiv:2201.01608 .

61.

Yoo

J. J.

Min

Koh

(2024). Cross-national news narratives of the Paralympic Games: Computational text analysis of the media coverage in the United States and South Korea. Communication & Sport, 12(2), 230–253. https://doi.org/10.1177/21674795221090420

62.

Yuan

Schuchard

R. J.

Crooks

A. T.

(2019). Examining emergent communities and social bots within the polarized online vaccination debate in Twitter. Social Media + Society, 5(3), 2056305119865465. https://doi.org/10.1177/20563051198654

63.

Ziady

, CNN Business (2021, June 3). Naomi Osaka: Nike backs tennis star after she withdraws from French open. CNN. https://edition.cnn.com/2021/06/01/business/naomi-osaka-nike-sponsors/index.html