Investigating Young Employee Stressors in Contemporary Society Based on User-Generated Contents

Understanding stressors is an effective measure to decrease employee stress and improve employee mental health. The extant literature mainly focuses on a singular stressor among various aspects of their work or life. In addition, the extant literature generally uses questionnaires or interviews to obtain data. Data obtained in such ways are often subjective and lack authenticity. We propose a novel machine–human hybrid approach to conduct qualitative content analysis of user-generated online content to explore the stressors of young employees in contemporary society. The user-generated online contents were collected from a famous Q&A platform in China and we adopted natural language processing and deep learning technology to discover knowledge. Our results identified three kinds of new stressors, that is, affection from leaders, affection from the social circle, and the gap between dream and reality. These new identified stressors were due to the lack of social security and regulation, frequent occurrences of social media fearmongering, and subjective cognitive bias, respectively. In light of our findings, we offer valuable practical insights and policy recommendations to relieve stress and improve mental health of young employees. The primary contributions of our work are two-fold, as follows. First, we propose a novel approach to explore the stressors of young employees in contemporary society, which is applicable not only in China, but also in other countries and regions. Second, we expand the scope of job demands-resources (JD-R) theory, which is an important framework for the classification of employee stressors.


Introduction
Stress is an integral part of employees' lives and occurs in a wide variety of work environments [1]. Employee stress is frequently defined as a work or personal related feeling of difficulty, frustration, depression, or tension. It is deemed as a harmful part of the work environment [2]. First, it can seriously undermine employee well-being, thereby provoking health-related impairments globally [3]. Second, employee stress may affect an employee's family or personal life [4]. Third, fueled by adverse psychosocial working conditions, employee stress increases employee absence [5]. According to the World Health Organization (WHO), more than 350 million working people worldwide suffer from depression, with an increase in about 18% in the past decade. The prevalence rate is higher among young people [6].
A good understanding of employee stressors is an effective measure to decrease employee stress. Therefore, a significant number of research on employee stressors across various occupations have been developed over the last forty years. According to [7], there are mainly two types of factors leading to employee stress: work-related stressors and family life-related stressors.
Work-related stressors can be divided into direct stressors and indirect stressors. In terms of direct stressors, performance stress is the top stressor for employees [8]. Performance stress is the urgency to improve performance in order to achieve desirable to the job demands that lead to constant overtaxing and, in the end, to stress and exhaustion. Resource process refers to the lack of resources complicating the meeting of job demands, which also further leads to stress and withdrawal behavior [37]. With the JD-R model framework as shown in Figure 1, any stressor for employees can be categorized [38]. Another research gap is that the extant literature generally uses questionnaires or interviews to obtain data. These research methods are conventional and have the following two advantages: (1) they save time, money and manpower; and (2) the results are easier to quantify. However, the problem with such measures is that the same test object provides all the information and therefore the statistical relationship between the structures may be inflated due to common source bias [32]. In addition, the data obtained through questionnaires and interviews are subjective and lack authenticity, which may make the study results unrepresentative. For this reason, we systematically investigated the multiple stressors based on user-generated contents (UGC). UGC has emerged as a promising source of data to ascertain the actual thoughts and opinions of individuals [39]. UGC is easily accessible, with detailed information that can be processed efficiently and cost-effectively using advanced natural language processing (NLP) and machine learning technologies. As existing technologies still cannot fully grasp human language, a hybrid approach that combines the machine and human was applied in our research. Machine learning can efficiently extract keywords and sentences while human interventions aggregate their meanings.
In this paper, we introduced a machine-human hybrid approach to conduct qualitative content analysis of UGC to obtain comprehensive stressors. We divided the obtained stressors into mainly four categories: working process, family life, social media, and subjective cognitive bias. Finally, we give feasible suggestions to alleviate stressors.
There are two contributions in this paper. First, we systematically studied the stressors for young employees in contemporary society and identified some new stressors. Second, to obtain comprehensive and objective data, we developed a machine-human hybrid approach to gain insights into stressors for Chinese young employees. Another research gap is that the extant literature generally uses questionnaires or interviews to obtain data. These research methods are conventional and have the following two advantages: (1) they save time, money and manpower; and (2) the results are easier to quantify. However, the problem with such measures is that the same test object provides all the information and therefore the statistical relationship between the structures may be inflated due to common source bias [32]. In addition, the data obtained through questionnaires and interviews are subjective and lack authenticity, which may make the study results unrepresentative. For this reason, we systematically investigated the multiple stressors based on user-generated contents (UGC). UGC has emerged as a promising source of data to ascertain the actual thoughts and opinions of individuals [39]. UGC is easily accessible, with detailed information that can be processed efficiently and cost-effectively using advanced natural language processing (NLP) and machine learning technologies. As existing technologies still cannot fully grasp human language, a hybrid approach that combines the machine and human was applied in our research. Machine learning can efficiently extract keywords and sentences while human interventions aggregate their meanings.
In this paper, we introduced a machine-human hybrid approach to conduct qualitative content analysis of UGC to obtain comprehensive stressors. We divided the obtained stressors into mainly four categories: working process, family life, social media, and subjective cognitive bias. Finally, we give feasible suggestions to alleviate stressors.
There are two contributions in this paper. First, we systematically studied the stressors for young employees in contemporary society and identified some new stressors. Second, to obtain comprehensive and objective data, we developed a machine-human hybrid approach to gain insights into stressors for Chinese young employees.
The rest of this paper is organized as follows. Section 2 describes the study design and the methods applied for collecting data. Section 3 presents the empirical results derived from the UGC analysis. Section 4 discusses the implications for management practitioners. Section 5 concludes the paper with limitations and suggests future research directions.

Study Design
We developed a machine-human hybrid approach to conduct qualitative content analysis for the identification of stressors for Chinese young employees from UGC. The proposed approach can be divided into five stages (see Figure 2). First, we collected the raw UGC from a Chinese Q&A site. Second, we identified the informative sentences that can directly represent the stressors for Chinese young employees from the raw UGC. Third, using a language model named word2vec, we converted text into word embeddings that can be calculated to train word embeddings. Fourth, we clustered word embeddings and randomly sample words from each cluster in order to find the stressor-related keywords. Final, we extracted the original sentences of the keywords from the UGC, manually reviewed the chosen sentences, and analyzed the details of the stressors.

Stage 1: Data Collection
We collected the raw UGC from "Zhihu", the largest Chinese social question and answering platform. Zhihu, which started their business in December 2010, provides longterm detailed answers and discussions raised by the users. As of December 2020, the total number of questions and answers on Zhihu had exceeded 44 million and 240 million, respectively. Users of Zhihu can openly and freely express their opinions on various topics. In addition, Zhihu's data are available to the public and can be used for opinion mining [40,41].

Stage 1: Data Collection
We collected the raw UGC from "Zhihu", the largest Chinese social question and answering platform. Zhihu, which started their business in December 2010, provides long-term detailed answers and discussions raised by the users. As of December 2020, the total number of questions and answers on Zhihu had exceeded 44 million and 240 million, respectively. Users of Zhihu can openly and freely express their opinions on various topics. In addition, Zhihu's data are available to the public and can be used for opinion mining [40,41].
According to the literature, we choose several keywords that could directly point out that young employees were under stress. We searched the questions containing these keywords on Zhihu, and crawled responses to the three questions: "What makes young people feel tired?", "Why do so many employees work overtime in China? and what are they doing?", and "Why do young people in China have no desires?". Here, "burn out", "work overtime", and "desirelessness" can all infer people are under stress [17,37,42]. There are many responses to the three questions, which contain rich detailed information about the things young employees have suffered, and the stress on young employees from the things suffered. There were 3336 responses (a total of 25,044 sentences) to the first question, 2211 responses (a total of 16,589 sentences) to the second question, and 7073 responses (a total of 49,955 sentences) to the third question.

Stage 2: Identify Informative Content
UGC contains a substantial amount of sentences that do not relate to the stressors. To improve the accuracy of subsequent machine-learning, we identified the informative sentences that can directly express stressors for employees. We treated the identification process as a classification task: the classifier will identify the informative and noninformative sentences, and label them as 1 and 0, respectively. The Convolutional Neural Network (CNN) model was chosen to train the classifier. As an important basic model in deep learning, CNN can automatically extract key features and has a short training time to avoid the tedious process of manual feature extraction [43,44]. In addition, large-scale network implementation is much easier with CNN than with other neural networks [44]. The process of classification can be divided into three steps. In step 1, we manually labeled a small set of sentences as informative/noninformative sentences to build a set of training data. We randomly selected 1000 sentences from the answers of each question, with a total of 3000 sentences. We manually judged whether each sentence represented the stressors for young employees: if a sentence represents the stressors for young employees, it is labeled as 1; otherwise it will be labeled as 0. In step 2, we trained the text classifier. Since texts cannot be directly calculated, we mapped the raw UGC data onto sentence embeddings using Bidirectional Encoder Representation from Transformers (BERT) model. With the architecture of multi-layer bidirectional transformer encoder, BERT is empirically powerful in many natural language processing (NLP) tasks. Unlike other traditional language models such as RNN, it is an unsupervised model and does not require manual intervention [45]. The sentence embeddings are input in the CNN model to train the classifier. In step 3, the trained classifier automatically identifies informative sentences from the UGC.

Stage 3: Preprocess UGC and Train Word Embeddings
As mentioned earlier, texts cannot be directly calculated, so we needed to preprocess the informative sentences and map the words onto word embeddings. The process of preprocessing can be divided into two steps. In step 1, we segmented the Chinese sentences. Since the sentences were written in Chinese, it was necessary to split the sentences into words. We used "Jieba", a Python software for Chinese text segmentation, to segment the Chinese sentences into a corpus. It is worth noting that, as mentioned earlier, we collected responses from three questions. We segmented the responses to the same question into one corpus. In this way, we constructed three corpora. In step 2, we removed the stop words. In the corpus, there were many stop words (e.g., "but" and "and") that would significantly affect the efficiency of NLP and need to be deleted. Table 1 illustrates the comparison between the raw and processed data. Table 1. Examples of the raw and processed data.

Raw Data Processed Data
As a freshman, I have no experience at all, and every time I fail to do well, I will be considered incapable. Freshman, experience, every time, fail, to do well, incapable We no longer discuss houses, just the current housing price. This is a desperate, meaningless sad topic.
No longer, discuss, house, current housing price, desperate, meaningless, sad topic Under continuous high-intensity and high-pressure work, the nerves are already very fragile, and the body is also suffering from major and minor problems.
Continuous, high-intensity, high-pressure work, nerves, fragile, body, suffering, major, minor, problems We mapped the preprocessed words onto numerical word embeddings using the skip-gram model. The skip-gram model is a kind of language model, which is based on the artificial neural network. It can process large-scale textual data and capture the contextual information of the text [46]. We implemented the skip-gram model through the Word2Vec package in Python. The training results were saved as word embeddings for use during the subsequent stages.

Stage 4: Cluster Key Words
To efficiently review the informative sentences, we clustered the words in the corpus into several groups based on word embeddings. We applied the "k-means" algorithm to cluster the keywords. This algorithm was implemented with the following steps. First, several words were randomly designated as the cluster centers of each group. Second, the distance between words and each cluster center were measured. Third, the words were classified into the nearest cluster. Fourth, each cluster center was recalculated. The iterations of the second and fourth steps continued until there was no further change in each cluster center. We clustered the words into 30 groups for each question, with a total of 90 groups. The words in the same group usually had similar semantics. The clustering results were saved for use in the next stage.

Stage 5: Manually Extract the Details of Stressors for Young Employees
Since the existing NLP technology is not able to really understand the semantics of the text, we have to manually analyze the meaning of the sentence. We randomly selected four sample words from each group, and extracted the sentences that contained the sample words from the UGC. We manually reviewed the extracted sentences and identified the details of stressors for young employees. Two professors and two graduate students participated in this operation. They conducted an evaluation based on the four eyes principle, which can effectively reduce serious error in judgment [47]. Initially, they reviewed the sentences separately, and then verified the consistency. If no agreement was reached, they continued their discussion until agreement was reached.

Results
After reviewing the chosen sentences, we identified four main categories of stressors for Chinese young employees mentioned frequently by users of "Zhihu": working process, family life, social media, and subjective cognitive bias. Figure 3 illustrates a holistic view with detailed information emerging from the data.  The dark blue central node represents employee stress. The light blue nodes linked with the central dark node represent four categories. The keywords included in the four categories are listed nearby. The number in parentheses express the counts of responses that mention the keywords. Table 2 show the detailed information on each category, its keywords, and related original example sentences.  The dark blue central node represents employee stress. The light blue nodes linked with the central dark node represent four categories. The keywords included in the four categories are listed nearby. The number in parentheses express the counts of responses that mention the keywords. Table 2 show the detailed information on each category, its keywords, and related original example sentences.

Working Process
The first was working long hours. There were four main keywords related to working long hours: "working overtime", which can stress the body as well as the mind; "health" and "rest", where there is talk that working long hours will cause health problems and reduce the amount of rest time; and "stay up late", where people mentioned that they felt a lack of entertainment due to their working long hours, so had fun at night, sacrificing their rest time. The second subcategory was dead-end job. There were four main identified keywords related to dead-end jobs: "hope", which is at the top of the keywords mined from the answers, where many people mentioned that even if they worked very hard, there was little hope in their jobs; "graduation", where new employees complained that they were always confronted with job stress; "COVID-19 epidemic" and "competition", where due to the outbreak of COVID-19, many people felt higher competition pressure. The third subcategory was a rough time during work. There were mainly four keywords related to this subcategory: "supervisor" and "push", where many people complained that they were often pushed to work overtime by their supervisors; "reward", where many people complained that they were not rewarded enough for what they produced; and "resign", where many people said that they dare not resign even if the job was painful. The fourth subcategory was the social life at the workplace. There were two main keywords related to this subcategory: "colleagues" and "boss", where many people said that they did not get along with their colleagues and bosses. The fifth subcategory was a dislike of their jobs. There were two main keywords related to this subcategory: "dream" and "love", where many people said that they did not love their jobs and they chose to suppress their inner desires and dream. The sixth subcategory was the commute to work. There were three main keywords related to this subcategory: "stop", "bus", and "commute", where many people complained that they struggle through the long commutes. The seventh subcategory was a lack of social security. There were two main keywords related to this subcategory: "squeeze" and "exploit", where some people said that they faced exploitation and pressure caused by their jobs due to the lack of social security.

Family Life
Keywords emerged for family life were categorized under nine subcategories, see Table 3. It's hard to feed myself, let alone my family.
No one will give me living expenses in vain to feed my whole family.

Marriage 39
In my opinion, marriage is a waste of money.
Maintaining marriage needs money.

Brother 24
My parents and brother want to squeeze me, although their income is more than mine.
I have to feed my brother and my family.

High housing costs
Buying houses 675 I cannot afford to buy a house by myself in Guangzhou, where the housing prices are the lowest in the first-tier cities.
I strive to save money to buy a house.
It is impossible for me to buy a house and a car with such a low salary.
Can't afford 309 I can't afford such expensive houses.
I can't afford a house and no one wants to marry me.

House prices 234
Take Jinan's house prices as an example, it has increased ten-fold in 20 years.
The housing prices are desperate.
Due to the high house prices, I can just rent a shabby house rather than buying a house.

High cost to get married
Get married 330 I am so poor that I can't get married or feed a child.
In my opinion, getting married will cost too much money.

Betrothal gifts 147
The expensive betrothal gifts make us give up to get married.
Betrothal gifts are too expensive.

Daily consumption
Car loans 138 I am overwhelmed by mortgage and car loans.
It is difficult for us to pay the mortgage and the car loans, and I even want to sell my car.

Consumption 104
With the soaring prices, it is difficult for families of ordinary consumption level to bear.
Consumption levels have exceeded the income levels of ordinary families.

Education cost
Kindergarten 96 In my small city, a kindergarten costs 3800 in a quarter.
My children have to take part in the "specialty training class" since they are in kindergarten.

Kids 93
I need to earn money to support my family and make my kids live a better life.
We must take care of our parents, wives, and kids.

Tuition fees 35
The mortgage, milk powder money, and tuition fees are all overwhelming me. The first was low salary. There were two main keywords related to this subcategory: "salary" and "income", where many people complained about the low salary. The second subcategory was high cost of family life. There were three main keywords related to this subcategory: "feed", "marriage", and "brother", where many people said that they had to spend a considerable sum of money supporting their families. The third subcategory was high housing costs. There were three main keywords related to this subcategory: "buying houses", "can't afford", and "house prices", where a great number of people complained that the house prices were too high. Housing price was mentioned repeatedly in 5357 responses. The fourth subcategory was the high cost of getting married. There were two main keywords related to this subcategory: "get married" and "betrothal gifts", where many people complained that getting married in China will cost too much. The fifth subcategory was daily consumption. There were two main keywords relating to this subcategory: "car loans" and "consumption", where many people complained that daily consumption was too high to afford. The sixth subcategory was educational cost. There were three main keywords related to this subcategory: "kindergarten", "kids", and "tuition fees", where many people complained that the educational cost was also too high to afford. The seventh subcategory was supporting aging parents. There were two main keywords related to this subcategory: "supporting aging parents" and "the elderly", where many people mentioned that they had to spend a lot of money to support their aging parents. The eighth subcategory was health care expenses. There were two main keywords related to this subcategory: "health care" and" doctors", where many people said that drugs, treatments, and seeing a doctor were very expensive, and they even dreaded being ill. The ninth subcategory was social pressure. There were two main keywords related to this subcategory: "parents", where many people complained that they were under parental pressure to get married to anyone and "social situation", where some people said that they had significant anxiety and depression due to poor ability to communicate in social situations.

Social Media
Keywords that emerged for social media can be categorized under three subcategories, see Table 4. The first is the consumption of time and energy caused by social media. There were three main keywords related to this subcategory: "mobile phone", where many people said that they spent a lot of time every day with their mobile phones unconsciously; and "video" and "Tik Tok", where many people said that they watched videos on social media every night before bedtime, even when they came home late. The second subcategory was desire for materials aroused by social media. There were two main keywords related to this subcategory: "actual life", where many people said that they dreamt of the high-end lifestyle spreading on social media, but were unable to make it happen in actual life, which would drag them into a deep depression; and "friend circle", where many people said that they envied the lives of friends seen on social media. The third subcategory was lives occupied by busy working. There were three main keywords related to this subcategory: "punch a time clock", "WeChat", and "message", where many people complained that since social media made it easier to contact them after work hours, their lives were further occupied by busy work.

Subjective Cognitive Bias
Keywords emerged for subjective cognitive bias were categorized under three subcategories, see Table 5. Even my smallest wishes still cannot be satisfied.

Consumerism 62
We are influenced by consumerism and a money-first culture.
Some people accept the consumerist and take out a loan.

Sub-Category Keywords Count Examples of Sentences
The gap between rich and poor The gap between rich and poor 71 I know that no matter how hard I work, I cannot get my ideal life given the gap between the rich and poor.
Since the gap between the rich and poor has gradually widened, we are dissatisfied with our current situation.
Poor 62 Most young people who are born in poverty cannot change their fates even if they do their best.
Poor kids are still poor when they grow up. The first subcategory was insatiable desire. There were three main keywords related to this subcategory: "desire", "satisfy", and "consumerism", where many people said that they were passionate about the products out of their affordability. If their desires are not satisfied, they will feel frustrated. The second subcategory was the gap between the rich and poor. There were two main keywords related to this subcategory: "the gap between the rich and poor" and "the poor", where many people said that through the information spreading on social media, they realize that there exists a growing gap that separates them from the richest people, and they are increasingly dissatisfied with their current situation. The third subcategory was the gap between dream and reality. There were two main keywords related to this subcategory: "gap" and "frustration", where many people said that there exists a huge gap between their dreams and real life, which will lead to a sense of failure.
Based on the above results, we found that working process was usually at the top of the list of stressors for young Chinese employees. These have centered on fears that their jobs have no future: the chance of being promoted is low, and it is even possible to get fired. Long working hours and heavy tasks also make them feel tired and under great strain. On the other hand, high living expenses such as buying houses, supporting aging parents, and so on, weigh heavily on their minds. In addition, it is worth noting that besides external pressures, many young employees are also constantly under the stresses resulting from the gap between their dreams and real lives. People have an easy access to luxurious lifestyles through social media, which can produce the illusion that they are inferior and lead to people's discontentment with their current situations.
The above results also provide further support for the JD-R theory. The identified stressors for young Chinese employees correspond precisely to the job demands, job resources, and personal resources. Job demands refer to those physical, social, or organizational aspects of the job that require sustained physical or mental effort and are therefore associated with certain physiological and psychological costs (e.g., exhaustion) [37]. The identified keywords related to "working long hours" and "lives occupied by busy working" express that these jobs require a lot of people's time and energy. Job resources refer to those physical, psychological, social, or organizational aspects of the job that may do any of the following: (a) be functional in achieving work goals; (b) reduce job demands at the associated physiological and psychological costs; and (c) stimulate personal growth and development [37].
The identified sentences related to "dead-end job", "social life at workplace", "dislike their jobs", "commute to work", and "the consumption of time and energy caused by social media" all express that employees feel very tired of work due to their lack of job resources such as job skills, social skills, and so on, which can create strain on employees. Meanwhile, all of the identified sentences related to "family life" express that employees have to pay a considerable sum of money to raise their families, thus they have to work flat out to earn enough income from work. Personal resources refer to the beliefs people hold regarding how much control they have over their environment [38]. The identified keywords related to "desire for material aroused by social media", "insatiable desire", "the gap between the rich and poor", and "the gap between dream and reality" only express that employees feel extremely miserable about the gap between their dreams and real lives. Additionally, our results also suggest that there are other stressors for employees that cannot be explained by JD-R theory: the sentences related to "rough time during the work" and "lack of social security" express that employees are often squeezed by their supervisor due to the lack of social guarantee of employee rights, which makes them feel exhausted. Table 6 compares our identified stressors with stressors used in the literature. Table 6. Comparison with stressors used in the literature.

Stressor Ours References Stressor Ours References
Working long hours Yes [31,36,42] Health-care expenses Yes [31] Dead-end job Yes Social pressure Yes Rough time during the work Yes Consumption of time and energy caused by social media Yes [22] Social life at workplace Yes [17] Desire for material aroused by social media Yes Dislike their jobs Yes [17,28,36] Lives occupied by busy working Yes [22] Commute to work Yes [10] Insatiable desire Yes Lack of social security Yes The gap between rich and poor Yes Low salary Yes [31,36] The gap between dream and reality Yes High cost of family life Yes [31] Emotional demands No [12,25,36] High housing costs Yes [31,34,35] Organizational injustice No [14,15,36] High cost to get married Yes [31] Performance stress No [8] Daily consumption Yes [31] Work-family conflict No [36] Education cost Yes [31,32] Being "infected" from other unhealthy organizations No [27] Supporting aging parents Yes [30]

Discussion
Through qualitative content analysis conducted by a machine-human hybrid approach, we developed a comprehensive understanding of stressors for young Chinese employees. Our results matched the JD-R model exactly. In particular, we systematically integrated employee stressors and identified three kinds of new stressors that have not been mentioned in the previous literature.
First, we found that being squeezed by supervisors, as mentioned in Table 2, is a new kind of employee stressor that is related to employee well-being. To the best of our knowledge, this is a stressor that has not been mentioned in the previous studies. Abusive behavior of supervisors has been identified as a stressor for young employees that cannot be neglected [48]. With gradual improvement in the legal system, the abusive behavior of supervisors has gradually disappeared. However, in fact, in many enterprises, being squeezed by supervisors is still an inescapable stressor for young employees. Many supervisors are constantly putting pressure on the employees to increase labor output. This leads to excessive physical and mental output that puts tremendous pressure on employees in the environment.
Second, we found that affection from others (see Tables 3 and 4) could also put a strain on employees. Previous research in organizations has demonstrated that the hard work of co-workers is for the affections from others, which are related to the employees' psychological stress [49]. However, importantly, we found clear evidence that employees were influenced not only by colleagues, but also by social media and loved ones. Our results indicate that social media is one of the most important ways through which employees can be affected by others. In recent years, the use of social media has become a global phenomenon. Over two-thirds of Internet users have active accounts on social networking sites [50]. This leads to the frequent occurrence of social media fearmongering, which causes individuals to indulge in frequent negative comparisons between themselves and others, resulting in many unrealistic material desires. In addition, we also found that pressure from parents such as pushing for marriage could also put psychological strain on employees.
Finally, we found that subjective cognitive bias (see Table 5) plays an important role in causing employee stress; the ability to cope with such subjects cognitively can be seen as a kind of personal resource. In this respect, we extended the research results of the JD-R model. Most of the previous research on personal resources in the JD-R model emphasize the people's perceptions of their control over their environment [38]. That is, it emphasizes the employees' subjective feelings about whether they can finish the work on schedule. However, through data analysis, we found that personal resources should also be extended to one's subjective feelings about the social environment. Due to subjective cognitive bias, employees feel that their life is worse than others, that there is a gap between dream and reality, and between material desire and reality. The frequent occurrences of social media fearmongering only aggravate these subjective cognitive biases. When these problems are not solved, they develop anxiety.
In light of our findings, several strategic directions have emerged to reduce employee stress and improve the employees' physical and mental health.
First, we need to focus on the two kinds of stressors of being squeezed by supervisors and a dead-end job ( Table 2) as they account for a large percentage of our results. Inadequate mechanism for employee protection is responsible for most of these stressors. Consequently, there is an urgent need to improve forms of interactions between employees and employers in the system of social and labor relations and develop social partnerships. Socially oriented employers should create favorable working conditions, improve the motivation and incentive systems, and ensure career growth [51]. In addition, other measures including social support, institutional guarantees, and legal protections should also be carried out to relieve employee stress.
Second, the negative impact of pressure from others and subjective cognitive bias are serious concerns. Strategies for limiting such negative impacts are therefore necessary. Employees should be guided to establish clear self-cognition, correct values and beliefs, and judgement ability for Internet information. Besides, websites should also be able to automatically identify, filter, and remove misinformation. Such methods are an effective means of alleviating employee psychological stress and promoting their psychological health.
Third, efforts should be made to reduce the mental stress, mental fatigue, and exhaustion experienced by employees with a family life. Suggested policy recommendations include improving social welfare protections for wage-earners. The provision of practical support for raising children, supporting the elderly, and buying houses could help employees to focus on their work more easily. In general, interactive mechanisms between employment and social security help to vigorously ensure and improve people's well-being, maintain social stability, and promote social harmony and progress.

Conclusions
In conclusion, through the technology of natural language processing and machine learning to mine UGC on Zhihu, we proposed a novel approach to investigate the details on stressors for young Chinese employees. The proposed approach for stressor identification is applicable not only in China, but also in other countries and regions. We also contribute to JD-R theory by expanding the scope of resources (i.e., mechanism for protecting employee rights and personal cognition). In light of our findings, we offer valuable practical insights and policy recommendations to relieve stresses and improve the physical and mental health of employees.
Our study can be extended in several directions. First, this research lies in the fact that our conclusions are based on the answers to the three questions from one social question and answer platform. The UGC data are from a single source, and do not contain socio-demographic and socio-economic characteristics of the users or the specific professions. In the future, we can utilize UGC data from various sources and conduct research together with other traditional social science methods. Second, our approach can be further improved, since it still needs human reading sentences for categorization. In the future, we can utilize automatic text classification techniques and text generation techniques for categorization and interpretation. Third, this research only discussed the stressors from the lack of resources and the increase in job demand. We encourage future research to further explore the factors in the context of the entire JD-R model and examine the direct and indirect effects on and of resources on the stressors. Finally, all of our samples were collected from China. We encourage future research to further explore the employee stressors in different countries and compare the differences.

Data Availability Statement:
We collected the raw UGC from "Zhihu", the largest Chinese social question answering platform. Zhihu, which started in December 2010, provides long-term detailed answers and discusses various questions raised by users. The Zhihu platform is a social question answering platform for people to share their knowledge or opinions. Zhihu users know that their answers are open to the public and the reason why they answer questions is to provide the public with specific knowledge or their views on social issues. There is no acquiescence bias involved as we did not acquire the answers at all.

Conflicts of Interest:
The authors declare no conflict of interest.