Keywords

1 Introduction

Job anxiety is a stimulus-related type of anxiety, occurring in the workplace or when thinking of the workplace [8]. Whilst the workplace can provide support [11], it can provoke anxiety too. High standards and expectations can lead to feelings of insufficiency. Social anxiety might be heightened due to monitoring and sanctioning, whilst competition with colleagues might lead to persecution fears [4]. With leadership, greater responsibilities can lead to overburdening, provoking anxiety [5].

Work anxiety is correlated with lower job commitment [1] can cause absenteeism and impact performance [7], further associated with minimized efficiency and increased expenses for businesses, insurances, and public pension funds.

Symptoms of work anxiety include typical anxiety symptoms, like blushing, trembling, or palpitations [8]. At present, despite its clinical and economical importance, only 1 questionnaire is available to measure job anxiety, known as the job anxiety scale (JAS) [6]. However, JAS’s validity is difficult to assess since it is a self-report scale, such that it consists of shortcomings due to acquiescence and social desirability [9]. Therefore, with no specific ICD codes for work anxieties, it can be difficult to characterize and thus diagnose them dynamically.

In context of these challenges, this study examines and develops a novel solution based on the previous metrics of job anxiety. The research here analyzes social media as a tool to express work-related concerns and thus to predict work anxiety. As people use Twitter to express their thoughts, social media can become a data source for job anxiety. By training an AI-based model on this data, work anxiety can effectively be detected.

This study aims to investigate the following question: can a mathematical model identify and quantify the relationship between a healthy corporate professional’s use of language in comparison with that of a corporate professional suffering from job anxiety?

The hypothesis pursued is, if a healthy corporate professional’s use of language is compared to that of a corporate professional suffering from job anxiety, then there exists a difference in terms of the linguistic markers, which can be identified by a mathematical model such that the use of these linguistic markers is further heightened in individuals suffering from higher levels of work anxiety.

The main contributions in this paper include

  1. 1.

    A Python script using the Twitter API and Selenium was developed to create a comprehensive dataset with tweets from over 15,000 Twitter users

  2. 2.

    Sentiment analysis of tweets in relation to levels of job anxiety (as expressed linguistically) over 3-month periods in 2019, 2020, and 2021

  3. 3.

    An SVM regression model to predict levels of job anxiety from tweets

The solution proposed in this study, the SVM regression model, combats the problems associated with other ways to measure levels of work anxiety, as it avoids response bias by not requiring a conscious undertaking like a survey and is also specific to job anxiety than anxiety generally. The solution thus has the potential to increase economic productivity and help combat mental health problems.

2 Background Literature

Generally, only some research has been conducted regarding work anxiety, as researchers often adapt different pre-existing instruments or tools. Nonetheless, one empirical way to measure perceived job anxiety is through the job anxiety scale (JAS) [6]. JAS is a self-rating scale to measure job anxiety, containing 5 main dimensions, (i) stimulus-related anxiety and avoidance behaviour, (ii) social anxieties, (iii) health-related anxieties, (iv) cognitions of insufficiency, and (v) job-related worries. However, the extensivity of the JAS renders it uneconomical and inefficient and places job anxiety in a very specific, unmalleable context. Therefore, to consider job anxiety as a wider construct for this study, a shortened, empirically derived version of the JAS, JAS-15 [10], was considered. The JAS-15 retains the 5 theoretically meaningful dimensions.

Whilst no research has been conducted on language use and work anxiety specifically, the previous research helps identify possible correlations between linguistic structures and anxiety. For example, the use of first-person singular pronouns during negative memory recall is associated with individuals suffering from anxiety [2]. It was also found that the use of unigrams like “withdrawal” and “severe” on Twitter indicated anxiety [3].

Whilst the background literature is useful in defining the study’s terms, they either do not consider job anxiety specifically or have not been implemented in a context relevant to social media. This lack of data suggests that further research is needed to determine linguistic markers of work anxiety, along with an efficient and accurate solution to realistically predict work anxiety levels.

3 Data Collection

3.1 Data Collection

A list of Twitter usernames of corporate professionals had to be collated.

First, a list of 10 keywords that corporate professionals might use in their tweets was created, identified based on prior knowledge.

Thereafter, using “search” on Twitter, it was manually checked that, for each keyword, most initial tweets from search results were by corporates. Then, using Web scraping with Python and Selenium (tool automating browsers), each keyword was searched and first 3000–3200 tweets collected. The no. of tweets could not be standardized due to memory overflow issues on Microsoft Edge (driver browser), but no less than 3000 and no more than 3200 tweets were collected from each keyword. The content, username, and name associated with tweets in each keyword were stored in CSVs.

However, not all usernames collected were those of corporate professionals, so irrelevant data had to be removed using filtering on Microsoft Excel. First, all duplicate usernames were removed and usernames containing the words “news” or “job” or links were removed (generally portals and not individuals). Then, going over each keyword file, general elimination trends (Table 1) were identified for each keyword.

Table 1 Elimination patterns for usernames
Full size table

Finally, a total of 20,358 usernames was collected.

Next, tweets by these usernames had to be collected for further analysis. To ensure that further algorithms and analyzes were not limited to heightened anxiety arising due to work-from-home during COVID-19, tweets from the identified users were collected over 3 periods of time—May–August 2019, May–August 2020, and January-April 2021. The Tweepy API (library for communication between Python and Twitter) was used to collect tweets. A script was run for 32 days, and >1 million tweets were collected.

However, not all users had sufficient Twitter activity from all time periods, so minimum 30 tweets were needed per period. Finally, data were retained from 8598 users.

3.2 Pre-Processing

Once the Twitter data were read into Python, pre-processing was conducted. All Unicode characters, emoticons, and links were replaced with an empty string to ensure that these special characters/symbols did not contribute to model training or sentiment analysis.

3.3 Sentiment Analysis

For each user, 3 values were calculated separately for each period. These 3 values were (i) positive sentiment score, (ii) negative sentiment score, and (iii) work anxiety score.

The positive sentiment and negative sentiment scores were calculated using the TextBlob (Python library for common NLP tasks) sentiment analysis algorithm.

For the work anxiety score, some words associated with work anxiety, adapted from JAS-15, were identified, likely to be used by people suffering from work anxiety. Each time a pair of words was used in a tweet, the raw work anxiety score was incremented, and finally, the raw work anxiety score was divided by total number of words by user to obtain final work anxiety score. Whilst this was a primitive approach before a machine learning approach, it helped in understanding the linguistic potential of JAS-15.

  1. 1.

    “always worry” AND “work”

  2. 2.

    “minor matters at workplace” OR “minor matters at work” AND “worry”

  3. 3.

    “workplace” OR “job” OR “work” AND “restricting my capacities for achievement”

  4. 4.

    “suffer” OR “miserable” OR “ill health” OR “bad health” OR “not feeling well” OR “unfairness” OR “avoid” AND “workplace”

  5. 5.

    “work” AND “nervous” OR “my health” OR “suffer” OR “miserable” OR “ill health” OR “bad health” OR “not feeling well” OR “unfairness” OR “avoid”

  6. 6.

    “anxiety” OR “suffer” OR “miserable” OR “ill health” OR “bad health” OR “not feeling well” OR “unfairness” OR “avoid” AND “job”

  7. 7.

    “problem” AND “superiors” OR “boss” OR “colleagues”

    Thereafter, for each period, the Pearson correlation coefficient was calculated between negative sentiment or positive sentiment and work anxiety.

3.4 Data Labelling

After conducting sentiment analysis, data from some users were manually read and labelled on the severity of work anxiety expressed on a scale of 0 to 1, according to JAS-15. The adaptation of JAS-15 for identifying tweets is shown below (Table 2).

Table 2 JAS-15 for Twitter
Full size table

Ground truth train data were collected and labelled from 40 random users, test data were collected and labelled from 15 random users, and data from 20 users (with high work anxiety indexes) were used in creating the token bag (bag of words).

Bag of words contained 50 most common words (no stop words) from those 20 users.

3.5 Model Training

A support vector machine (SVM) algorithm was used for creating the work anxiety regressor, trained on 6 feature vectors. SVM was chosen given the small dataset (with less noise), high prediction accuracy, support for kernels, easy implementation, and memory efficiency. The dimensions of the JAS-15 were used to identify 5 features. The training and testing data also took the same feature vectors into account since they were labelled according to the JAS.

The 6 feature vectors included

  1. 1.

    No. of times workplace mentioned (stimulus-related anxiety)

  2. 2.

    No. of times colleagues are mentioned (social anxiety and cognition of mobbing)

  3. 3.

    No. of times health is mentioned (in relation to health- and body-related anxieties)

  4. 4.

    No. of times ambition is mentioned (in relation to cognition of insufficiency)

  5. 5.

    No. of times deadline/pressure is mentioned (in relation to job-related worrying)

  6. 6.

    No. of times words from bag of words used

    A corpus was created for each feature vector and then values found.

Thereafter, an SVM regressor was trained and tested.

4 Results—Sentiment Analysis

The Pearson correlation coefficients, across the 3 time periods, between scores have been shown in Table 3.

Table 3 Correlation between sentiments and work anxiety
Full size table

It can be seen that generally, there is a low negative correlation between positive sentiment and work anxiety, whilst there is a moderate positive correlation between negative sentiment and work anxiety. This is consistent with the researchers’ expectations as work anxiety is more strongly correlated with negative outcomes.

The correlation between negative sentiment and work anxiety became stronger between 2019 and 2021, which could be due to rising negative sentiment amidst COVID-19 and expected increase in work anxiety due to general increase in uncertainty.

5 Results—SVM Algorithm

The trained SVM regressor achieved a root mean squared error (RMSE) of 0.2 and an accuracy of 83%, which indicates that whilst the algorithm might be successful in predicting work anxiety in a certain range, it needs more data for higher accuracy and lowered error.

6 Discussion

There are several limitations to this study that might have acted as sources of error.

Firstly, the labelling of ground truth and test data was relative, due to the relative nature of the JAS-15 itself, and subject to human error. Furthermore, only 40 users were used in testing and 15 users in training, which is a small sample size. To appropriately train and evaluate, at least 100 users should have been used in testing and training alike.

There were limitations in the sentiment analysis as well. As was seen with username collection, a lot of irrelevant data had been collected. Whilst it was attempted to remove most irrelevant data, it could not be ensured that all of it was removed, and thus, lowered noise in the dataset could not be guaranteed. Furthermore, other algorithms besides TextBlob could have been used, as TextBlob only has an accuracy of ~56% [12].

Future directions could include training the algorithm with more data or implementing an app with a GUI to enable user understanding of work anxiety as well.

7 Conclusions

The hypothesis was supported in that specific linguistic markers (through the dimensions of the JAS) can be used to predict levels of work anxiety, based on JAS-15, as there is a measurable difference in language use between a healthy individual and an individual suffering from some level of work anxiety. Twitter, therefore, serves as a good source of data to predict a corporate professional’s level of work anxiety.