Early Detection of Workplace Stress Using NLP and Sentiment Analysis on Employee Communication Data

Abstract

One of the major concerns of the contemporary workplace scenario is the issue of job or work stress, which impacts the wellbeing and success of employees to a great extent. An enormous amount of textual communication data is produced every day owing to the extensive usage of digital communication platforms like social media, email communication, and instant messaging services among others. Information about the emotional state of mind of the employees is conveyed implicitly through digital communication media. Hence, stress prediction and intervention is possible using textual communication data.
This paper suggests a proposed approach for early prediction of workplace stress using NLP and describes a case study where sentiment analysis and various supervised learning approaches are applied on textual communication data. A total of 8,897 messages are gathered from publicly accessible Twitter and Reddit datasets. As a proxy to represent professional communication data, a set of preprocessing steps including text normalization, lowercase conversion, punctuation removal, stop words removal, tokenization, and lemmatization is conducted to enhance data quality and reduce any language noise in the text corpus. TF-IDF is one of the methods employed to preprocess textual data, making it possible for further processing by the machine learning model. In this research, different types of supervised learning models are examined, namely Support Vector Machines (SVM), Random Forest, Naive Bayes, and Logistic Regression. Among the employed supervised learning algorithms, Random Forest model is suggested to be the best classifier (82.75%), followed by Logistic Regression (82.24%).

Keywords