Abstract
We introduce text mining to study work engagement by using this method to classify employees' survey-based self-narratives into high or low work engagement and analyzing the text features that contribute to the classification. We used two samples, representing the 2020 and 2021 waves of an annual survey among healthcare employees. In the first study, we used exploratory sample 1 (N = 5591) to explore which text features explain work engagement (unigrams, bigrams, psychological, or linguistic). In the second study, we confirmed whether features persisted over time between exploratory sample 1 and confirmatory sample 2 (N = 4470). We find that psychological features classify employees across two samples with 60% accuracy. These features partly validate the literature: High-engaged employees refer more to affiliation and positive emotions, and low-engaged employees refer more to negative emotions and power. We extend the literature by studying linguistics: High-engaged employees use more first-person plural ("we") than low-engaged employees. Finally, some results question the literature, like the finding that low-engaged employees refer more to their managers. This study shows text mining can contribute by confirming, extending, or questioning the literature on work engagement and explores how future research could build on our findings with survey-based or in vivo applications.