Detecting Expressive Writing in Online Health Communities by Modeling Aggregated Empirical Data

Although expressive writing has proved to be beneficial on physical, mental, and social health of individuals, it has been restrained to lab-based experimental studies. In the real world, individuals may naturally engage with expressive writing when dealing with difficult times, especially when facing a tough health journey. Health blogging may serve as an easy-to-access method for self-therapy, if spontaneous expressive writing occurs. However, many posts may not be expressive enough to provide the therapeutic power. In this study, we build a Gaussian naive Bayes model to detect expressive writing in an online health community, CaringBridge. Because we lack full text data as training data, we use a method to learn model parameters from meta-analysis of the literature. The classifier reaches reasonable accuracy on the test set annotated by the authors. We also explore factors that may influence users' spontaneous expressive writing. We find gender, health condition, author type, and privacy settings can affect individuals' spontaneous expressive writing. Finally, we reflect on our methodology and results and provide design implications for online health communities.