A Bayesian hierarchical mixture model for extracting process features in a newly developed cloud-based writing...

In a computer-based writing platform, keystroke log data can document a student's writing process (e.g., typing behaviors). A set of tools has been developed to collect the resulting log data. However, these tools originate from decades ago, and none of them has been adapted for use with cloud servers. In this research, we describe a cloud-based writing platform that we have developed, Clourite. It offers a series of benefits including improved scalability and automatic data collection. It is easy to use and requires only a few mouse clicks without installing any software. Moreover, with a sample of 309 college students, we collected empirical keystroke log data through Clourite and extracted writing process features using a Bayesian hierarchical mixture model. The non-hierarchical mixture model may enlarge the standard error when the number of pause events in that component is small, as shown in our previous work. In contrast, the proposed hierarchical model in a fully Bayesian framework has the advantage that essays are considered similar data units, meaning the model can borrow strength from longer essays to enhance the efficiency of estimation for shorter essays. We identified a distinct writing process pattern differentiating stronger and weaker writers, along with the association between process features and writing quality measures (i.e., human scores). These findings underscore the potential of keystroke logging analysis for characterizing student writing processes.