As self-directed online anxiety treatment and e-mental health programs become
more prevalent and begin to rapidly scale to a large number of users, the need to develop
automated techniques for monitoring patient progress and detecting early warning signs is
at an alltime high. While current online therapy systems work based on explicit quantitative
feedback from various survey measures, little attention has been paid thus far to the large
amount of unstructured free text present in the monitoring logs and journals submitted by
patients as part of the treatment process. In this paper, we automatically categorize patients’
internal sentiment and emotions using machine learning classifiers based on n-grams,
syntactic patterns, sentiment lexicon features, and distributed word embeddings. We report
classification metrics on a novel mental health dataset.