I liked this presentation's direction.
2 things I would do:
1) Instead of theorizing which linguistic components might predict therapy outcome, just let the system find them. Search engine technology can do much more than just spot future tense and first person, for example, analyzing the semantics to determine the subjects in the text.
2) Effect measurement is not clear. I did not read the full article, but since it is not specified I assume nothing new. I would try to measure the results online - to see a timeline interaction between writing of the texts and well being, and also measure them in indirect ways such as biofeedback and response time to stroop tasks, since self report may interact with personal writing. Even if I am wrong, self report should be crossed with implicit measures.
So cool :)