Proceedings of the 2nd Workshop on Computational Linguistics and Clinical Psychology: From Linguistic Signal to Clinical Reality, 2015.

Abstract

Motivational Interviewing (MI) is an efficacious treatment for substance use disorders and other problem behaviors (Lundahl and Burke, 2009). However, little is known about the specific mechanisms that drive therapeutic change. A growing body of research has focused on coding within-session language to better understand how therapist and patient language mutually influence each other and predict successful (or unsuccessful) treatment outcomes. These studies typically use human raters, requiring considerable financial, time, and training costs for conducting such research. This paper describes the development and testing of a recursive neural network (RNN) model for rating 78,977 therapist and patient talk turns across 356 MI sessions. We assessed the accuracy of RNNs in predicting human ratings for client speech and compared them to standard n-gram models. The RNN model showed improvement over ngram models for some codes, but overall, all of the models performed well below human reliability, demonstrating the difficulty of the task.

Links

Bib Entry 

@inproceedings{tanana2015recursive,
  author = {Tanana, Michael and Hallgren, Kevin and Imel, Zac and Atkins, David and Smyth, Padhraic and Srikumar, Vivek},
  title = {{Recursive Neural Networks for Coding Therapist and Patient Behavior in Motivational Interviewing}},
  booktitle = {Proceedings of the 2nd Workshop on Computational Linguistics and Clinical Psychology: From Linguistic Signal to Clinical Reality},
  year = {2015}
}