Deep Neural Architectures for Discourse Segmentation in E-Mail Based Behavioral Interventions

Communication science approaches to develop effective behavior interventions, such as motivational interviewing (MI), are limited by traditional qualitative coding of communication exchanges, a very resource-intensive and time-consuming process. This study focuses on the analysis of e-Coaching sessions, behavior interventions delivered via email and grounded in the principles of MI. A critical step towards automated qualitative coding of e-Coaching sessions is segmentation of emails into fragments that correspond to MI behaviors. This study frames email segmentation task as a classification problem and utilizes word and punctuation mark embeddings in conjunction with part-of-speech features to address it. We evaluated the performance of conditional random fields (CRF) as well as multi-layer perceptron (MLP), bi-directional recurrent neural network (BRNN) and convolutional recurrent neural network (CRNN) for the task of email segmentation. Our results indicate that CRNN outperforms CRF, MLP and BRNN achieving 0.989 weighted macro-averaged F1-measure and 0.825 F1-measure for new segment detection.