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

Area of Science:

• Behavioral science
• Communication science
• Computational linguistics

Background:

• Behavioral interventions like motivational interviewing (MI) are crucial for health.
• Traditional qualitative coding of communication exchanges is resource-intensive and time-consuming.
• Automating the analysis of e-Coaching sessions, which deliver MI-based interventions via email, is needed.

Purpose of the Study:

• To develop an automated method for segmenting emails in e-Coaching sessions.
• To identify email fragments corresponding to specific MI behaviors for qualitative coding.
• To evaluate machine learning models for email segmentation in behavior interventions.

Main Methods:

• Framed email segmentation as a classification problem.
• Utilized word and punctuation embeddings with part-of-speech features.
• Evaluated Conditional Random Fields (CRF), Multi-layer Perceptron (MLP), Bi-directional Recurrent Neural Network (BRNN), and Convolutional Recurrent Neural Network (CRNN).

Main Results:

• The CRNN model outperformed CRF, MLP, and BRNN models.
• Achieved a 0.989 weighted macro-averaged F1-measure for email segmentation.
• Demonstrated strong performance in new segment detection with an 0.825 F1-measure.

Conclusions:

• CRNN is a highly effective model for automated email segmentation in MI-based e-Coaching.
• This approach significantly advances the potential for scalable, automated qualitative coding of digital behavior interventions.
• Automated segmentation facilitates more efficient analysis of communication data in behavioral science research.