Incorporating Structural Information in Text Classification for Motivational Interviewing

Hosein Rezaei, Tommy Yuan, Tarique Anwar

Abstract: Speech act classification is a crucial task in the application of dialogue systems for mental healthcare. Motivational Interviewing (MI) in particular, requires at least two speech act classifiers for classifying the utterances of both client and clinician. State-of-the-art MI classifiers, despite their good performance, aren’t still accurate enough for being applied in a dialogue system. In order to improve their performance, we propose the use of different kinds of graphs such as Abstract Meaning Representation (AMR) or dependency graphs, either as direct input to graph neural networks or as a new attention layer in transformers. The impact of incorporating the structural information distilled in these graphs on classification will be reported.