Xin Sun, Jos A. Bosch, Jos Dobber and Gert-Jan de Bruijn
Dialogue and behavioural analysis of psychotherapy conversations can effectively assess and guide psychotherapy counselling. In practice, the conversational data in the field of psychotherapy is difficult to collect because of privacy issues, which makes it difficult to analyse and evaluate real-life psychotherapy conversations. In this work, we focus on the study of Motivational Interviewing (MI, Miller and Rollnick, 2003, 2012), a psychotherapy strategy that is effective for behaviour change, especially for addictive and lifestyle behaviours. We obtain a Dutch MI corpus from clinics, and these MI conversations are manually parsed and coded by MI experts using the Motivational Interviewing Skill Codes (MISC, Miller et al., 2003). We translate the Dutch corpus into English, reorganise the MISC codes and expand them with more psychological labels, resulting in a complete MISC annotated bilingual (Dutch-English) MI conversation corpus. Subsequently, we perform the following tasks on this fully annotated MI corpus:
(1) Classification: Dynamically identifying the MI behavioural codes of ongoing patient utterances;
(2) Prediction: Predicting the most appropriate upcoming MI behavioural code for the therapist based on the ongoing therapy dialogue history;
(3) Analysis: Exploring the implicit correlation on MI behavioural codes between therapist and patient, also the correlation between the therapist utterances, patient utterances and motivation for behaviour change by analysing both word and utterance-level attention;
For the above three tasks, we model the dialogue with pre-trained language models and analyse the correlations through both attention and self-attention mechanisms. Possible contributions of this work can be:
(1) Creation of bilingual Motivational Interviewing conversation corpus with full MISC annotation and extra psychological labels;
(2) Identified patient’s MI behavioural codes can enable therapists to monitor the progress and efficacy of the psychotherapy;
(3) Predicted therapist’s MI behavioural codes can provide real-time assistance and guidance to therapists, and can also be used to evaluate the therapist’s behaviours;
(4) Linguistic analysis of MI dialogue on both word and utterance levels can help interpret the implicit correlation between the therapist’s codes/utterances, patient’s codes/utterances, motivation, and MI effectiveness;