Laughter Valence Prediction in Motivational Interviewing based on Lexical and Acoustic Cues

摘要

Motivational Interviewing (MI) is a goal oriented psychotherapy counseling that aims to instill positive change in a client through discussion. Since the discourse is in the form of semi-structured natural conversation, it often involves a variety of non-verbal social and affective behaviors such as laughter. Laughter carries information related to affect, mood and personality and can offer a window into the mental state of a person. In this work, we conduct an analytical study on predicting the valence of laughters (positive, neutral or negative) based on lexical and acoustic cues, within the context of MI. We hypothesize that the valence of laughter can be predicted using a window of past and future context around the laughter and, design models to incorporate context, from both text and audio. Through these experiments we validate the relation of the two modalities to perceived laughter valence. Based on the outputs of the prediction experiment, we perform a follow up analysis of the results including: (i) identification of the optimal past and future context in the audio and lexical channels, (ii) investigation of the differences in the prediction patterns for the counselor and the client and, (iii) analysis of feature patterns across the two modalities.