AN APPROACH USING AFFECTIVE COMPUTING TO PREDICT INTERACTION QUALITY FROM CONVERSATIONS

John Gottman’s mathematical models have been shown to accurately predict a couple’s style of interaction using only the sentiments found in the couple’s conversations. I derived speaker sentiment slopes from 151 recorded dyadic audio conversations from the IEMOCAP dataset through an IBM Watson emotion recognition pipeline and assessed its accuracy as input for a Gottman model by comparing the cumulative speaker sentiment slope for each conversation produced from predicted emotion codes to that produced from groundtruth codes provided by IEMOCAP. Watson produced sentiment slopes strongly correlated with those produced by groundtruth emotion codes. An abbreviated pipeline was also assessed consisting just of the Watson textual emotion recognition model using IEMOCAP’s human transcriptions as input. It produced predicted sentiment slopes very strongly correlated with those produced by groundtruth. The research demonstrated that artificial intelligence has potential to be used to predict interaction quality from short samples of conversational data.