Natural language understanding for computational psychotherapy applications

Abstract

In this thesis, I tackle the problems in recently emerging technology-mediated psychotherapy through computational methods. Among them, I focus on developing natural language understanding techniques since verbal interaction is most important factor to understand the interaction. Specifically, I develop models and resources to understand languages of clients in counseling, languages of people in suicidal risk, languages of people in various emotional state. For language understanding of clients in text-based counseling, I develop a categorization method as a labeling scheme for client utterances. I also propose a new model, Conversation Model Fine-Tuning to classify the utterances with small size of labeled data. This allows us to understand client's language and automatically extract meaningful information from them. For language understanding of people in suicidal risk, I build a dataset of social media posts written by military personnel with corresponding expert annotations of suicidal risk levels. Various pretrained language models are further fine-tuned by using the dataset to classify the risks for developing simple yet effective baselines, achieving high classification performance. This could be applied to help them in time. For language understanding of people in various emotional states, I propose a framework which enables a model to learn to predict dimensional emotions as well as categorical emotions, only trained from corpus annotated with categorical emotion labels, to give better emotional feedback in self-help psychotherapy without labeled data. Dimensional emotion predicted by a model trained using our framework shows significant positive correlations to corresponding ground truth without direct supervision. Through these contributions, our knowledge could advance in understanding dynamics in technology-mediated psychotherapy and relevant information seeking behaviors of people in need. Clients and therapists could be supported in practice by automatized computational models as well.