Automatic classification of behaviour change language can enhance conversational agents’ capabilities to adjust their behaviour based on users’ current situations and to encourage individuals to make positive changes. However, the lack of annotated language data of change-seekers hampers the performance of existing classifiers. In this study, we investigate the use of semi-supervised learning (SSL) to classify highly imbalanced texts around behaviour change. We assess the impact of including pseudo-labelled data from various sources and examine the balance between the amount of added pseudo-labelled data and the strictness of the inclusion criteria. Our findings indicate that while adding pseudo-labelled samples to the training data has limited classification impact, it does not significantly reduce performance regardless of the source of these new samples. This reinforces previous findings on the feasibility of applying classifiers trained on behaviour change language to diverse contexts.

My PhD focuses on conversational agents for behaviour change, with a focus on the feasibility of applying Large Language Models (LLMs) such as GPT-4 in this context.

Health behaviour change is a difficult and prolonged process that requires sustained motivation and determination. Conversa- tional agents have shown promise in supporting the change process in the past. One therapy approach that facilitates change and has been used as a framework for conversational agents is motivational interviewing. However, existing implementations of this therapy approach lack the deep understanding of user utterances that is essential to the spirit of motivational interviewing. To address this lack of understanding, we introduce the GLoHBCD, a German dataset of naturalistic language around health behaviour change. Data was sourced from a popular German weight loss forum and annotated using theoretically grounded motivational interviewing categories. We describe the process of dataset construction and present evaluation results. Initial experiments suggest a potential for broad applicability of the data and the resulting classifiers across different behaviour change domains. We make code to replicate the dataset and experiments available on Github.

In this description paper we outline the system architecture submitted to Task 4, Subtask 1 at SemEval-2022. We leverage the generative power of state of the art generative pretrained transformer models to increase training set size and remedy class imbalance issues. Our best submitted system is trained on a synthetically enhanced dataset with 10.3 times as many positive samples as the original dataset and reaches an F1 score of 50.62%, which is 10 percentage points higher than our initial system trained on an undersampled version of the original dataset. We explore possible reasons for the comparably low score in the overall task ranking and report on experiments conducted during the post-evaluation phase.