Modeling Low-Resource Health Coaching Dialogues via Neuro-Symbolic Goal Summarization and Text-Units-Text Generation

TL;DR

The paper tackles the high cost and data demands of health coaching by introducing a neuro-symbolic goal summarizer that operates without predefined schemas and a text-units-text dialogue generator that leverages discrete dialogue units to generate coaching conversations. It demonstrates that these data-efficient, interpretable models achieve state-of-the-art semantic frame accuracy and superior response quality on health coaching data, including a new dataset enriched with Fitbit progress. A novel extended PVI-based data difficulty metric supports coach alerts and curriculum-guided training, addressing unconventional patient responses in deployment. The work provides a practical framework for scalable, low-resource health coaching dialogue systems with potential applicability to broader behavioral change and patient education contexts.

Abstract

Health coaching helps patients achieve personalized and lifestyle-related goals, effectively managing chronic conditions and alleviating mental health issues. It is particularly beneficial, however cost-prohibitive, for low-socioeconomic status populations due to its highly personalized and labor-intensive nature. In this paper, we propose a neuro-symbolic goal summarizer to support health coaches in keeping track of the goals and a text-units-text dialogue generation model that converses with patients and helps them create and accomplish specific goals for physical activities. Our models outperform previous state-of-the-art while eliminating the need for predefined schema and corresponding annotation. We also propose a new health coaching dataset extending previous work and a metric to measure the unconventionality of the patient's response based on data difficulty, facilitating potential coach alerts during deployment.

Figures (2)

• Figure 1: A simplified demonstration of Neuro-Symbolic Goal Summarization. The health coach discusses the goal for week $w_t$ by referring to the goal set in the previous week $w_{t-1}$ ("same days?"). Our model is trained to generate an executable instruction (Copy {Days}) and to extract the partial goal ("Walk 2,500 steps") from the dialogue of the current week. The model then edits the partial goal by applying the instruction to the reference previous goal, resulting in the comprehensive goal summary: "Walk 2,500 steps from Monday to Friday."
• Figure 2: Human evaluation of generated response by health coaches. Our model's outputs have a $14.29\%$ preference over previous state-of-the-art zhou-hc-22.