Towards the Use of Language Models in Interventions of Vaccine Concerns

Abstract

Increased vaccination rates over the past 50 years have saved hundreds of millions of lives worldwide. Many of the lives saved have been those of children in developing countries. Vaccines have led to the eradication of smallpox and drastic reduction of polio. They are one of the most effective tools for improving global health. However, they depend on high coverage. Vaccine hesitancy, the delay or refusal of recommended vaccinations, has increased within developed countries in recent years. This troubling trend has led to increased efforts within public health towards direct interventions of vaccine attitudes. At the same time, language models have shown incredible potential in recent years, becoming a versatile means to easily implement artificial intelligence in new technological innovations. It is therefore natural to see language models as a potential tool towards the intervention of vaccine hesitancy, and to understand the specific concerns which might lead to it. However, the eventual adoption of language model-based systems on high-stakes tasks—such as those central to public health—still requires significant evaluation and thoughtful system design. This represents a gap in the research literature. There is a need for a deeper understanding of the consequences of using language models in these high-stakes tasks. This dissertation explores key considerations required to effectively, efficiently, and ethically leverage language models within the domain of interventions for vaccine concerns, a growing contributor of vaccine hesitancy. Specifically, the investigations span across mapping the landscape of concerns and developing the approaches needed for continuous, automatic monitoring; tools and best practices for text annotation within this high-difficulty, high-subjectivity set of tasks; methods for characterizing the confidence we can place in automatic evaluation; as well as the generation of tailored, text-based interventions and how humans perceive such interventions. Collectively, these investigations help move the field towards the use of language models in interventions of vaccine concerns, while outlining the remaining limitations and challenges.