Public health surveillance from social media relies on being able to separate what people say about their own health from other discussions that use the same keywords for other reasons. Many applied studies in the area have failed to take this into account, leading to biased results that could overrepresent the importance of certain health issues by location or over time and lead to poor quality decision-making about policy or resource allocation. We directly address this problem here by introducing new ways to more explicitly capture the different ways people use disease and symptom keywords on Twitter, improving the overall performance when detecting what people say about their own health, and creates new opportunities to understand how figurative language shapes conversations on social media and other user-generated text.

People often discuss their health on social media platforms. Discussion of personal experiences with diseases and symptoms can be useful in public health applications like adverse event surveillance. A major challenge comes from the need to distinguish personal health mentions from other uses of those terms, including figurative use, where words are used to mean something different. Public health applications require the separation of personal health mentions from other uses. Prior approaches incorporate some elements of context but could be improved to capture relationships between the linguistic characteristics of figurative expressions and the representations of the context. In this work, we investigate the role of context representation for identifying personal health mentions on social media and measure the impact of different representation choices on detecting figurative use of a range of disease and symptom words. We present an end-to-end approach that selects representations a...