| dc.description.abstract | Public perception plays an increasingly important role in policy decision-making and resource allocation. Conventional studies on public perception have mainly relied upon formal surveys, which are limited in their ability to shed light on the matter because they are time- consuming, and the findings can become stale in the face of the rapid evolvement of the situation. Social media platforms have enabled people to report on their experiences and express their perspectives on a wide scale. It provides opportunities to investigate public perception in a timely manner.
This dissertation investigates how to utilize user-generated data to understand public perception of societal concerns. Our investigation revolves around two core questions: 1) Given the societal concern of interest, what are the topics of public concern, and 2) what is the public sentiment about the societal concern and related topics. These questions are answered through three related, but computational distinct, tasks. First, we investigate the behavior of online research cohort membership disclosure. We gathered and analyzed 4,020 tweets and uncovered over 100 tweets disclosing the individuals’ memberships in over 15 medical research programs. Through sentiment analysis, we learned that 45.3% of self-disclosed users have a positive attitude towards the joined research project. In the second task, to gain a deeper understanding of the self-disclosure behavior, we investigated the face image sharing trend in a Direct-to- Consumer genetic testing forum. Through topic modeling and statistical inference on over 15,000 Reddit posts, we found that posts including a face received 60% more comments and had karma scores (upvotes – downvotes) 2.4 times higher than other posts. The topics in posts including a face were primarily about sharing, discussing ancestry composition, or sharing family reunion photos. In the final part of this dissertation, we focus on a topic that affects the general population: public sentiment about COVID-19 and related topics. In this investigation, we combined word embedding models with clustering strategies to identify topics closely related to COVID-19, and relied upon the similarity between topic hashtags and opinion adjectives to infer the sentiment with respect to the identified topics. We discovered a significant difference between US urban and rural users in their sentiment about COVID-19 prevention strategies, misinformation, politicians, and the economy. The sentiment analysis approach is notable in that it can readily be extended to other topics of interest without additional data collection or model training. | |
