RAPID: Automated Extraction and Validation of the Gist of Social Media Messages about COVID-19

当新型冠状病毒席卷全球时，民主社会面临着一个困境，即如何鼓励牺牲以降低风险，即使威胁是看不见的，而且似乎正在消退。当风险很高时，个人可能需要长期保持极端形式的社交距离，但这种选择会带来经济和个人成本。然而，不参与这些减少风险的活动可能会导致数百万人死亡。因此，公共卫生传播者必须了解拒绝这些行动的理由，并根据经验支持的原则进行风险沟通，包括向做出这些具有挑战性选择的个人传达这些行动的要点（底线意义）的必要性。本项目由三个研究组成，每个研究在不同的时间点有两个样本，最终目的是将社交媒体信息的gist转化为有效的风险沟通干预措施。该方法的动机是模糊痕迹理论——一种基于证据的关于风险下健康决策的解释——它假设决策是基于刺激的定性基本表征，这些刺激在上下文中编码了基本含义。学者们使用自下而上的无监督机器学习算法来测试对时间敏感的假设，以表征有关COVID-19的社交媒体信息中的主题，以及自上而下的人类判断，即理论上可预测的感知风险的心理表征和降低风险行为的好处。研究1通过从Twitter和Facebook公开帖子中提取数百万条社交媒体信息，量化了与社交距离和其他降低风险行为（例如洗手）有关的gist的流行程度。这包括使用概率主题模型对正在进行的数百万条社交媒体消息的语料库进行自动编码。研究2提供了一种系统的方法，通过人工判断来解释主题模型在不同时间点提取的主题。研究3涉及一组不同的人类评委，评估主题是否与风险和收益的主要表示的目标集一致，以及表达价值的主要原则。研究人员对先前研究中成功使用的项目进行了改造。类似的主旨表示已经预测了许多健康状况下自我报告的风险降低行为。因此，在研究3中，研究人员通过实地进行理论动机调查来验证主题学家，分析其信度和结构效度。重点是了解风险和益处的要点，特别是能够促进个人和社会健康的要点。总体而言，该项目引入了从社交媒体中获取专家的新技术，这些技术可用于产生有意义的公共卫生交流。该项目直接为公共卫生传播者现有的尝试提供信息，以表达民众应采取的风险、利益和行动，以减轻COVID-19大流行的传播。这些度量是针对人类用户进行验证的，从而使团队能够实现准确性和规模。该项目可作为一项更大努力的基础，该努力可以提高要点提取自动化的程度，帮助公共卫生传播者迅速了解在疫情期间应向哪些社区传播哪些要点。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

While the novel coronavirus sweeps the globe, democratic societies face a quandary, namely, how to encourage sacrifices to reduce risk, even when the threat is invisible and seems to be receding. When risk is high, individuals may need to engage in extreme forms of social distancing for long periods, but this choice comes with economic and personal costs. Not engaging in these risk-reduction activities, however, could cause millions of deaths. It is therefore crucial for public health communicators to understand rationales for refusing these actions and to base risk communication on empirically supported principles, including the need to convey the gist (bottom-line meaning) of these actions to individuals making these challenging choices.This project consists of three studies, each with two samples at different time points, with the ultimate goal in subsequent studies of converting the gists of social media messages into effective risk-communication interventions. The approach is motivated by Fuzzy Trace Theory–an evidence-based account of health decision making under risk–which posits that decisions are based on qualitative gist representations of stimuli that encode basic meaning in context. The scholars test time-sensitive hypotheses using bottom-up unsupervised machine-learning algorithms to characterize the topics in social media messages about COVID-19 and top-down human judgments about theoretically predictable mental representations of the gist of perceived risks and benefits of risk-reduction behaviors. Study 1 quantifies the prevalence of gists pertaining to social distancing and other risk-reduction behaviors (e.g., hand washing) by extracting millions of social media messages from Twitter and public Facebook posts. This involves automatic coding an ongoing corpus of millions of social media messages using probabilistic topic models. Study 2 provides a systematic approach with human judges to interpreting the topics extracted by a topic model at different time points. Study 3 involves a different group of human judges assessing topics for their consistency with a targeted set of gist representations of risks and benefits, along with gist principles that express values. The researchers adapt items used successfully in prior research. Similar gist representations have predicted self-reported risk-reduction behaviors for numerous health conditions. Thus, for Study 3, the researchers validate topic gists by fielding a theoretically motivated survey analyzed for reliability and construct validity. The focus is on perceptions of the gist of risks and benefits, especially gists that can promote the health of individuals and society. Overall, this project introduces novel techniques for eliciting gists from social media that can be used to generate meaningful public health communications. The project directly informs existing attempts by public health communicators to express the risks, benefits, and actions that members of the population should take to mitigate the spread of the COVID-19 pandemic. The measures are validated against human users, enabling the team to achieve both accuracy and scale. This project serves as the basis for a larger effort that can increase the extent to which gist elicitation may be automated, helping public health communicators to quickly understand which gists should be communicated to which communities during outbreaks.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.