III: Medium: Adaptive Information Extraction from Social Media for Actionable Inferences in Public Health

社交媒体是几乎所有产品和服务的非策划的用户生成反馈的主要来源。用户越来越多地依赖社交媒体来披露严重的现实生活事件，例如餐馆的食物中毒事件，而不是访问官方沟通渠道。这种有价值的用户生成信息如果得到可靠的识别，可能会对与公共卫生相关的关键应用程序（本项目感兴趣的应用程序系列）及其他应用程序产生巨大的积极影响。例如，如果可以从社交媒体中推断出支持调查的令人信服的证据，当地卫生部门可能会对餐馆的潜在食源性疾病爆发进行调查。该项目将解决与处理社交媒体相关的基础研究挑战，以产生可操作的推论，该过程的输出导致真实的世界中的具体行动。除了产生广泛适用的研究成果外，该项目还将把一个关键的公共卫生应用作为其核心，即检测和应对餐馆中的食源性疾病爆发。总体而言，该项目将开发（1）以实体为中心的建模和社交媒体选择策略，以涵盖用户产生的大量内容。（2）非正式、嘈杂和不合语法的文本上的非传统信息提取策略，以及基于学习的方法，以产生可操作的、以实体为中心的公共卫生应用推理;以及（3）用于一般在线主动学习和搜索的方法，其被调整用于检测可操作推断所需的罕见和不频繁的发生。此外，该项目将围绕检测和采取行动的食源性疾病爆发的应用，在哥伦比亚大学和纽约市卫生和心理卫生部（DOHMH）之间的联合合作。此次合作将为应用于大规模数据科学问题的新研究成果提供一个强大的真实世界平台，这是评估计算机科学研究的难得机会。这种合作将包括开发和部署一个对公共卫生和社会有直接影响的系统。一个概念验证原型已经在DOHMH使用，并帮助识别和应对几次以前未知的疫情。该项目的公共卫生调查结果将按照卫生和人力部的最佳做法在政府机构之间共享。将通过项目网站（http：//publichealth.cs.columbia.edu/）与其他研究人员和机构分享已开发的编码和附加说明的数据集。

Social media is a major source for non-curated, user-generated feedback on virtually all products and services. Users increasingly rely on social media to disclose serious real-life incidents, such as a food poisoning incident at a restaurant, rather than visiting official communication channels. This valuable user-generated information, if identified reliably, may have a dramatic positive impact on critical applications related to public health -- the family of applications of interest in this project -- and beyond. For example, a local health department might launch an investigation of a potential foodborne disease outbreak at a restaurant if compelling evidence supporting the investigation can be inferred from social media. This project will address fundamental research challenges associated with processing social media to produce actionable inferences, where the output of the process leads to concrete actions in the real world. In addition to producing broadly applicable research results, the project will have as its centerpiece a critical public health application, namely, detecting and acting on foodborne disease outbreaks in restaurants.Overall, this project will develop (1) strategies for entity-centric modeling and selection of social media, to cover the vast volumes of user-produced content across sources; (2) non-traditional information extraction strategies over informal, noisy, and ungrammatical text, as well as learning-based approaches to produce actionable, entity-centric inferences for public health applications; and (3) methods for general online active learning and search that are tuned for detecting the rare and infrequent occurrences required for actionable inferences. Furthermore, this project will center around an application of detecting and acting on foodborne disease outbreaks, in a joint collaboration between Columbia University and the New York City Department of Health and Mental Hygiene (DOHMH). This collaboration will provide a robust, real-world platform for a continuous, end-to-end evaluation of the novel research results as applied to a large-scale data science problem, a rare opportunity in the evaluation of Computer Science research. This collaboration will include the development and deployment of a system with a direct impact on public health and society. A proof-of-concept prototype is already in use at DOHMH and has helped identify and act on several previously unknown outbreaks. The public health findings from the project will be shared across governmental agencies, following DOHMH's best practices. Developed code and annotated datasets will be shared with other researchers and agencies via the project web site (http://publichealth.cs.columbia.edu/).