Doctoral Dissertation Research in DRMS: Judgments of Answerers and the Answers They Give

人类社会生活的一个基本特征是提问和回答问题，这是一项复杂的任务，需要一个人决定问谁，如何表达问题，以及如何解释答案并判断其价值。然而，尽管这一活动具有中心地位，但人们对这一过程的每一步所涉及的一般原则却知之甚少，因此很难设计出更好的方法来帮助人们提出问题并找到他们所需要的信息。在信息时代，这个问题变得更加紧迫：人们问的问题比以往任何时候都多，涉及的话题也更广泛，他们不仅把这些问题问给朋友和家人，还问给网上公告板和社交媒体上的陌生人和熟人。这篇论文的工作解决了我们的理解差距，使用心理学和计算机科学的方法相结合。机器学习和人工智能的工具将用于分析人们在在线公告板系统上产生的大规模问题和答案集合，结果将用于构建解释和预测人们如何以及何时找到问题的好答案的一般理论。在测试这些理论的实验室实验中，参与者将看到具有不同基本属性的问题和答案，结果将确定对信息收集特别重要的特征。数据科学调查的大规模性质提供了基于问题和答案之间关系的微妙模式的理论发展，而实验室工作允许调查问题的因果性质：对于提问者来说，什么使答案好？人们探索和了解世界的最常见方式之一是通过提问从他人那里寻求信息。这种行为需要两种判断：（1）应该从谁那里寻求答案;（2）应该如何判断答案是否满足问题？这两种判断是相互依赖的，要求人们在不确定的情况下做出决策。这种对如何做出这种判断的双管齐下的调查不仅包括基于实验室的实验，这是社会和决策科学的传统优势，而且还包括数据科学的组成部分，研究这些决策是如何在野外做出的。这种方法的组合将被用来检查好的答案和那些给他们的定义特征。研究的重点是人与人之间的语言对话中的问题和答案。数据科学和计算语言学的技术将用于分析在线讨论论坛上产生的问题和答案，从Stackexchange上的技术帮助请求到Mumsnet上父母提出的更微妙的社会问题。问题和答案之间的关系的两个关键属性，在何种程度上，他们重叠，并在何种程度上回答集中的问题所提出的解决方案空间，假设对应的答案是如何评估。受控实验室实验将用于阐明那些回答问题的人的偏好。提问者对潜在回答者的偏好类型直接影响到所收到信息的相关性和质量，反过来又影响到后续回答和进一步决策的效用。贝叶斯框架是用来正式面对的提问者的任务：推断什么类型的回答者的个人是，鉴于他们过去的回答行为。该奖项反映了NSF的法定使命，并被认为是值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估的支持。

A basic feature of human social life is asking and answering questions, a complex task that requires one to decide whom to ask, how to phrase the question, and how to interpret the answer and judge its worth. Despite the centrality of the activity, however, little is understood about the general principles involved in each step of the process, making it difficult to design better ways to help people ask questions and find the information they need. The problem becomes even more pressing in the information age: people ask more questions than ever before, on a wider range of topics, and they take those questions not only to friends and family, but also to strangers and acquaintances on the wider world of online bulletin boards and social media. This dissertation work addresses gaps in our understanding, using a combination of methods from psychology and computer science. Tools from machine learning and artificial intelligence will be used to analyze large-scale collections of questions and answers that people produce on online bulletin board systems and results will be used to build general theories that explain and predict how and when people find good answers to their questions. In laboratory experiments to test these theories, participants will be shown questions and answers with varying underlying properties, and the results will identify features that are particularly crucial for information gathering. The large-scale nature of the data science investigation affords the development of theories based on subtle patterns in the relationship between question and answer, while the laboratory work allows investigation into the causal nature of the problem: what, for the question-asker, makes an answer good?One of the most common ways people explore and learn about the world is by seeking information from others through asking questions. This behavior requires two kinds of judgments: (1) from whom one should seek answers and (2) how should one judge whether the answer satisfies the question? These two judgments are inter-dependent and require that people make decisions under uncertainty. This two-pronged investigation into how such judgments are made includes not only the lab-based experiments that are a traditional strength of social and decision sciences, but also a data science component that looks at how these decisions are made in the wild. This mix of methodologies will be used to examine the defining features of good answers and those who give them. The research focuses on questions and answers that are given in human-to-human linguistic dialogue. Techniques from data science and computational linguistics will be used to analyze collections of questions and answers produced on online discussion forums, ranging from requests for technical help on Stackexchange to more nuanced, social questions asked by parents on Mumsnet. Two key properties of the relationship between questions and answers—the extent to which they overlap, and the degree to which answers focus the solution space posed by the question—are hypothesized to correspond to how answers are evaluated. Controlled laboratory experiments will be used to elucidate preferences for those who answer questions. The types of preferences a questioner has for potential answerers has direct consequences for the relevance and quality of information received, and in turn affects the utility of subsequent answers and further decisions. A Bayesian framework is used to formalize the task facing the questioner: to infer what type of answerer an individual is, given their past answering behavior. Behavioral experiments will test key predictions of the model.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.