CAREER: Human-Computer Cooperation for Word-by-Word Question Answering

这个CAREER项目研究人类和计算机如何协同工作来回答问题。人类和计算机拥有互补的技能：人类对世界有广泛的常识理解，更擅长使用非常规语言，而计算机可以毫不费力地记住无数的事实，并在瞬间检索它们。这一提议有助于机器理解人、地点和人物；如何将这些信息传达给人类；以及如何让人类和计算机在使用有限信息的情况下合作回答问题。这个提议的一个关键组成部分是逐字回答问题：这迫使人类和计算机尽可能有效地利用信息回答问题。除了将这些技能嵌入到问答任务中，该提案还有一个广泛的推广计划，在高中生和大学生的互动问答比赛中展示这项技术。这项研究可以通过对实体之间关系进行编码的中维嵌入中的实体的新表示来实现（例如，“Goodluck Jonathan”和“Nigeria”的表示编码前者是后者的领导者），从而使系统能够回答有关尼日利亚的问题。我们通过传统的问答评估和与人类协作的互动实验来验证这些表征的有效性，以确保我们能够有效地可视化这些表征。除了帮助训练计算机回答问题之外，我们还使用对手建模和强化学习来帮助训练人类更好地回答问题。

This CAREER project investigates how humans and computers can work together to answer questions. Humans and computers possess complementary skills: humans have extensive commonsense understanding of the world and greater facility with unconventional language, while computers can effortlessly memorize countless facts and retrieve them in an instant. This proposal helps machines understand who people, places, and characters are; how to communicate this information to humans; and how to allow humans and computers to collaborate in question answering using limited information. A key component of this proposal is answering questions word-by-word: this forces both humans and computers to answer questions using information as efficiently as possible. In addition to embedding these skills in question answering tasks, this proposal has an extensive outreach program to exhibit this technology in interactive question answering competitions for high school and college students.This research is possible by a new representations of entities in a medium-dimensional embedding that encodes relationships between entities (e.g., the representation of "Goodluck Jonathan" and "Nigeria" encodes that the former is the leader of the latter) to enable the system to answer questions about Nigeria. We validate the effectiveness of these representations both through traditional question answering evaluations and through interactive experiments with human collaboration to ensure that we can visualize these representations effectively. In addition to helping train computers to answer questions, we use opponent modeling and reinforcement learning to help train humans to better answer questions.