CAREER: Using Fiction to Improve Real-World Information Systems

在更高的层面上，这个项目旨在设计计算方法来推理小说世界，并反过来从小说中学习，为现实世界中的系统设计提供信息。虽然人工智能的许多工作都是从新闻和维基百科等相对较短的事实来源了解世界，但小说为改进现有信息系统和创新新的应用程序提供了一系列负担。与新闻等事实来源不同，小说捕捉情感、日常行为和常识，为引导知识库提供大量信息来源，这些知识库可以为问题回答系统、对话代理和下一代人工智能提供动力。该项目将提高自然语言理解在小说领域的表现，并利用它来探索两个案例研究：推断人们生活中日常事件的结构，包括宏观事件(如吃早餐)和低水平微观事件(如坐在桌子前，再倒一杯咖啡，把盘子放在水槽里)之间的关系；以及学习文本中所描述的观察到的行为与他们的代理人广泛覆盖的心理态度(如喜悦，悲伤，惊讶)之间的关系。这个项目旨在吸引社会科学和人文科学领域的学生和研究人员，他们在计算机领域的代表性历来偏低。虽然根据该项目开展的技术研究直接说明了社会科学和人文科学的专业知识如何为信息系统的计算设计提供信息，但该奖项下的初级教育计划将研究一个基本问题：如何使STEM领域以外的学生学习并提高他们在自然语言处理、机器学习和数据科学方面的技能。这项工作将使人文和社会科学的研究人员参与技术研究，向没有技术背景的学生传授技能，并将计算方法的进步转化为领域知识的进步。该项目的基础工作旨在通过提供两个案例研究，说明从小说中描绘的世界学习如何改进关于现实世界的推理系统，从而弥合计算与人文和社会科学之间的差距。这是一个新的前沿，不仅可以告诉我们当前文本蕴涵和情感分析系统的局限性，而且还可以在这个交叉点开辟新的研究领域。这项工作将在小说带来的两个任务上取得进展：推断普通动作的顺序和等级顺序，其中单个宏观事件由几个微观事件组成；以及推断文本中提到的人的潜在态度。这两个案例研究都将虚构作为知识的来源，并需要开发优化的计算模型，以弥合虚构和现实之间的差距。具体地说，这项工作将导致出版当代小说的新数据集，为实体和实体之间的相互参照(这有可能为该领域的嵌套实体识别和相互参照解析产生新的技术水平)，从小说中提取的日常动作知识库，用于对分层事件进行建模和从观察到的动作中学习心理态度的开源软件，以及在学术场所详细介绍在该项目范围内创建的方法的出版物。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

At a high level, this project aims to design computational methods to reason about the world of fiction, and, in turn, learn from fiction to inform the design of systems in the real world. While much work in artificial intelligence learns about the world from relatively short factual sources like news and Wikipedia, fiction offers a range of affordances for improving existing information systems and innovating new applications altogether. Unlike factual sources like news, fiction captures emotion, everyday action and commonsense, offering a vast source of information to bootstrap knowledge bases that can power question answering systems, conversational agents, and the next generation of artificial intelligence. This project will improve the performance of natural language understanding on fiction as a domain, and use it to explore two case studies: inferring the structure of everyday events in people's lives, including the relation between macro-level events (such as eating breakfast) and low-level micro-events (sitting down at the table, pouring another cup of coffee, putting the dishes in the sink); and learning the relationship between observed actions depicted in text and the broad-coverage mental attitude (such as joy, sadness, and surprise) of their agents. This project aims to draw in students and researchers in the social sciences and humanities, who have historically been underrepresented in computing. While the technical research carried out under this project directly speaks to how expertise in the social sciences and humanities can inform the computational design of information systems, the primary educational plan under this award will investigate one fundamental question: how to enable students outside STEM fields to learn and improve their skills in natural language processing, machine learning and data science. This work will engage researchers in the humanities and social sciences in technical research, teaching skills to students without technical backgrounds, and translating advances in computational methodology to advances in domain knowledge. The fundamental work in this project aims to bridge the gap between computation and the humanities and social sciences by providing two case studies of how learning from a depicted world in fiction can improve systems that reason about the real world. This is a new frontier that can not only teach us about the limitations of current systems for textual entailment and sentiment analysis, but can also open up new areas of research at this intersection. This work will make progress on two tasks enabled by fiction: inferring the sequential and hierarchical order of commonplace actions, in which a single macro-event is comprised of several micro-events, and inferring the latent attitudes of people mentioned in text given observations of their actions. Both case studies draw on fiction as a source of knowledge, and require the development of computational models optimized to bridge the gap between fiction and reality. Concretely, this work will result in the publication of a new dataset of contemporary fiction, labeled for entities and coreference between them (which has the potential to yield a new state of the art for nested entity recognition and coreference resolution for this domain), a knowledge base of everyday actions extracted from fiction, open-source software for modeling hierarchical events and learning mental attitudes from observed actions, and publications at academic venues detailing the methodologies created under the scope of this project.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.