CAREER: Utilizing Fine-Grained Knowledge Annotations in Text Understanding and Retrieval

我们信息社会的所有成员都不断地需要快速获取知识，以用于工作、教育和个人兴趣。通常，在一个网页或一篇文章中无法找到必要的知识，但需要一个更长的、针对具体问题的复杂答案。例如，政治事件的起因、巧克力的营养价值或风能的经济可行性等问题。然而，今天的搜索引擎，仅仅提供了一个列表的多个来源，但离开用户自己的综合来源的知识。该项目开发了新的算法，提取相关的关键概念，文本段落和关系信息，为用户提供综合信息的单一摘要。摘要分为不同的部分，每个部分涵盖一个复杂主题的不同方面。该项目的重点是确定相关的事实和联系，使用户能够形成自己的意见，并作出战略决策。嵌入在一个自我导向的学习环境中，它允许用户以自己的节奏学习新的主题。综合教育活动将利用该软件激发中学生和其他学科的本科生对STEM的兴趣。目前的算法不足以识别复杂主题的相关资源，因为这些主题需要有关世界的知识。利用知识图进行信息检索已经取得了一些重要的进展，例如实体查询特征扩展模型（EQFE）。该项目采用一种新颖的整体方法来开发知识图中的知识表示，相应的文本注释算法和携手工作的检索算法。该项目侧重于三个方面：（1）实体方面链接，确定实体提及的主题上下文，以促进高精度的段落排名。(2)利用关系抽取的开放领域的信息检索，其中存在许多不相关的关系是明确解决。(3)选择知识图的查询特定子图，其适于通过长距离依赖性来识别相关实体，同时避免概念漂移。所有这三个方面都导致了对实体的相关连接和方面的细粒度机器理解，通过为其相关性提供出处的支持段落进行对齐。所有这三个方面的影响都超出了信息检索的范围：许多建立在实体链接上的应用程序将从实体方面链接中获得主题精度。当在上下文中考虑关系信息时，大多数利用关系的技术都将得到改进。任何从知识图结构中提取信息的方法在消除虚假边缘时都会表现得更好。总的来说，这项研究工作将奠定基础，确定查询相关的复杂信息在自然大集合。这个奖项反映了NSF的法定使命，并已被认为是值得通过评估使用基金会的智力价值和更广泛的影响审查标准的支持。

All members of our information society are constantly in need for quick access to knowledge for work, education, and personal interests. Often the necessary knowledge cannot be found on a single web page or article but deserves a longer, query-specific complex answer. Examples are questions about causes of political events, about nutritional benefits of chocolate, or about the economic viability of wind energy. However, today's search engines that merely provide a list of multiple sources, but leave users on their own to synthesize sources into knowledge. This project develops novel algorithms that distill relevant key concepts, text passages, and relational information to provide users with a single summary of comprehensive information. The summary is structured into different sections, each covering a different facet of a complex topic. The focus of this project is to identify the relevant facts and connections that will enable users to form their own opinions and make strategic decisions. Embedded in a self-directed-learning environment, it allows users to learn about new topics at their own pace. Integrated educational activities will make use of the software to inspire STEM-interest among middle school students and undergrads of other disciplines.Today's algorithms fall short to identify relevant resources for complex topics, as these require knowledge about the world. Utilizing knowledge graphs for information retrieval has led to several important advances, such as the Entity Query Feature Expansion model (EQFE). This project takes a novel holistic approach towards developing representations of knowledge in a knowledge graph, corresponding text annotation algorithms, and retrieval algorithms that work hand-in-hand. The project focuses on three thrusts: (1) Entity aspect linking, which determines the topical context of entity mentions to facilitate a high-precision paragraph ranking. (2) Utilizing relation extractions for open domain information retrieval, where the presence of many non-relevant relations is explicitly addressed. (3) Selecting the query-specific subgraph of the knowledge graph that is suitable to identify relevant entities through long-range dependencies while avoiding concept drift. All three thrusts lead to fine-grained machine-understanding of relevant connections and aspects of entities, aligned through supporting passages that provide provenance for their relevance. The impact of all three thrusts extends beyond information retrieval: many applications that build on entity links will gain topical precision from entity aspect links. Most technology that utilizes relations as-is, will be improved when relational information is considered in context. Any method that extracts information from the knowledge graph structure will perform better when spurious edges are eliminated. Overall, this research effort will lay the foundation for identifying query-relevant complex information in natural large collections.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.