Robust and Scalable Knowledge Extraction from the Web

一种新的调查、解决问题和决策模式已经在我们的社会中变得普遍，包括应用计算、数学和统计模型从大量原始数据中推断出可操作的信息。这种范式，通常被称为“大数据分析”，正在推动我们的业务，改变我们的教育体系，并掌握着科学和工程领域下一个突破的关键。尽管人们对这一模式寄予厚望，但仍然存在许多挑战，使这一承诺无法成为现实。大多数组织中的绝大多数信息，无论是公共的还是私人的，都是以文本形式出现的（电子邮件、备忘录、报告、新闻文章、立法等），目前处理结构化数据的分析工具无法触及。至于与分析越来越相关的公开信息，情况也有问题。尽管存在在线编码和分享知识的适当标准，但网上的绝大多数内容仍然只使用表示标记（例如HTML）。再加上不断增加的内容量（包括组织内部和在线），从这些来源提取可操作信息的有效和高效工具的需求从未如此重要。这些技术的应用开始出现，例如Google的知识图，它被提供来补充常规Web搜索的结果。本研究的目标是推进国家的最先进的非传统信息资源，包括主要的文本源和半结构化的来源（在线知识库和数据库）的信息提取。我们的目标是：（1）从这些来源推断潜在结构;（2）找到来源内和来源之间的联系;（3）因此能够对这些来源进行数据分析。我们将首先利用我们在从文档中无监督提取关系以及对信息提取系统进行系统和大规模评估方面的工作，这些工作已经发现了一长串改进的机会，然后探索人工智能和自然语言处理中更深层次的模型，以提取和链接数据。

A new mode of inquiry, problem solving, and decision making has become pervasive in our society, consisting of applying computational, mathematical, and statistical models to infer actionable information from massive quantities of raw data. This paradigm, often called "Big Data Analytics", is driving our businesses, changing our educational systems, and holds the key to the next breakthroughs in science and engineering. While the hopes for this paradigm are understandably high, many challenges remain that prevent this promise to become a reality. The vast majority of information in most organizations, both public and private, comes in textual form (email, memos, reports, news articles, legislation, etc), remaining out of reach of current analytics tools which deal with structured data. As for publicly available information, which are increasingly more relevant for analytics, the situation is also problematic. Despite the existence of suitable standards for encoding and sharing knowledge online, the vast majority of content on the Web is still made available using only presentation markup (e.g. HTML). Coupled with the ever increasing volume of content (both internal to organizations and online), the need for effective and efficient tools for extracting actionable information from such sources has never been so critical. Applications of these techniques are starting to emerge, such as Google's knowledge graph which is being offered to complement the results of regular Web searches. The goal of this research is to advance the state-of-the-art in information extraction from non-traditional information resources, including both primarily textual sources and semistructured sources (online knowledge bases and databases). Our goals are: (1) to infer latent structure from these sources; (2) to find linkages within and across sources; (3) to enable data analytics over these sources as a result. We will start by leveraging our work on unsupervised extraction of relations from documents and on the systematic and large-scale evaluation of information extraction systems, which have uncovered a long list of opportunities for improvement, and then explore deeper models in AI and NLP for the purpose of extracting and linking data.