III: Small: Improving Technical Paper Database Search through Math-Aware Search Engines

今天的搜索引擎使用复杂的技术进行基于单词的搜索，但不能使用微妙的数学符号。该项目旨在允许科学家、工程师、数学家和学生使用文字、数学符号或两者中的一部分来定位技术信息。例如，学习图论的数学家可以使用这些新功能在物理、生态和社会网络分析中找到相关应用，尽管这些学科使用的符号和术语有任何差异。在给定大量技术文档的情况下，我们将应用机器学习技术来构建用于解释数学思想的公式和单词之间的关联，并确定如何在这两种表达形式之间自动翻译。然后，这些联想和翻译可以被学生使用，他们用单词写下他们想要的东西，搜索引擎会找到表达同样想法的文档，即使只是用数学符号。这些新的具有数学意识的搜索引擎将通过允许搜索者发现跨技术学科的信息，并通过使用数学符号作为支点，甚至跨人类语言来加速创新。为了实现这些目标，该项目将开发新的可扩展技术，用于索引和检索技术文档中的数学内容。这些方法将适应广泛的符号约定、格式和编码。将使用新的基于上下文的方法来推断公式与相关文本之间的关联，以建立丰富而灵活的内容等值模型。这些等价性模型将用于新的排名算法，该算法将使用单词或数学符号找到的结果整合到单个排名列表中。开源参考实现将被公开共享，为评估这些实现而创建的新测试集合将与其他研究人员共享。为了获得使用这些新功能的经验，该项目将在CiteSeerX科学文献数字图书馆中添加数学感知搜索。CiteSeerX是一种开放的Web服务，可用于比较实际使用的替代检索方法。有关更多信息，请参阅项目网页：https://www.cs.rit.edu/~dprl/math-aware-search.html.

Today's search engines make use of sophisticated techniques for searching based upon words, but are not able to make nuanced use of mathematical notation. This project aims to allow scientists, engineers, mathematicians, and students to locate technical information using words, mathematical notation, or some of each. For example, a mathematician studying graph theory could use these new capabilities to find related applications in physics, ecology, and social network analysis, despite any differences in the notation and terminology used in those disciplines. Given a large collection of technical documents, we will apply machine learning techniques to construct associations between the formulae and words used to explain mathematical ideas, and determine how to translate automatically between those two forms of expression. These associations and translations can then be used by students who write what they are looking for using words, with the search engine finding documents that express those same ideas, even if only in mathematical notation. These new math-aware search engines will accelerate innovation by allowing searchers to discover information both across technical disciplines and, by using mathematical notation as a pivot, even across human languages.To accomplish these goals, the project will develop novel scalable techniques for indexing and retrieval of mathematical content in technical documents. These methods will accommodate a broad range of notational conventions, formats, and encodings. New context-based methods for inferring associations between formulae and related text will be used to build rich and flexible models of content equivalence. These equivalence models will be used in new ranking algorithms that integrate results found using words or using mathematical notation into a single ranked list. Open-source reference implementations will be shared publicly, and new test collections created to evaluate these implementations will be shared with other researchers. To gain experience with the use of these new capabilities, the project will add math-aware search to the CiteSeerX digital library of scientific literature. CiteSeerX is an open Web service that can be used to compare alternative retrieval methods in actual use. For further information see the project Web page: https://www.cs.rit.edu/~dprl/math-aware-search.html.