Elements: Towards a Robust Cyberinfrastructure for NLP-based Search and Discoverability over Scientific Literature

该项目创建了一个开放的平台，用于从科学文本中访问和挖掘信息，提供对一系列软件、计算资源和出版数据的访问。目前的搜索技术通常会找到许多相关的文档，但没有对这些文档的信息内容进行提取和组织，也没有根据这些组织好的内容提出新的科学假设。自然语言处理（NLP）策略是解决这一问题的公认手段，该项目开发了网络基础设施，以支持从科学出版物中进行复杂的搜索和检索，使用和增强先进和完善的自然语言处理和机器学习工具的设施，以及从科学出版物中提取和汇总数据。该项目利用了两个nsf资助的项目：语言应用（LAPPS）网格，它已经被证明是开发NLP应用程序的有效平台；以及威斯康星大学的xDD（前身为GeoDeepDive），这是一个可扩展的、可靠的基础设施，能够快速增长科学出版物的数字图书馆，目前包括来自多个分布式商业和开放获取提供商的1300多万份文件。通过为挖掘科学出版物提供服务，并降低因许可、再分配和知识产权问题而导致的进入壁垒，这项工作显著提高了这些现有的nsf资助的基础设施的价值。科学家可以通过基于web的界面使用威斯康星大学的高性能计算（HPC）基础设施进行大规模的文本检索和挖掘。迭代的领域适应能力允许科学家轻松地将现有的服务适应到专门的领域，而无需配置或安装额外的组件。网络基础设施的潜在影响适用于任何依赖计算工具来挖掘大型文本数据集的社区，包括社会学、心理学、经济学、教育、语言学、数字媒体和人文学科的研究人员。该项目扩展了LAPPS网格，为UW- xdd的科学出版物和UW的高性能计算设施提供访问，以及快速适应现有的、完善的自然语言处理和机器学习软件工具到新领域和评估结果的手段。LAPPs网格提供了大量的NLP工具，这些工具来自各种各样的web服务来源，以及多个常用资源和一个前端文档检索引擎，该引擎目前配置用于访问PubMed/PubMedCentral以及CORD-19数据集的夜间更新。LAPPS Grid是开源的，可以从web、在用户的笔记本电脑或台式机上、在云中运行，也可以在需要保护敏感数据或许可数据时、在没有可用的网络连接时、或在远程HPC设施上部署时，作为自包含的docker映像运行。所有工具和资源都可以互操作地使用，从而消除了转换输入和输出格式以同时使用一组工具或资源所需的工作。xDD是世界上最大的科学出版物单一存储库之一，它涵盖所有知识领域，自动合并新文档，并每小时更新API端点。xDD已经积累了来自多个商业和开放访问出版商的数百万个文档（超过1300万个出版物）。xDD基础设施是正在开发的UW-COSMOS管道的一个组成部分，该管道由一套支持文档处理的服务组成，包括pdf的摄取和解析；提取单个文档对象，如文本部分、图形、表格和标题；和recall，它在上下文和对象上创建可搜索的Anserini和ElasticSearch索引，以便检索信息。具体的项目活动包括有效地检索和分析xDD的大量科学出版物；扩展LAPPS网格在科学出版物挖掘和领域适应方面的NLP能力；开发网格和xDD/COSMOS之间的完全互操作性；扩展LAPPS网格服务，以处理来自UW-xDD的非常大的文本数据集；以及测量可视化技术并将其整合到网格中。该奖项由先进网络基础设施办公室颁发，由计算机和信息科学与工程理事会下属的NSF信息和智能系统部门以及NSF公共访问计划共同支持。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This project creates an open platform for accessing and mining information from scientific texts that provides access to an array of software, computing resources, and publication data. Current search technologies typically find many relevant documents, but do not extract and organize the information content of these documents or suggest new scientific hypotheses based on this organized content. Natural Language Processing (NLP) strategies are a recognized means to approach this problem, and this project develops the cyberinfrastructure to support sophisticated search and retrieval from scientific publications, use and augmentation of facilities for advanced and well-established natural language processing and machine learning tools, and extraction and aggregation of data from scientific publications. The project leverages two NSF-funded projects: the Language Applications (LAPPS) Grid, which has already proven to be an effective platform for development of NLP applications; and University of Wisconsin’s xDD (formerly, GeoDeepDive), a scalable, dependable infrastructure capable of rapidly growing a digital library of scientific publications, currently including over 13 million documents from multiple distributed commercial and open-access providers. The effort significantly enhances the value of these existing NSF-funded infrastructures by providing access to services for mining scientific publications and lowering the barriers to entry resulting from licensing, redistribution, and intellectual property issues. Scientists may perform large-scale text retrieval and mining using the University of Wisconsin’s high performance computing (HPC) infrastructure through a web-based interface. Iterative domain adaptation capabilities allow scientists to easily adapt existing services to specialized areas without configuring or installing additional components. The potential impact of the cyberinfrastructure is applicable to any community that relies on computational tools for mining large textual datasets, including researchers in sociology, psychology, economics, education, linguistics, digital media, and the humanities.This project extends the LAPPS Grid to provide access to UW-xDD’s collection of scientific publications and UW’s High Performance Computing facilities, as well as means to rapidly adapt existing, well-established natural language processing and machine learning software tools to new domains and evaluate results. The LAPPs Grid provides a large collection of NLP tools from a wide variety of sources exposed as web services, together with multiple commonly used resources and a front-end document retrieval engine currently configured to access PubMed/PubMedCentral as well as nightly updates of the CORD-19 dataset. The LAPPS Grid is open source, and can be run from the web, on a user’s laptop or desktop, in the cloud, or as a self-contained docker image when it is necessary to protect sensitive or licensed data, when there is no network connection available, or for deployment on remote HPC facilities. All tools and resources can be used interoperably, eliminating the effort required to convert input and output formats to use a set of tools or resources together. xDD is one of the world’s largest single repositories of scientific publications that spans all domains of knowledge, incorporates new documents automatically and updates API endpoints every hour. xDD has accumulated millions of documents from multiple commercial and open-access publishers (over 13M publications). The xDD infrastructure is an integral part of the developing UW-COSMOS pipeline, which consists of a suite of services supporting document processing, including ingestion and parsing of PDFs; extraction of individual document objects such as text sections, figures, tables, and captions; and recall, which creates searchable Anserini and ElasticSearch indexes on the contexts and objects to enable retrieval of information. Specific project activities include implementing efficient retrieval and analysis of xDD’s vast holdings of scientific publications; extending the NLP capabilities of the LAPPS Grid for scientific publication mining and domain adaptation; developing full interoperability between the Grid and xDD/COSMOS; scaling LAPPS Grid services to handle the very large textual datasets available from UW-xDD; and surveying visualization techniques and integrating them into the Grid.This award by the Office of Advanced Cyberinfrastructure is jointly supported by the NSF Division of Information and Intelligent Systems within the Directorate for Computer and Information Science and Engineering, and the NSF Public Access program.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.