CRII: III: Towards Reasoning Augmented Searching for Domain-Specific Knowledge Screening

在大数据时代，对于领域专家（如医生和制药科学家）来说，从不同的数据库中有效检索相关信息以支持其决策越来越具有挑战性。基于传统关系数据库的知识筛选，包括搜索和过滤，可能会遇到几个问题。例如，基于关键字的搜索和基于规则的过滤只能处理有限类型的问题，缺乏灵活性。在过去的几年里，现代深度学习方法通过自动将自然语言问题翻译成结构化查询语言来缓解这个问题。然而，如果将不同类型的数据（如表格、文本和异构图形数据）整合到具有许多表的关系数据库中，结构化查询可能会变得非常复杂，这对深度学习模型提出了新的挑战。该项目将通过设计一种基于非关系数据库的新模式来应对这些挑战，该数据库以更灵活的非表格形式存储数据。新范式可以轻松地将推理纳入问题到查询的翻译中，使深度学习模型能够处理更复杂的问题，这将有利于许多特定领域的应用程序。该项目还将促进教学和辅导活动，例如开发新课程和培训机器学习、自然语言处理、数据管理和健康信息学方面的下一代专家。该项目的成果和观察结果将开放供公众使用，该项目将为非关系数据库上自然语言驱动的知识筛选开辟新的研究方向。虽然有许多著名的，高效的，可扩展的非关系数据库和搜索引擎，很少有人致力于开发自然语言查询方法，并利用它们的潜力。该项目旨在通过设计新的基于自然语言的搜索和查询的基础框架来填补这一空白，包括非关系数据库中的数据整合，数据库和查询模板中的推理集成，以及人在环模型开发和评估。基于流行的搜索引擎ElasticSearch，将开展两项主要研究活动：（1）研究人员将开发新的深度学习模型，用于将自然语言问题转换为ElasticSearch查询，并创建新的数据集用于训练和评估模型。(2)研究人员将提出一种统一的方法，标准格式和可扩展的方式来创建知识“金块”来存储多模态数据，并开发新的问题生成模型来自动从嵌套的知识中生成问题。该项目将产生各种成果，例如用于模型开发的数据，用于模型训练和推理的算法，以及用于创建训练数据的注释工具。这些产品将有利于数据管理和筛选，并支持医疗保健，生物信息学和科学研究的决策。该奖项反映了NSF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。

In the era of big data, it is increasingly challenging for domain experts, such as physicians and pharmaceutical scientists, to efficiently retrieve relevant information from different databases to support their decision-making. Knowledge screening, including searching and filtering, based on traditional relational databases may suffer from several problems. For example, keyword-based searching and rule-based filtering can only handle limited types of questions and lack flexibility. Over the past few years, modern deep learning methods have alleviated this problem by automatically translating natural language questions into structured query languages. However, if different types of data such as tabular, textual, and heterogeneous graph data are consolidated into a relational database with many tables, structured queries can become very complex, which poses new challenges to deep learning models. This project will tackle these challenges by designing a new paradigm based on non-relational databases that store data in a more flexible non-tabular form. The new paradigm can easily incorporate reasoning into question-to-query translation, enabling deep learning models to handle more complex questions, which will benefit many domain-specific applications. The project will also promote teaching and mentoring activities, such as developing new courses and training of next generation experts in machine learning, natural language processing, data management, and health informatics. The project outcomes and observations will be open for public use.The project will forge a new research direction for natural language-driven knowledge screening on non-relational databases. Although there are many well-known, efficient, and scalable non-relational databases and search engines, little effort has been devoted to developing natural language querying methods for them and exploiting their potential. This project aims to fill this gap by designing new underlying frameworks for natural language-based searching and querying, including data consolidation in non-relational databases, reasoning integration in both databases and query templates, and human-in-the-loop model development and evaluations. Two primary research activities will be undertaken based on a popular search engine known as ElasticSearch: (1) The investigator will develop new deep learning models for translating natural language questions into ElasticSearch queries and create new datasets for training and evaluating the models. (2) The investigator will propose a unified approach, standard format, and extensible way to create knowledge “nuggets” to store multi-modal data and develop new question-generation models to automatically generate questions from nested knowledge. The project will produce a variety of outcomes, such as data used for model development, algorithms for model training and inference, and annotation tools used for creating training data. These products will benefit data management and screening, and support decision-making in healthcare, bioinformatics, and scientific research.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.