Reasoning over Large Amounts of Data in Ontologies via Abstraction and Refinement

基于本体的数据访问（OBDA）是知识表示和信息系统领域日益流行的一种范式。在此上下文中，本体是具有背景领域知识的TBox和包含有关应用程序领域元素的事实的ABox的组合。TBox用于丰富和集成大型的、不完整的、可能是半结构化的数据，然后用户可以通过查询访问这些数据。例如，维基百科的很大一部分以机器可处理的形式提供，它与本体TBox一起，是许多应用程序的重要信息源。为了有效地处理大型box， OBDA方法假设数据存储在数据库中。然而，通常在数据库中做出的完整数据假设（封闭世界假设）并不成立，回答查询需要推理。一种标准的推理方法是物化，即在系统接受查询之前，将所有需要的结果添加到ABox中。然而，对于大型箱子，物质化可能需要几个小时。在这个项目扩展中，我们建议采用一种新的方法来实现具体化，在这种方法中，我们不直接在（通常很大的）ABox上计算具体化，而是在数据的一个较小的“抽象”上工作。对于抽象，我们定义了一些标准，在这些标准下，来自ABox的个体被认为是等价的。这样难以区分的个体在抽象中只被表现一次。对于与ABox相比较小的tbox，抽象通常比原始ABox小得多，因此，所需的结果可以在主存中有效地计算。通过必然的结果，无法区分的个体可能变得可以区分。为了说明这一点，最初的抽象被迭代地细化，直到达到一个定点。到目前为止所获得的结果将在几个方向上得到扩展：1)开发的处理析取的技术将被扩展到更具表现力的本体语言（同时仍然保证健全性和完整性）。2)抽象的相关部分必须被细化，必须被识别和增量处理，以最小化与数据库后端的通信。3)基于增量改进，我们计划开发处理本体更新的技术。4)抽象方法似乎非常适合于改进本体调试过程，特别是对于通过直接从抽象中生成解释从文本中学习的大型abox。拟议的项目通过以一种新颖的方式将成熟的数据库技术与基于内存的推理技术相结合，支持对不断增长的结构化数据源的有效利用。

Ontology based data access (OBDA) is an increasingly popular paradigm in the area of knowledge representation and information systems. An ontology in this context is a combination of a TBox with background domain knowledge and an ABox, which contains facts about elements of the application domain. The TBox is used to enrich and integrate large, incomplete, and possibly semi-structured data, which users can then access via queries. For example, a large part of Wikipedia is available in machine-processable form, which, together with an ontological TBox, is an important information source for many applications. To efficiently handle large ABoxes, OBDA approaches assume that the data is stored in a database. Nevertheless, the assumption of complete data that is typically made in databases (closed world assumption) does not hold and reasoning is required to answer queries. A standard reasoning approach is materialization, i.e., all entailed consequences are added to the ABox before the system accepts queries. For large ABoxes, however, the materialization can take several hours.Within this project extension, we suggest a novel approach to materialization, where we do not compute the materialization directly on the (usually large) ABox, but where we work instead on a smaller ``abstraction'' of the data. For the abstraction, we define criteria under which individuals from the ABox are considered equivalent. Such indistinguishable individuals are then represented just once in the abstraction. For TBoxes that are small compared to the ABox, the abstraction is usually significantly smaller than the original ABox and, hence, the entailed consequences can be computed efficiently in main-memory. Through the entailed consequences individuals that were indistinguishable may become distinguishable. To account for that, the initial abstraction is iteratively refined until a fix-point is reached. The results obtained so far are to be extended in several directions: 1) The developed technique for handling disjunctions is to be extended to more expressive ontology languages (while still guaranteeing soundness and completeness). 2) Relevant parts of the abstraction that must be refined, are to be identified and incrementally treated in order to minimize the communication with the database backend. 3) Based on the incremental refinements we plan to develop techniques for handling updates to the ontology. 4) The abstraction approach seems well-suited for improving the ontology debugging process in particular for large ABoxes that are learned from text via the generation of explanations directly from the abstraction. The proposed project supports the efficient use of the ever growing sources of structured data by combining well-established database technologies with in-memory-based reasoning techniques in a novel way.