Matching Representations at different Levels of Granularity

点击翻译按钮获取中文摘要

Conceptual models of information structures and information flows are a central concept in computer science. They play a crucial role in the design and maintenance of information systems and the task of identifying mappings between different models as a basis for integrating different systems has become more and more important.Automatically identifying semantically correct mappings between models that describe a domain at different degrees of granularity pose some problems that cannot be adequately handled by existing matching approaches: (1) The mapping can be partial, that means that only some elements from one model actually do have counterparts in the other model and (2) the mapping can be n to m, meaning that one element in the first model can correspond to a combination of elements in the second model and vice versa. Existing methods for complex matching either require a high number of models as basis for correlation statistics or provide heuristic solution that only apply in a very restricted setting. Optimization-based matching algorithms that try to maximize the similarity between mapped elements from two models have proven to be the method of choice for one-to-one matching, due to their benefits over purely heuristic methods. Despite this fact, there have been no attempts so far to extend these approaches to the problem of complex matching. The goal of the project is (1) to develop new optimization-based methods for solving complex matching problems, i.e., matching problems that require the detection of n-to-m correspondences between elements in the models to be matched and (2) to show the practical benefit of the methods by applying them to the problem of matching real-world conceptual models, in particular ontologies and process models.We will approach this problem by combining a deep linguistic analysis of complex labels found in conceptual models with the idea of computing one-to-one correspondences not at the level of model elements, but on the level of meaningful linguistic components found in the labels of elements. We then generate complex mappings between model elements from the resulting mappings between parts of element labels. In the course of the project, we will develop and implement linguistic methods for analyzing complex class labels in ontologies and extend exsting work for analyzing activity labels in process models. We will further investigate efficient and effective formulations of complex matching task as an optimization problem and efficient and scalable methods for computing solutions taking our previous work on computing most probably consistent models as a starting point. Finally, we will spend significant effort on creating benchmark data for complex ontology matching and making it available to the scientific community.