Financial data in enterprises is often stored using di erent data models, yet, it needs to be integrated in order to foster comprehensive evaluations. Conceptually, each of those data models can be understood as an ontology, and automated ontology matching can be applied as a rst step towards data integration. In this paper, we analyze the performance of existing ontology matching tools for matching nancial data models. The data has been provided by SAP SE and consists of real data schemas that are used in the nancial services area and mappings between them. We have created ve data sets by translating enterprise data schemas to ontologies and expert mappings to ontology alignment gold standards. We evaluate state of the art ontology matchers on our newly created data set. Our experiments show that current matching systems struggle to handle enterprise data sets and achieve signi cantly lower scores compared to data sets of other evaluation initiatives.
For nancial services enterprises, an understanding of the company's nancial
standing as well as its risk exposure is crucial for business decisions. Naturally,
there is an endogenous motivation to federate data. Additionally, regulators
emerge to be an exogenous driver for this process by obligating nancial
institutions to report risk KPIs in a timely manner and even by regulating the IT
infrastructure (like BCBS 2392 [
Studer et al. de ne an ontology as \a formal, explicit speci cation of a shared
conceptualization" [
Ontologies have already been used in enterprise settings before [
Approach
Concepts and data structures can be described using various notations and syntaxes. At SAP Financial Services, for example, data sources and data consumption layers are described, among others, by conceptual data models, physical data models, API documentation, or simply by SQL DDL statements. Depending on the notation abstraction chosen, the expressiveness varies. Ontologies can be used to describe data structures, since they are more expressive than the aforementioned notations.
In a rst step, the available data was collected, and data structures were
translated into ontologies using the Web Ontology Language (OWL). In a second
step, the known mappings were transformed into the alignment format as de ned
by the Alignmnet API [
To address the problem of heterogeneity of notations, all schemas were
transformed into ontologies by schema-speci c adapters. We have adopted the
approach for translating entity relationship models to ontologies introduced in [
The data has been provided for research by SAP SE. The SAP Financial Services Data Platform (FSDP) is a solution with the purpose to help nancial institutions with their data management. It includes a semantically rich conceptual data model (CDM) and a performance optimized physical data model (PDM) that can be deployed on a column-based database. As analytical (OLAP) and transactional (OLTP) applications run on the platform, inbound and outbound mappings are required. Data sources and consumers are mapped to the CDM. All mappings used here were manually created by multiple experts from the banking and insurance domain within SAP and map to the FSDP CDM.
The rst data set (D1) is derived from the mapping between the conceptual and the physical data model of FSDP. This is the largest data set. Because of performance improvements and implementation adaptions, the entities of the models are di erent. The second data set (D2) consists of a mapping between the FSDP CDM and a regulatory reporting application which brings its own data model. The third data set (D3) maps between the FSDP CDM and an SAP accounting solution. The fourth data set (D4) is a mapping between the FSDP business partner and the business partner of SAP ERP. The last data set (D5) maps between a loans management system and FSDP. Data sets D2, D3, and D5 are work in process and only the mapped structures were kept in the corresponding ontology. Data sets D1 and D4 are complete. Table 1 gives an overview over the data sets used. For a rst analysis, all OAEI 2018 matchers were ran on the data set. In addition, a simple string matcher3 has been used as baseline. The individual matcher performance is given in Table 2. For an overall statistic, macro average was chosen due to the di erent size and di culty of the data sets. Macro averages can
be found in Table 3. All statistics were calculated using the MELT framework4
[
Out of the matchers evaluated, only 5 matchers returned non-empty
alignments. Out of those, ALOD2Vec [
Early experiments indicate that current matchers struggle to match real world industry data schemas. Likely explanations are missing background knowledge, shallow and weakly structured ontologies, and potential over tting to publicly available benchmarks. In addition, most OAEI data sets and matchers focus on a 1-1 alignment arity while the data sets evaluated here are more complex. While the current prototypical set-up shows how ontology matching can be applied in a real enterprise setting, there are still many challenges that need to be addressed. The current data sets presented in this paper give a rst indication of the performance of current state of the art matchers on real nancial services data models. However, the data sets are yet small and incomplete. We plan to extend the current data base by increasing the amount of data and to improve
5 Note that all matchers in Table 2 could process each data set { i.e., the problems are rather semantic than technical. its quality. When the data base is grown to a more signi cant size and a high level of quality can be ensured, we consider o ering a blind alignment track at the OAEI. Since the results show that nancial services data models cannot be matched without background knowledge, future work will also focus on evaluating suitable sources of background knowledge, and on developing robust matchers that can handle loosely structured data schemas.
Acknowledgements. Acknowledgements go to Gaurav Sharma and Stephan Schub for helping compiling the mappings and overcoming technical obstacles.