Servicenavigation

Detailseite

Zurück

Projekt Druckansicht

Identifikation semantischer Relationen zwischen Modellen auf unterschiedlichen Granularitätsstufen

Fachliche Zuordnung Bild- und Sprachverarbeitung, Computergraphik und Visualisierung, Human Computer Interaction, Ubiquitous und Wearable Computing
Förderung Förderung von 2014 bis 2017
Projektkennung Deutsche Forschungsgemeinschaft (DFG) - Projektnummer 257850598
 
Erstellungsjahr 2018

Zusammenfassung der Projektergebnisse

Within this project, we tried to develop, implement and evaluate methods that are specifically designed to create mappings between entities described at different levels of granularity. This refers to mappings that do not express equivalence between the matched elements, e.g., subsumption correspondences, or equivalence correspondences between an element and a complex that consists of several elements. We argued that this requires a modelling style that distinguishes between the layer of labels and the layer of logical entities (concepts, activities). We have continued to work on the idea first introduced in the proposal and finally developed a matching system called MAMBA. By participating in the OAEI 2015 we have proven that the approach outperforms the majority of alternative approaches. With a focus on process model matching, we have developed an approach for ensemble matching, which takes the mappings generated by different systems as input to generate a mapping as output, which is a subset of the ingoing matching hypothesis. This process needs to be controlled by a set of constraints, including cardinality constraints. We have focussed on a cardinality constraint that is still restrictive enough to have a significant impact, while it still supports the generation of subsumption correspondences. The overall model, which is in accordance to MAMBA, based on a Markov Logic formalization, achieved better results than each of the members of the ensemble. This shows that we were able to relax the original 1:1 constraint in the appropriate way. We have successfully developed two important approaches, one for process model matching and one of ontology matching. We evaluated both approaches via the PMMC and OAEI datasets and compared them against the results of these evaluation campaigns. The results show that our approaches outperform the majority of the current state of the art. However, we have not yet developed a generic matching system that combines both approaches. Such a combined solution would also require a more expressive label analysis which takes care of the specifics (process model vs. ontology), while the results of this analysis can be expressed within the same formalisation. Unfortunately, the results of WP 1 were not as good as expected. In particular, the attempts to use frames failed and the use of distributional semantics generated more noise than useful information. For that reason, we could not implement a component that creates a useful and expressive label representation. This is definitely a point that we have to work on in the future. With respect to scalability we detected an interesting way to guide the search for an optimal solution by first solving a relaxed, reversed version of the optimization problem. We implemented an automated generation of the relaxed problem which is very much tailored to the specifics of the matching domain. It is an open question, whether the approach can be used in a general way to speed up certain kind of optimization problems that suffer from a large search space which can be down‐sized significantly based on the solution of the relaxed problem.

Projektbezogene Publikationen (Auswahl)

 
 

Zusatzinformationen

© 2024 DFG Kontakt / Impressum / Datenschutz

Textvergrößerung und Kontrastanpassung