CroMatcher results for OAEI 2015

CroMatcher is an ontology matching system based on parallel composition of basic ontology matchers. There are two fundamental parts of the system: first, automated weighted aggregation of correspondences produced by different basic matchers in the parallel composition; second, an iterative final alignment method. This is the second time CroMatcher has been involved in the OAEI campaign. Basic improvement with respect to the previous version has been implemented in order to speed up the system. 1 Presentation of the system CroMatcher is an automatic ontology matching system for discovering correspondences between entities of two different ontologies. This is the second version of the system. The first version [1] was presented in the OAEI campaign held in 2013. In this second version, the system architecture remained unchanged but the system implementation was modified as well as the implementation of several basic matchers in order to speed up the system. Our goal was to prepare the system for the following test sets: Benchmark, Anatomy, Conference and Large Biomedical Ontologies. The system is fully prepared for the Benchmark, Anatomy, and Conference. It is partly prepared for the Large Biomedical Ontologies (only for the 10% fragments of ontologies). We are currently working to speed up our system even more and we expect to present it in the next OAEI campaign. 1.1 State, purpose, general statement As stated before, the architecture of the new version of the system remained unchanged according to the first version [1] from 2013. To recapitulate, CroMatcher contains several terminological and structural matchers connected through sequential-parallel 1 Presently at Ericsson Nikola Tesla, the research was done while working at the University of Zagreb composition. First, the terminological basic matchers are executed. These matchers are connected through a parallel composition. After the execution of terminological matchers, the weighted aggregation is performed in order to determine the aggregated correspondence results of these matchers. These aggregated results are used in the execution of the structural matchers as initial values of entity correspondences. Structural matchers are also executed independently of each other in another parallel composition. Again, weighted aggregation is performed in order to determine the aggregated correspondence results of the structural matchers. Before the final alignment, the aggregated correspondence results of the terminological matchers and the aggregated correspondences’ results of the structural matchers need to be aggregated using weighted aggregation. Eventually, the method of the final alignment is executed. This method iteratively takes the best correspondences between two entities into the final alignment. 1.2 Specific techniques used In this section, only the modified components will be described in detail. The rest of the main components are described in the first version of the system [1]. We modified some terminological and structural matchers in order to speed up the matching process. These matchers are modified for the test sets Anatomy and Large Biomedical Ontologies because the ontologies in these test sets contain a large number of entities. Our matcher first counts the number of entities. If the ontologies contain more than 1000 entities than the modified versions of some matchers are activated instead of the original versions of matchers. Furthermore, we modified one terminological basic matcher in order to read entity information from components oboInOwl#hasRelatedSynonym and oboInOwl#hasDefinition. These components are implemented within ontologies of the Anatomy test set and contain considerable information about entities. The modified basic matchers are the following: 1. Terminological matchers:  Matcher that compares ID and annotation text of two entities (classes or properties) with the n-gram matcher [2] is extended in a way that also compares the text obtained from components oboInOw#hasRelatedSynonym and oboInOwl#hasDefinition. As stated before, these components are implemented within ontologies in the Anatomy test set. Our system first checks whether these components are implemented. If these components are not implemented within compared ontologies, the matcher compares only the ID and annotations like before.  Matcher that compares textual profiles of two entities with TF/IDF [3] and cosine similarity [4] is modified for the ontologies that contain more than 1000 entities in order to speed up the matching process. A textual profile is a large text that describes an entity (text obtained from annotations of compared entity and its all sub entities) therefore the matching was very slow because the TF/IDF method need to load the text of all entities before starting comparing two entities. When a target ontology contains more than 1000 entities, a modified implemented matcher is activated. This matcher compares textual profiles of two entities with the string metric described in [5]. This metric calculates similarity based on adjacent character pairs that are contained in both strings. This string metric is much faster than the TF/IDF method but the matching results are a bit worse than the results obtained with TF/IDF method. It is acceptable because the system performs the matching process faster enough to match ontologies with many entities.  Matcher that compares individuals of two entities by applying TF/IDF and cosine similarity is modified for the ontologies that contain more than 1000 entities. If the ontology contain more than 1000 entities, a modified implemented matcher with string metric described in [5] is activated like in the previous basic matcher.  Matcher that compares extra individuals of two entities with TF/IDF and cosine similarity is modified like two previous matchers in order to speed up the matching process. 2. Structural matchers:  All structural matchers described in the first version of our system [1] are executed iteratively. In order to speed up the matching process, we also made modification when comparing ontologies that contain more than 1000 entities. All structural matchers are executed just once (instead of being executed iteratively many times) when comparing the ontologies with more than 1000 entities. This speeds up the matching process but decreases the quality of matching process when comparing large ontologies. In the next version of the system, our major concern will be to solve the problem of slow iterative execution of structural matchers.