ServOMBI (ServO based Mapping with Binary Indexing) is an ontology matching system [1] which is designed by reusing the overall workflow followed by the ServOMap large scale ontology matching system [2] grounded on top of the ServO Ontology Repository (OR) system [3]. ServOMap is able to handle ontologies which contain several hundred of thousands entities. To deal with large ontologies, the system relies on terminological indexing strategy provided by the ServO OR to reduce the search space and computes an initial set of candidate mappings based on the terminological description of the entities of the input ontologies.
With ServOMBI, new components and variant algorithms have been introduced in this new version in regards to the ServOMap system. Among these new features we have : -
• a binary indexing strategy to complement the terminological indexing of ServOMap for optimizing ontology navigation,
• a modified contextual similarity computation thanks to the introduction of the binary indexing strategy during the Machine Learning (ML) step,
• a new ML algorithm during the contextual similarity
• the introduction of parallelization of some tasks for optimization purposes
• the selection of the final mappings using a variant of a stable marriage problem algorithm [8] Figure 1: Overall process of ServOMBI.
In the 2015 edition, ServOMBI participated in the entity level matching tasks apart from the Multifarm task. In the following sections, we described the main characteristics of the system and the overall results obtained during this year edition of OAEI.
ServOMBI has been built on the basis of the ServOMap system. ServOMap is designed with the purpose of facilitating interoperability between different systems which are based on heterogeneous knowledge organization systems (KOS). The heterogeneity of these KOS may have several causes ranging from the language format they use to the level of formalism of the terminology which describe the entities they involve. Our system relies on Information Retrieval (IR) techniques [4] and a dynamic description of entities of different KOS for computing the similarity between them.
ServOMBI implements new features and strategies and reuse some components of the ServOMap system.
The overall process followed by the ServOMBI system is depicted on Figure 1. The initialization phase is modified by introducing the Binary Ontology Indexing (BOI).
1. Ontology Loading: ServOMBI following the ServOMap approach relies on IR techniques for ontology matching. Each ontology to process is seen as a corpus of semantic documents to process. Each entity of the ontology is a document in the sense of IR. It is therefore necessary to identify the useful descriptors for indexing entities. The loading step perform the task of generating documents from entities. ServOMBI uses different reasoners (ELK [9], Hermit [11]) according to the size of the ontology to process.
2. Metadata and Metrics Generation: This step reuse the component implement in the ServOMap system and and identify 4 categories of matching tasks that are used to classify the input ontologies that are being processed : entity matching task (small, medium, big) and instances matching task. 3. Terminological Indexing: Following a generic metamodel provided by the ServO OR, a terminological based inverted index is built from the documents generated during the loading step. ServOMBI introduces a tasks parallelization using multithreading as the input ontologies terminology indexing could be done separately.
To complement the terminological indexing and to optimize the performance in terms of processing times due to the contextual similarity computation (surrounding concepts lookups), we have introduced a binary indexing strategy for the input ontologies. This technique of taxonomical representation consists of representing each concept of the ontology by a binary code, is inspired by the CEDAR system [5]. A binary code is a number of n bits, with n the number of concepts within the processed ontology. Thus, each concept has a code (a bit vector) carrying a "1" in the position corresponding to his index and the index of any other elements that it subsumes. These bit vectors must be encoded as the reflexive transitive closure of the "is-a" relation obtained from subsort declarations. Concepts are represented by a graph. Figure 2 gives an example of a binary representation of the extract of an ontology in the academic domain.
In this phase two main steps are performed: the selction of the final mappings and consistency checking. The selection of final mappings implement an algorithm of the stable marriage problem [8].
ServOMBI use the Lucene Apache IR library. Lucene provides functionalities for indexing and searching textual documents. The actual version of the matching system is based on the version 4 while the uploaded version for OAEI is based on the version 3.6. Their index format is slightly different. We have implemented the initial interactive matching [12] in ServOMBI using an oracle by modifying the validation process of the candidate mappings . This is performed after each round of candidate retrieving.
The wrapped SEALS client for ServOMBI version used for the OAEI 2015 edition is available at http://lesim.isped.u-bordeaux2.fr/servo/ServOMBI. The instructions for testing the tool is described in the tutorial dedicated to the SEALS client1 .
The results obtained by ServOMap during OAEI 2015 are available at http://lesim.isped.u-bordeaux2.fr/servo/ServOMBI/oaei2015.zip/.
We summarize in this section the results obtained by ServOMBI during the 2015 edition of OAEI.
The Anatomy track consists of finding an alignment between the Adult Mouse Anatomy and a part of the NCI Thesaurus (describing the human anatomy).
The results achieved by ServOMBI are summarized by Table 1.
The Conference track contains 16 ontologies from the same domain (conference organization). They have been developed within the OntoFarm project year the different tools are evaluated using i) crisp reference alignments where the confidence values for all matches are 1.0, ii) the uncertain version of the reference alignment where confidence values reflect the degree of agreement of a group of twenty people on the validity of the match [6] and iii) logical reasoning using violations of consistency and conservativity principles [15] [16]. Various reference alignments and evaluation modalities (R.A.M.) are used to assess the performance of the tooms. Thus, ra1 is the original reference alignment of the Conference track, ra2 is entailed reference alignment generated as a transitive closure computed on the original reference alignment (ra1) and rar2 is violation free version of reference alignment. Three different modalities are provided for these reference alignments, M1, M2 and M3 which contain respectively only classes, only properties and classes and properties. The results obtained by ServOMBI according to these different modalities on the crisp reference alignments where the confidence value is 1.0 are summarized on table 2. The value of β is respectively set to 0.5, 1 (harmonic measure) and 2.
The Large BioMed track consists of finding alignments between the Foundational Model of Anatomy (FMA), SNOMED CT, and the National Cancer Institute Thesaurus (NCI). The results obtained by ServOMBI for the small fragments of FMA-NCI task and FMA-SNOMED ontologies are summarize in Table 3 2
This track aims at offering a systematic and automated evaluation of matching systems with user interaction to compare the quality of interactive matching approaches in terms of F-measure and number of required interactions. For the 2015 edition, the Conference, Anatomy and Largebio tracks dataset are used for the evaluation. Moreover, this year a domain experts with variable error rates, respectively 0.1, 0.2 and 0.3 are considered in addition to the perferct emmulated user (oracle) with error rate 0.0. ServOMBI participated for the first year to this track. The interaction implemented currently in the system is mainly to allow the user validating the provided candidate mappings. Tables 4, 5 and 6 give respectively the results obtained by the system on the Anatomy, Conference and Largebio dataset for the Interactive track. We note that for the Largebio interactive track, the ServOMBI was only able to match the FMA-NCI small fragments and FMA-SNOMED small fragments. Overall ServOMBI improved its performance when compared to the results obtained with the normal Anatomy, Conference and Largebio track. However, the system make a greater number of requests compared to the other participating systems in the Interactive track.
This track does not follow the usual OAEI tasks for evaluating the performance of participating systems [14]. Precision and Recall are calculated with respect to the ability of the generated alignments to answer a set of queries in a ontologybased data access scenario where several ontologies exist. This track uses the Conference dataset for the evaluation with two reference alignments, the publicly available Conference track alignment (RA1) and the repaired one (RAR1).
We have participated in the 2012 and 2013 edition with the ServOMap system which achieved overall good results. The performance of this system is very good in particular for the tasks involving large ontologies. The new features implemented within ServOMBI did not lead to overall improved performances according to the results of the ServOMap system as expected. The contextual similarity computation, which is performed iteratively, is very time consuming and did not improved the overall recall of the system. In addition, while there is a gain in terms of computation times with concepts lookups, the BOI does not impact the overall performance of the system in terms of times taken to perform the matching tasks.
We have described in this paper the main functionalities of the ServOMBI ontology matching system and the overall results obtained during the 2015 OAEI edition. ServOMBI introduces a binary indexing strategy to complement the usual terminological indexing strategy used by the ServOMap system. The system achieved performance lower than expected according to the introduced features for the contextual similarity coputation. However it succed improving the F-Measure whith the interaction strategy. ServOMap continues to be developed in parralel and now include graph-based visualization.
As of future work, we envision to investigate an improved integration of the binary indexing and the contextual similarity computing. In addition, we plan to use combine multiple learning algorithms to improve the candidate selection during the contextual similarity computing.
This work has been partly supported by the project DRUG-SAFE funded ANSM (Agence Nationale de la Sécurité du Médicament). We also thank the organizers of OAEI with providing test dataset and the evaluation infrastructure.