Using AgreementMaker to align ontologies for OAEI 2011

The AgreementMaker system is unique in that it features a powerful user interface, a flexible and extensible architecture, an integrated evaluation engine that relies on inherent quality measures, and semi-automatic and automatic methods. This paper describes the participation of AgreementMaker in the 2011 OAEI competition in four tracks: benchmarks, anatomy, conference, and instance matching. After its successful participation in 2009 and 2010, the goal in this year’s participation is to explore previously unused features of the ontologies in order to improve the matching results. Furthermore, the system should be able to automatically adapt to the matching task, choosing the best configuration for the given pair of ontologies. We believe that this year we have made considerable progress in both of these areas. 1 Presentation of the system We have been developing the AgreementMaker system since 2001, with a focus on realworld applications [5,9] and in particular on geospatial applications [4,6,8,10,11,12,13,15]. However, the current version of AgreementMaker, whose development started in 2008, represents a whole new effort. The code base has more than doubled since then, with the AgreementMaker framework being expanded to accomodate many types of ontology matching techniques. ? Partially supported by NSF Awards IIS–0513553 and IIS-0812258 and by the Intelligence Advanced Research Projects Activity (IARPA) via Air Force Research Laboratory (AFRL) contract number FA8650-10-C-7061. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of IARPA, AFRL, or the U.S. Government. 1.1 Purpose and state of the system The AgreementMaker system [1,2,3,7] is an extensible ontology matching framework that has been expanded to include many types of matching algorithms in order to handle many different matching scenarios. At the heart of the system is its ability to efficiently combine the results from several matching algorithms into one single and better result [2]. This capability allows us to focus on developing new matching algorithms and later combine them with our existing approach in order to improve our results. 2 Schema Matching Techniques Introduced in OAEI 2011 As compared to previous years, we have introduced several matching techniques in order to improve our matching algorithms. 2.1 Automatic Configuration Selection via Ontology Profiling Metrics The AgreementMaker system can be run with different configurations that optimize the system accuracy and coverage depending on the specific ontologies to be aligned. Changing the composition of the matcher stack (e.g. an instance matcher is used only when instances are available) has a high impact on the system performance. We developed an approach to adaptively optimize the configuration of AgreementMaker depending on the ontologies to be aligned. The approach we adopted can be sketched as follows: the ontologies to be aligned are profiled using several metrics proposed in the literature (e.g. relationship richness, inheritance richness, WorldNet coverage and so on [16] ). The metric-based profiles are used to automatically classify the matching task into a configuration class with specific settings. The classification is based on a supervised machine learning framework trained with a subset of the OEAI dataset for which a reference alignment is available. Our learning framework is very flexible: we can use many combinations of matchers and parameters, various types of classifiers (KStar, Naive Bayes, Multilayer Perceptron etc.) and new metrics. The experimental results show that the use of the automatic configuration methods improved the overall performance of AgreementMaker in the competition. In particular, in this paper we show the importance of this method for the significant improvements we achieved in the Benchmark and Conference tracks. The new AgreementMaker ’s matching process follows the steps represented in Figure 1: the ontology pair to be matched is classified by the ontology-profiling algorithm; based on the classification, a run configuration is created, and an ontology matching algorithm is instantiated to create an alignment. Fig. 1. AgreementMaker OAEI2011 Automatic configuration selection. 2.2 Lexicon Expansion via a Mediating Ontology Fig. 2. Using a mediating ontology. One approach to matching two domain specific ontologies is to use a third ontology from the same domain as a mediating ontology, with the mediating ontology to provide missing information relevant to the matching task. Shown in Figure 2, the source and target ontologies, OS and OT respectively, are first matched with the mediating ontology OM . Mappings between the source and target ontologies are then created based on the distance between the concepts in the mediating ontology to which they have been mapped previously. For the specific problem of matching the Mouse Anatomy ontology to the Human Anatomy ontology a successful approach has been to use the UBERON cross species anatomy ontology as a mediating ontology [14]. We have adapted this approach to our lexicon framework, using the BSM lex to match the MA and HA ontologies with UBERON thereby making use of the extra synonyms defined in UBERON. 2.3 Extension of Synonyms This strategy relies on synonyms defined in the OWL ontology itself, currently via the hasRelatedSynonym property, to generate a lexicon of synonym terms (single or multi-word terms). This is done by finding common terms between ontology synonyms to infer synonyms terms. For example, in the Human Anatomy ontology, the concept NCI C12275 (“Maxillary Sinus”) has the synonyms “Antrum, Maxillary” and “Sinus, Maxillary”. Our algorithm infers that “sinus” and “antrum” are synonyms as well without any external reference. These synonym terms are then used to create novel synonyms, by interchanging terms in existing synonyms and labels with their synonymous term. 2.4 Alternate Hierarchy Support In addition to the subclass hierarchy defined as part of the OWL ontologies of the Anatomy track, there is also a “part of” hierarchy defined using the UNDEFINED part of property. Taking into account this hierarchy in the VMM lex increases the percision and recall of the matching algorithm.