MaasMatch results for OAEI 2011

This paper summarizes the results of the first participation of MaasMatch in the Ontology Alignment Evaluation Initiative (OAEI) of 2011. We provide a brief description of the techniques that have been applied, with the emphasis being on the application of virtual documents and information retrieval techniques in order effectively utilize linguistic ontologies. Also, we discuss the results achieved in the tracks provided under the SEALS modality: benchmark, conference and anatomy. 1 Presentation of the system 1.1 State, purpose, general statement Sharing and reusing knowledge is an important aspect in modern information systems. Since multiple decades, researchers have been investigating methods that facilitate knowledge sharing in the corporate domain, allowing for instance the integration of external data into a company’s own knowledge system. Ontologies are at the center of this research, allowing the explicit definition of a knowledge domain. With the steady development of ontology languages, such as the current OWL language [3], knowledge domains can be modeled with an increasing amount of detail. Unfortunately, since ontologies of the same knowledge domain are commonly developed separately or for different purposes, transferring information across different sources becomes challenging as the heterogeneities between the ontologies need to be resolved. Several types of heterogeneities can emerge between two ontologies, commonly divided into syntactic, terminological, semantic and semiotic heterogeneities [1]. MaasMatch is an ontology matching tool with a focus on resolving terminological heterogeneities, such that entities with the same meaning but differing names and entities with the same name but different meanings are identified as such and matched accordingly. Given this focus, the tool has been primarily tested using the conference data set, since the ontologies of this data set are more likely to contain these heterogeneities. 1.2 Specific techniques used In this section we will present the techniques applied in MaasMatch. The overall structure of MaasMatch is simple, being a combination of a string similarity measure and our WordNet similarity, and using the combination of the two similarity matrices to extract the final alignments. However, most of our research so far has been invested into advancing the effectiveness of WordNet similarities. WordNet makes it possible identify concepts that have the same meaning but different names, since synonyms are grouped into sets, called synsets. However, a more challenging task is the identification of concepts have similar names, but different meanings. As an example, if an ontology contains a concept ’house’, then WordNet contains 14 different meanings for this word, and hence 14 different synsets that can be described by this name. One is thus faced with the challenge of automatically identifying the synset that denotes the correct meaning of the ontology entity. To do this, we applied a combination of information retrieval techniques and the creation of virtual documents in order to determine which synset most likely denotes the correct meaning of an entity. That way, only synsets which resulted in a high document similarity with their corresponding concept are subsequently used for the calculation of the WordNet similarity. The approach can be separated into 5 distinct steps as follows: Given two ontologies O1 and O2 that are to be matched, where O1 contains the sets of entities E x = {e1, e2, ..., em}, where x distinguishes between the set of classes, properties or instances, and O2 contains the sets of entities E y = {e1, e2, ..., en}, and where C(e) denotes a collection of synsets representing entity e, the main steps of our approach, performed separately for classes, properties and instances, can be described as follows: 1. Synset Gathering: For every entity e in E x, assemble the set C(e) with synsets that might denote the meaning of entity e. 2. Virtual Document Creation: For every entity e in E x, create a virtual document of e, and a virtual document for every synset in C(e). 3. Document Similarity: For every entity e in E x, calculate the document similarities between the virtual document denoting e and the different virtual documents originating from C(e). 4. Synset Selection: For every collection C(e), discard all synsets from C(e) that resulted in a low similarity score with the virtual document of e, using some selection procedure. 5. WordNet Similarity: Compute the WordNet similarity for all combinations of e ∈ E x and e ∈ E x using the processed collections C(e) and C(e). The first step of the procedure is fairly straightforward, where all corresponding synsets are collected if the complete name of an entity is present in WordNet and string processing techniques such as word stemming or finding legal sub-strings in the name are applied if the complete name is not present in WordNet. Figure 1 illustrates steps 2 5 of our approach for two arbitrary ontology entities e and e: Once the similarity matrix, meaning all pairwise similarities between the entities of both ontologies, are computed, the final alignment of the matching process can be extracted or the matrix can be combined with similarity matrices stemming from other approaches. Virtual Documents The second step of the approach consists of the creation of virtual documents for an ontology entity and several synsets that might denote the actual meaning of the entity. When constructing the virtual document, one must collect information from the ontology, or WordNet if a virtual document of a synset is constructed, in such a way that the resulting document adequately describes the meaning of the entity. An Fig. 1. Visualization of step 2-5 of the proposed approach for any entity e from ontology O1 and any entity e from ontology 2. expressive ontology such as OWL allows for the collection from various sources of information. In addition to its own name, an entity can also contain comments, which usually are written descriptions of the entity, and multiple labels. Providing context information is also beneficial. To do this, the names of the parent and child entities are also added to the document. Different details are added given the different types of entities. For virtual documents of classes, the names of all its properties are added and for properties the names of all the classes inside their range and domain are added. Once all the information for a virtual document is collected, several post-processing techniques such as word-stemming and stop-word removal are applied, before the document is transformed into the vector-space model. Using the document vectors, the similarity between the entity document and the different synset documents is then computed using the cosine similarity. Synset Selection Based on the document similarities between an entity and the potential synsets, some of the synsets are then discarded based on their similarity value. Several selection procedures have been tested, for instance using a cut-off value by computing the arithmetic or geometric mean of the similarities. Another tested method consisted of retaining only synsets whose document similarity had a higher value than the sum of the mean and standard deviation of the similarity value, which had the intriguing property that only few synsets remained if their similarity values were distinctly higher than the others, and more if this wasn’t the case and thus it was uncertain which synset was actually appropriate. Experimentation revealed that stricter selection methods performed better than lenient methods, with the simple method of using only the synset with the highest document similarity to compute the WordNet distance resulting in the highest scoring alignments [4]. 1.3 Adaptations made for the evaluation For experiments unrelated to the actual OAEI competition, but using some of the OAEI datasets, a cut-off confidence value of 0.7 has been used for the alignments, since this is one of the standard values that has previously been used for OAEI evaluations. For the purpose of OAEI participation, however, this value was altered to 0.95 to improve the final F-Measures achieved by the system, especially for the conference data set. 1.4 Link to the system and parameters file MaasMatch and its corresponding parameter file is available on the SEALS platform and can be downloaded at http://www.seals-project.eu/tool-services/browse-tools. 1.5 Link to the set of provided alignments (in align format) The computed alignments for the preliminary evaluation can be found at http://www.personeel.unimaas.nl/frederik-schadd/MaasMatchOAEI2011results.zip.