Result of Ontology Alignment with RiMOM at OAEI'06

In this report, we briefly describe our system RiMOM and its underlying techniques. Given two ontologies, RiMOM intends to combine multiple strategies, aiming at finding the “optimal” alignments from the source ontology to the target one. RiMOM integrates multiple strategies: edit-distance based strategy, statistical-learning based strategy, and three similaritypropagation based strategies. Each strategy is defined based on one kind of ontological-information/approach. RiMOM conducts alignment finding as follows. It first estimates two factors respectively approximately representing the structure similarity and the label similarity of the two ontologies. The two factors are used in strategy selection to determine which strategies will be used in the alignment task. Then, we apply the selected strategies to find the alignment independently and combine the alignment results. Finally we employ the alignment refinement to prune “unbelievable” alignments. This report presents our results based on the evaluation. We also share our thoughts on the experiment design, showing specific strengths and weaknesses of our approach. 1. PRESENTATION OF THE SYSTEM Ontology alignment is the key point to reach interoperability over ontologies. In semantic web environment, ontologies are usually distributed and heterogeneous and it is necessary to find the mapping between them before processing across them. In recent years, much research work has been conducted for finding the alignment of ontologies [1] [4]. RiMOM is a tool for ontology alignment by combining different strategies, aiming at finding the “optimal” alignment results [5]. Each strategy is defined based on one kind of information or one type of approach. In our current version, there are five strategies defined: edit-distance based strategy, statistical-learning based strategy, and three similarity-propagation based strategies (including concept-to-concept propagation strategy (CCP), property-to-property propagation strategy (PPP), and concept-to-property propagation strategy (CPP)). 1.1 State, purpose, general statement We here define ontology alignment as a directional one. Given an alignment from ontology O1 to O2, we call ontology O1 as source ontology and O2 as target ontology. We call the process of finding the alignment from O1 to O2 as (Ontology) alignment discovery or alignment finding. Challenges for automating ontology alignment include: 1) how to automatically find alignments of high quality; 2) how to find the alignments efficiently; 3) how to make full use of the user interaction, since entirely automatic alignment is usually not possible; 4) how to automatically adjust the strategies for finding the alignments in a specific task, since the characteristics of the ontologies to be aligned are different in different tasks; 5) how to ease parameterizing, as the accuracy of alignments may vary largely with different parameters. In this campaign, we focus on dealing with the problems of 1), 2), and 4) with our system RiMOM. 1.2 Specific techniques used There are six major steps in the alignment process of RiMOM: 1) Similarity factors estimation. Given two ontologies, it estimates two similarity factors, which respectively approximately represent the structure similarity and the label similarity of the two ontologies. The two factors are used in the next step of strategy selection. 2) Strategy selection. The basic idea of strategy selection is if two ontologies have high label similarity factor, then RiMOM will rely more on linguistic based strategies; while if the two ontologies have high structure similarity factor, then we will employ similarity-propagation based strategies on them. See Section 1.2.2 for details. 3) Strategy execution. We employ the selected strategies to find the alignment independently. Each strategy outputs an alignment result. 4) Alignment combination. It combines the alignment results obtained by the selected strategies. The combination is conducted by a linear-interpolation method. 5) Similarity propagation. If the two ontologies have high structure similarity factor, RiMOM employs an algorithm called similarity propagation to refine the found alignments and to find new alignments that can not be found using the other strategies. Similarity propagation makes use of structure information. 6) Alignment refinement. It refines the alignment results from the previous steps. We defined several heuristic rules to remove the “unbelievable” alignments. 1.2.1 Multiple strategies The strategies defined in RiMOM can be classified into two categories: linguistic based strategies and structure based strategies. 1. Linguistic based strategies RiMOM contains two kinds of linguistic based strategies: edit-distance based strategy and statistical-learning based strategy. In our current version of RiMOM, for the statistical-learning based strategy, we use the classification method of K-Nearest Neighbor (KNN). For facilitating the description, we hereafter write the two strategies as ED and KNN. In ED, we calculate the edit distance between labels of two entities. In KNN, we formalize the problem of alignment as a problem of text classification. We view e2∈O2 as a class and its label, comment, and instances as a ‘document’ of the class. The text in a ‘document’ is tokenized into words. Then we employ stemming and stop words removing on the words and view the remains as features to train a text classification model. We also add some other general features which prove to be very helpful. For a concept, the features include: the number of its sub concepts, the number of properties it has, and the depth of the concept from “OWL:Thing”. For finding the alignment, we use the same method to generate a ‘document’ for a concept e1∈O1 and also add the general features as that in building the classification model. Then we use the trained classification model to identify which class the document should be classified. In this way, we are able to find which entity in O2 is the most possible one for an entity e1∈O1 to be aligned. The two strategies can be used for finding alignments independently. They can also be used together. In the latter case, we combine alignments of different strategies by: ( ) ( ) ( ) 1 2 1... 1 2