This paper summarizes the results of PRIOR system, which is an ontology mapping system based on Profile pRopagation and InfOrmation Retrieval techniques, for OAEI 2006 campaign. The PRIOR system exploits both linguistic and structural information to map small ontologies, and integrates Indri search engine to process large ontologies. The preliminary results of the experiments for four tasks (i.e. benchmark, web directories, anatomy and food) are presented. A discussion of the results and future work are given at the end.
1.1
The World Wide Web (WWW) makes a large number of digital resources publicly accessible. However, finding relevant information, i.e. searching for digital resources from various sources and manually organizing them for relevance, becomes more and more intractable. Semantic interoperability research is aimed at enabling different information systems to communicate information consistently with the intended meaning. Ontology mapping is one critical mechanism to achieve semantic interoperability.
Different communities have proposed different approaches to ontology mapping. The
techniques that have been applied to solve mapping problems include linguistic
analysis of terms [5][
Our approach begins with the belief that the combination of linguistic analysis and
graph theory will lead to successful mapping. It explores information from two
perspectives, linguistic and structural, to determine the correspondences that identify
similar elements in different ontologies. In an ontology, linguistic information is the
descriptive information, such as name (i.e. ID), label, comment and property
restriction, of a concept (i.e. class, individual and property). Structural information
refers to relationships between concepts in the ontology. Such relationships include
hierarchy relation, inverse relation and so on. Since the filed of information retrieval
is highly relevant to ontology mapping, we also explore using classic information
retrieval method to support the mapping of large ontologies. Figure 1 depicts the
architecture of PRIOR system. The details of the approach are explained in next
section.
We introduce the term “profile”. Similar to the virtual document used in Falcon-AO
system [
The Profile Propagation exploits the neighboring information of each concept. That is, we pass the profile of the ancestors, children or siblings of the concept to the profile of the concept itself. The reason why we do profile propagation is based on the observation that if we see the taxonomic tree of an ontology as the index of a book, the super class in the ontology reflects the “context” of its subclasses and each subclass is the “content” of its super class. The process of profile propagation can be VNnew = ∑ w( N , N ')VN '
N '∈S represented as: , where N and N’ represent two concepts in the ontologies, S represents the set of all concepts in the ontologies, VNnew represents the new profile vector of the concept N, VN’ represents the profile vector of the concept N’, and w(N, N’) is a function that assigns different weights to the neighbors of the concept according to the distance between them. Two principles to assign the weight are applied: 1) The closer the two concepts are, the higher weight will be assigned, i.e. the weight of a parent is higher than the weight of a grandparents and the weight of a child is higher than the weight of a grandchild. 2) The weight of a parent is higher than the weight of a child and the weight of a child is higher than the weight of a sibling. This is because children inherit all characteristics of the parent and may extend some characteristics that parent doe not have, and sibling is usually a complementary of the concept.
For small ontologies, the Profile Mapper compares each concept of the ontologies by computing cosine similarity of the profile of each concept. Simultaneously, the String Mapper computes the similarity between the names of different concepts using Levenshtein distance. The profile similarity and the name string similarity are further integrated to obtain final similarities between concepts. However, if the ontology is too large, calculating the similarity matrix will require too many computing resources and it is time consuming. Based on the understanding that ontology mapping is also an information retrieval task, we turn to classic information retrieval method to solve the problem. Specifically, we integrated indri1 search engine into PRIOR system. First, the Indri Mapper uses Indri to index profiles of concepts in ontology A. Then queries are generated based on the profiles of the concepts in ontology B. After storing the top-ranked results returned by the queries, we switch two ontologies, i.e. this time ontology B is indexed and queries are generated based on ontology A. The Indri Mapper will pass two sets of search results to the Mapping Extractor. Having the similarity matrix obtained from small ontologies or Indri search results from large ontologies, the Mapping Extractor extracts all candidates of matched concepts and output the results in desired format. 1.3
We didn’t do any major adaptations in order to align the OAEI campaign ontologies. However, for food test, we treat <skos:broader> and <skos:narrower> as parent and child relations. 1.4
The system is available at: http://www.sis.pitt.edu/~mingmao/om06/ 1.5
The result file can http://www.sis.pitt.edu/~mingmao/om06/result.zip be downloaded from 2
In this section we present the results of alignment experiments on OAEI 2006 campaign. All tests are run on a stand-alone PC running Fedora 4 operating system. 1 http://www.lemurproject.org/indri The PC has Pentium 4, 3.0GHz processor, 1G memory, 100GB Serial ATA hard disk and SUN JAVA VM 1.5.0_06. 2.1
benchmark The benchmark tests can be divided into two types. Test 101-266 are systematically generated from reference ontology, in which some information are discarded, and test 301-304 are real bibliographic ontologies. Since our approach is relied on the linguistic information, we obtain high precision and recall where the test ontologies contain the same names (or name conventions) and/or comments as the reference ontology (i.e. test 101, 103, 104, 203, 204, 208, 221-247). However our approach fails in the recall where both name and comments are replaced or missing in the test ontologies (i.e. test 202, 248-266). For tests 201, 206-207 and 210, though the class name has been “removed” or expressed in another language, we can find some matched classed and properties due to the information of comments and instances. For tests 205 and 209 having name synonyms, the performance of our approach is not good because we do not use thesaurus. For real ontologies 301-304, they cover the same domain as reference ontology using similar descriptive information and different structural information. The result of these real tests shows the average performance of our approach is around 80%. The full result of all tests can be found in Appendix. 2.2
directory The directory real world case consists of aligning web sites directory. It has 4640 elementary tests. Each of them is represented by pairs of OWL ontologies, where classification relation is modeled as OWL subClassOf. Therefore all OWL ontologies are taxonomies, i.e., they contain only classes (without Object and Data properties) connected with subclass relation. We use the same set of parameters and approach as those of benchmark test to obtain alignment results. 2.3
anatomy 2.3
food The anatomy task is to find alignment between classes in two medical ontologies, FMA ontology and OpenGALEN ontology. FMA has 72559 classes and OpenGALEN has 9564 classes. Due to the huge size of the ontologies, we use Indri approach. Finally 2583 pairs of candidates have been found within 9 minutes. The food thesaurus mapping task requires to create alignment between the SKOS version of the United Nations Food and Agriculture Organization (FAO) AGROVOC thesaurus, which has around 16000 terms and is expressed in multilingual, and the United States National Agricultural Library (NAL) Agricultural thesaurus, which has around 41000 terms and is expressed in monolingual. AGROVOC has 28179 concepts, and NAL has 41594 concepts. Due to the similar reason as anatomy task that the size of food thesaurus is too large, we use Indri approach. Finally 11511 pairs of candidates have been found within 73 minutes. Although “narrowMatch” and “broadMatch” are allowed, we can only get “exactMatch”. 3
3.1
Since our approach relies on linguistic information such as name, label, comment, and other descriptive information, it can not handle pure graph matching task, like test 248-266 in benchmarks. Also we do not use external resources like WorldNet to process synonyms, which we believe is important in real cases. Furthermore, some ontology like AGROVOC contains labels in foreign languages; currently we do not use this type of information.
We use Alignment API to parse ontologies and generate alignments. When processing FMA ontology in anatomy test, the API reads each owl:Class as a class first and then as an individual one more time. In all properties of a class, only “ID” and “label” are assigned to the class, all other properties such as “part” and “constitutional_part” are assigned to the individual. Since only classes are alignment candidates, we miss all information in individual. 3.2
One possible improvement is to integrate external resources to increase recall. For instance, WordNet can be integrated to process synonyms and dictionaries can be used to process foreign languages. Another possible improvement is to find out a better way to adjust the propagation weights. It’s possible to train the weights with some training data. 3.3
The ontologies in anatomy and food tests are very large and in a different format (i.e. SKOS, Protégé exported RDF) other than benchmark tests. It will be better to have a small part of ontology as training ontology, for which alignments are provided to participants. So that participants can train their approach on this training ontology. We also would like to see the OAEI 2006 campaign to be the first one to provide reference alignment for real word large scale ontologies so that different approaches can be judged in systematic way. 3.4
Considering the mapping relations in food track, the evaluation process is more complex. If concept A is an “exactMatch” to concept B, and concept C is a “broader” concept of B, then we can say concept A and C has a “broadMatch” relation. First we don’t know whether A-exactMatch-B and A-broadMatch-C will both appear in reference alignment. Second, if they both appear in reference alignment, but only AexactMatch-B mapping is in an answer alignment, how do we calculate recall regarding A-broadMatch-C mapping? 4
In this paper, we briefly present a system for ontology mapping – PRIOR system, in which we explore linguistic and structural information and profile propagation method to process small ontologies. We also integrate classic information retrieval method to process large ontologies. The preliminary results are carefully analyzed and some future work are discussed.
Matrix of results
0.5 0.5 0.92 0.86 0.68 0.95