In this paper we give a brief explanation of how Anchor Concept and Sub-Ontology based Ontology M atching ( SOBOM ) gets the alignment results at OAEI2009. SOBOM deal with the ontology from two different views: an ontology with is-a hierarchical structure O' and an ontology with other relationship s O' ' . Firstly , from the O' view, SOBOM starts with a set of anchor concepts provided by linguistic matcher. And then it extracts subontologies based on the anchor concepts and ranks these sub-ontologies according to their depth. Secondly , SOBOM utilizes Semantic Inductive Similarity Flooding algorithm to compute the similarity of the concepts between the sub-ontologies derived from the two ontologies according the depth of sub-ontologies to get concept alignments. Finally , from the O' ' view, SOBOM gets relationship alignments by using the concept alignment results in O' ' . The experiment results show SOBOM can find more alignment results than other compared relevant methods with high degree of p recision.
Currently mo re and more ontologies are distributedly built and used by different
organizations. And these ontologies are usually light-weighted [
SOBOM is an automat ic ontology matching tool. There are three matchers
implemented in current version: linguistic matcher I-Sub [
O
O '
O '' View from O'' Step3:Matching Relationships 3.1 R-Matcher to match other relationships according SISF results 3.2 Combine the concept alignments with relationship alignments
Step4:Output alignments 1. 1:1 mapping between Classes <c1,c2,0.9,=> 2. 1:1 mapping between Relationships <r1,r2,1.0,=> For simp licity, we define some notations used in the report.
Ontolog y: An ontology O consists of a set of concepts C , properties/relations R , instances I , and A xio ms AO . We use entity e to denote either c C or r R . Each re lation r has a domain and range defined as follo wing: Domain(r) {ci | ci C and having the relationship r}
Range(r) {ci | ci C and can be value of r} Anchor c once pt: an anchor concept is the strongest semantic similarity between two entities in different ontologies. It is a pair of concepts fro m two ontology: a c1, c2 ，where sim(c1, c2 ) Sub-Ontology: Sub_ O is a concept hierarchy with a.ci as root，it satisfied that c Sub_ O , c O and c is a descendant of a.ci , Sub _ O O 。 1.2
SOBOM a ims to provide high quality of 1:1 a lignments between concept and
property pairs. We imp le mented SOBOM a lgorith m in java and had integrated three
distinguishing constitutional matchers, I-Sub, SISF and R-matcher. They are regarded
as independent components in core matcher library o f SOBOM. Due to the space
limitat ion, we only describe the key features of them. The details can be found in the
related paper [
I-Sub is a light-we ighted matcher simp ly based on the string comparison techniques. The innovation of I-Sub is not only the commonalities between the descriptions of doma in entities are calcu lated but also their diffe rences are e xa mined. Fu rthermore, it is stable to sma ll d iverges fro m the optima l threshold taking place and intelligent to identify all the differences between strings. In SOBOM, I-Sub is a core co mponent to generate anchor concepts. SISF uses the RDF statement to represent the ontology and utilizes the results of I-Sub to inducting the construction of similarity propagation graph from sub-ontologies. SISF and I-Sub handle the ontology fro m the v iew O' and only generate concept-concept alignment. R-matcher is a re lationship matcher base on the definition of the ontology. It combines the linguistic and semantic in formation of a re lation. Fro m the O' ' view, it utilizes the is -a hie rarchy to e xtend the do main and range of a relationship and uses the result of SISF to generate the align ment between relationships.
More importantly, SoE is integrated into the SOBOM and extract sub-ontologies
according to the anchor concept [
Rule 1： Upwards traversal of the hierarchy: c' O , if c' is an ancestor of a.ci , then c' Sub _ O .
Rule 2： Siblings classes of anchor concepts： c' O , if c' is a sibling concept of a.ci , then c' Sub _ O .
Rule 3： Downwards traversal of the hierarchy： c' O ， if c' is descendant concepts of a.ci , c' Sub _ O Rule 4： Other relat ionships of the anchor concepts： r O , if r is a relationship in O and r is _ a , then r Sub_ O Rule 5： Leaf Concept Nodes: if a.c1 O1 , a.c2 O2 and a.c1 , a.c2 are lea f nodes respectively in O1,O2 , then don’t extract Sub-Ontology. After e xt racting sub-ontologies, SOBOM will match these sub -ontologies according to their depth in original ontology. We first match the sub-ontologies with larger depth value. By using SoE, SOBOM can reduce the scale of ontology and make it easy to operate sub-ontologies in SISF. 1.3
We don’t make any specific adaptation for the tests in the OAEI 200 9 campaign. A ll the align ments outputted by SOBOM are based on the same set of parameters. 1.4
The current version of SOBOM and the align ment results for OA EI 2009 are available at http://mlg.hit.edu.cn:8080/Ontology/Download.jsp, and the parameters setting is illustrated in the reading me file. 2
In this section, we describe the results of SOBOM algorith m against the benchmark, directory and anatomy ontologies provided by the OAEI 2009 campaign. We use Jena-API to parse the RDF and OWL files. The experiments were carried out on a PC running Windows vista ultimate (32 bit) with Core 2 Duo processors (2.66 GHz) and 4-gigabyte memory. 2.1
On the basis of the nature, we can div ide the benchmark dataset into five groups: #101-104, # 201-210, #221-247, #248-266 and #301-304. We described the performance of our SOBOM algorith m over each group and overall perfo rmance on the benchmark test set in Table 1.
#101-104 SOBOM plays well for these test cases.
#201-210 In this group, some linguistic features of candidate ontologies are discarded or modified. SOBOM is a sequential matcher, if the linguistic matcher get no mappings, then the SISF will produce no mapping too. So in these test, the result is in high precision but low recall.
#221-247 The structures of the candidate ontologies are altered in these tests. However, SOBOM discovers most of the a lignments fro m the linguistic perspective via our linguistic matcher, and both the precision and recall a re pretty good.
#248-266 Both the linguistic and structural characteristics of the candidate ontologies are changed heavily. In most cases, SOBOM can get high precision but low recall.
#301-304 This test group are four rea l-life ontologies of bibliographic references. SOBOM can only find equivalence a lign ment re lations.
Table 1. The performance on the benchmark 101-104 201-210 221-247 248-266 301-304 Average H-mean Precision 0.98 0.99 0.99 1.0 0.86 0.96 0.98 Recall 0.97 0.48 0.95 0.43 0.52 0.67 0.43 2.2
NCI
MA The directory t rack requires matching two ta xono mies describing the web directories . It inc ludes 4639 matching tasks represented by pairs of OW L ontologies, where classification relat ions are modeled as rdfs:subClassOf relat ions. But in the experiments, we found there are so me ontologies have wrong structure, they have a loop such as 1603, 1704, 2114, 2184, 2241, 2252, 2416, 3045, 3135, 3166, 3183, 3301,3398, 3440, 3556, 3653, 3695, 3711, 4075, 4129, 4544, 851, 118, 148, 1550,1723, 1863, 1967,2, 2000, 2103, 2270, 2271, 2632, 2749, 2803, 3058, 3186,3310, 3455, 3461, 3891, 4048, 4089, 4116, 4341, 4556, 614, 726, 747, totally 50 ontologies. So SOBOM cannot deal with these tests. The experiment results shows in Table 3.
Precision 0.5931 3
Strengths SOBOM deals with ontology fro m two different v iews and co mbines
results of every step in sequential way. If the ontologies have regular literals and
hierarchical structures, SOBOM can achieve satisfactory align ments. And it can avoid
missing align ment in many block matching methods [
Weaknesses SOBOM needs the anchor concepts to ext ract sub-ontologies. So it heavily depends on the anchor concepts. if the literals of concept missed, SOBOM will get bad results. 3.2
SOBOM can be viewed as a frame of ontology matching. So many independent matchers can be integrated into it. Now anchor concepts generator is a weak matcher, our next p lan is to integrate a more powerful matcher to produce anchor concepts or develop a new method to get anchor concepts.
This paper reports our first participation in OA EI ca mpa ign. We present the alignment process of SOBOM and describe the specific techn iques for ontology matching. We also show the performance in different a lignment tasks. The strengths and the weaknesses of our proposed approach are summarized and the possible improve ment will be made for the system in the future. We propose a brand new algorithm to match ontologies.