Ontology matching produces correspondences between entities of two ontologies. The OMReasoner is unique in that it creates an extensible framework for combination of multiple individual matchers, and reasons about ontology matching by using external dictionary WordNet and description logic reasoner. It handles ontology matching in both literal and semantic level, and it makes use of the semantic part of OWL-DL as well as structure. This paper describes the result of OMReasoner in the OAEI 2014 competition in three tracks: benchmark, conference, and MultiFarm.
1.1
State, purpose, general statement The matching process can be defined as a function f.
A’=f(O1, O2, A, p, r)
Where O1 and O2 are a pair of ontologies as input to match, A is the input alignment between these ontologies and A’ is new alignment returned, p is a set of parameters (e.g., weight w and threshold τ) and r is a set of oracles and resources.
Alignments express correspondences between two entities. A correspondence must express two corresponding entities and the relation that is supposed to hold between them. Given two ontologies, a correspondence is a 5-tuple:<id,e1,e2,R,n>, where . id is a unique identifier of the given correspondence; . e1 and e2 are the entities of the first and the second ontology respectively; . R is a relation (e.g., equivalence(=), more general(>), less general(<), disjointness )) holding between e1 and e2. In OAEI campaign, equivalence is mainly considered; . n is a confidence measure (typically in the [0 1] range) for the correspondence between e1 and e2.
OMReasoner
O1
O2 C1’,C2’, R1’,R2’
C1,C2, R1,R2… p ( w,τ)
r dictionary 2 multi-matchers matcher1 . . . matchern
C1≡C1’ R1⊑R1’ Combination 3 reasoning
C2≡C2’ C2⊒C3’ R2⊓R2’ reference corresp.
C2≡C2’ C2≡C3’ R2≡R2’ evaluation results C1,C2, R1,R2 … C1’,C2’, R1’,R2’ … 1. Parsing: we can achieve the classes and properties of ontologies by using ontology API: Jena. 2. Combination of multiple individual matchers: the literal correspondences (e.g. equivalence) can be produced by using multiple match algorithms or matchers, for example, string similarity measure (prefix, suffix, edit distance) by stringbased, constrained-based techniques. Meanwhile, some semantic correspondences can be achieved by using an external dictionary: WordNet. Then the multiple match results can be combined by using specific strategy. The framework of multi-matchers combination is supported, which facilitates inclusion of new individual matchers. 3. Reasoning: the further semantic correspondences can be deduced by using DL reasoner, which uses literal correspondences produced in step 2 as input.
Finally, we evaluate the results against the reference alignments, and compute two measures: precision and recall.
In OMReasoner, the framework for multi-matchers is flexible, and any new individual matcher can be included. Now, the instances of multi-matchers include EditDistance and WordNet (see Fig.2). 1.2
Specific techniques used
Threshold is necessary for many matchers (especially syntactic ones) to determine whether the similarity is regarded as equivalence. For example, the edit distance of “book” and “booklet” is 3/7 (i.e., the similarity confidence measure is 1-3/7=0.57). If the threshold is 0.55, then these two entities are equivalent (with confidence measure 0.57); else if threshold is 0.6, they are not. So, we have to tune our tool via threshold. 2. Combination of confidence measure
Each individual matcher can produce correspondences with confidence measures. All these confidence measures will be normalized before combination. OMReasoner includes following flexible strategies to combine the multiple match results: (a) weighted summarizing algorithm (WeightSum) The confidence can be summarized by weighted similarity algorithm (see formula 1), where wk is the weight for a specific matcher k, and simk(e1,e2) is the confidence measure of similarity (mainly equivalence) by this method.
n sim(e1,e2) k 1 wk simk(e1,e2), n where k1 wk 1.0 (b) maximum method (Max) The maximum confidence measure is chosen among n matchers (see formula 2) . sim(e1,e 2) max(sim1(e1,e 2),...,simn(e1,e 2))
OMReasoner uses semantic matching methods like WordNet matcher and description logic (DL) reasoning.
WordNet1 is an electronic lexical database for English, where various senses (possible meanings of a word or expression) of words are put together into sets of synonyms. Relations between ontology entities can be computed in terms of bindings between WordNet senses. This individual matcher uses an external dictionary: WordNet to achieve semantic correspondences.
(
OMReasoner employs DL reasoner provided by Jena. OMReasoner includes external rules to reason about the ontology matching. However, reasoning is time consuming and only contributes a little to results. In this version, reasoning is skipped. 2
Results：a comment for each dataset performed In this section, we present the results obtained by OMReasoner in the OAEI 2014 competition. It participated in three tracks: benchmark, conference, and MultiFarm. Tests were carried out on a PC running Windows Server 2008 R2 Standard with Intel Core i5 processor running at 2.8 Ghz and 16 GB RAM. 2.1
Benchmark In this track, the ontologies can be divided into 3 categories(see Table 1) . In group 1, the lexical information have been altered to change their labels or identifiers. This alteration includes replacing the labels or identifiers with other names that follow a particular naming convention, a random name, a misspelled name or a foreign word. In group 2 have ontologies that have flattened hierarchies, expanded hierarchies or no hierarchies at all. In group 3 the ontologies are the most challenging ones to align. This is because labels have been scrambled such that they comprise a permutation of letters of a particular length. We tune our tool by using threshold T and combination strategy S, then get the better results (τwd=0.95, τed=0.9; S=Max). The results obtained by OMReasoner in the benchmarks track are summarized in Table 2.
Table 1. The categories of the Benchmark 2014 category tests cases concept 101-104 201-210 systematic 221-247 248-266 real ontology 301-304 The confidence data set consists of numerous real-world ontologies describing the domain of organizing scientific conferences. We use Combination strategy to run our system tool in Conference track. The results obtained by OMReasoner in the benchmarks track are summarized in Table 3 (τwd=0.9, τed=0.8; S=Max). MultiFarm track is composed of a subset of the Conference dataset, translated in eight different languages. In this track, the ontologies can be divided into 2 categories. In group 1 the alignments ontologies are the same. In group 2 the alignments ontologies are different.
Firstly, we take use of dictionary to translate different languages into English. Then, the translated English is imported in multi-matchers by using Max strategy. Finally we get the results. We tune our tool by using threshold, and the results can be seen in Table 4(τwd=0.8, τed=0.6; S=Max), which show that the F-Measures of the ontologies alignments in group 2 are obviously worse than those in group 1. We think the reasons are that OMReasoner is not well designed to match different ontologies which are written in completely different languages yet.
To choose better threshold, we compare the results (see Table 5) across several thresholds in Conference track. Still we use Max method to run our tool. From the results, we find that when thresholdτwd=0.9,τed=0.8, our tool performs best. So that we take use of thresholdτwd=0.9,τed=0.8 in Conference track. Using the same method, we get the better thresholds for Benchmark and MultiFarm track. Threshold
Discussions on the way to improve the proposed system The performance of inference relies on the literal correspondences heavily, so more accurate results which are exported from multi-matchers will greatly enhance the results of our tool. Some approaches to improving our tool are listed as follows: 1. Adopt more flexible strategies in multi-matchers combination instead of just weighed sum. 2. Add some preprocessing (see Fig.2), such as eliminating specific character (e.g., ‘-’, ‘_’) or separating compound words, before words are imported into matchers. 3. Take comments and label information of ontology into account, especially when the name of concept is meaningless. 4. Reexamine the use of an appropriate threshold value to optimize accuracy.
Another problem in our tool is that we ignore structure information among ontology at the present stage. And we will improve it in the future. 3.2
Comments on the OAEI procedure OAEI procedure arranged everything in good order, furthermore SEALS platform provides a uniform and convenient way to standardize and evaluate our tool. 3.3
Comments on the OAEI test cases The OAEI test cases involve all kinds of fields which include conference, anatomy, language, etc. The variety of tracks and the improvements introduced along the years makes the campaign very useful to test the performance of ontology aligners and analyses their strengths and weaknesses. Nevertheless, we miss blind tests cases in more tracks, which will allow a fair comparison between systems. 3.4
Proposed new measures After serious discussion, we believe that OMReasoner can improve a lot. Some new ways proposed as follows: 1. Enrich the semantic dictionaries because WordNet is not a professional dictionary, which cannot obtain more comprehensive semantic concepts. 2. Take into account hierarchical ones instead of only all concepts and properties. 3. Find NCI thesaurus for anatomy track. 4. Find different languages dictionaries for MultiFarm. 5. Improve the algorithm of some matchers.
6. Include more different matchers.
In this paper, we presented the results of the OMReasoner system for aligning ontologies in the OAEI 2014 competition in three tracks: benchmark, conference, and MultiFarm. The combination strategy of multiple individual matchers and DL reasoner are included in our approach. This is the third time we participate the OAEI, the results are still not satisfying and we will improve it in the future.