The RSDL workbench was developed as a part of a service composition platform for service markets and provides tools to specify structural and behavioral aspects of services based upon the Rich Service Description Language (RSDL). Such comprehensive service descriptions allow a multi-faceted matching of service requests and offers in terms of their data models, operations, and protocols. Domains and application contexts of such service requests and offers are not known to the matchers in advance. Our data model matcher exploits several background ontologies to find corresponding data model elements. Data model alignments are represented in the form of relational Query View Transformation (QVT) scripts that are used to normalize behavioral models, which is a prerequisite for operation matching. For the OAEI campaign, we excluded background ontologies, because the involved additional costs did not justify the gain yet. In this paper, we present our system and the results for the OAEI campaign.
Fig. 1 gives an overview of our approach: (
The list of operation mappings is the input for the protocol matcher, that checks if the operation invocation sequences requested by the requester match with the operation invocation sequences allowed by the provider. The data model matcher and the transformation script generation was previously presented in [9]. The system was realized as an Eclipse plug-in and implements the interface of EMF Compare3 in order to reuse its graphical user interface. This paper focuses on its data model matching techniques and the results of the OAEI campaign.
As explained in Sect. 1, the purpose of the system is to match heterogeneous class models. The system automatically matches two UML class models that are part of respective RSDL specifications and generates a relational QVT model transformation script, which acts as a mediator enabling the translation of behavioral models. If necessary, the generated script can be manually revised.
In context of our system, the relevant OAEI tracks that we aim to compete in, are as follows: benchmark, anatomy, and conference. In the future, we also plan to participate in the multifarm, library, and largebio track. The tracks interactive, instance matching, and ontology alignment for query answering are less relevant for RSDLWB and support for these tracks is not scheduled.
In our knowledge, none of the existing matching system fulfills all the requirements
of RSDLWB class model matcher, i.e. (
According to the classification of [3], RSDLWB uses the following matching techniques: 1. String-based (normalization), 2. Language-based (tokenization), 3. Constraintbased (type similarity), 4. Linguistic resources / domain specific ontologies (background ontologies)4, 5. Taxonomy-based (upward cotopic similarity)4.
RSDLWB matches classes (DataProperties), attributes (DataProperties), and associations (ObjectProperties) pairwise and independently. The similarity of a pair is basically determined on the basis of their labels. In case of attributes, their type similarities [10] are considered as tie breakers. Before labels of two concepts are matched, they are split into tokens. Each single token is normalized by lowercasing and suppression of non-alphabetical characters. Next, the tokens are matched for their part. The overall label similarity arises from the average similarity of the token matching. If two tokens have identical normalized strings, they are assumed to match and get the highest similarity value.
The rest of this section addresses techniques that were not used in the OAEI campaign for reasons that are explained in Sect. 3. When two tokens are not identical, their Upward Cotopic (UC) similarity [6] is computed. The UC similarity is the quotient of the number of the tokens’ shared hypernyms and the number of all their hypernyms according to a Background Ontology (BO). Such a BO is selected when it contains two concepts with the same normalized labels as the tokens to be matched. In particular, an individual BO is selected for each label pair. BOs are stored in a relational database. The transitive closure of the hypernyms is precalculated for each BO concept and also stored in the database. We imported different ontologies to our database like WordNet [8], DBpedia [7], etc.
aliases
oid INT(
ontologies
id INT(
concepts
oid INT(
Listing 1: Querying Background Ontologies hypernyms1 hypernyms2 id prio hierarchies, because shallow hierarchies produce UC similarity values that are close to each other. Tab. 1 shows the query result set that contains the hypernyms of person and author, the ontology id and the priority. Accordingly, the UC similarity is: σUC = |hypernyms1 ∩ hypernyms2| = 7 |hypernyms1 ∪ hypernyms2| 10 = 0.7
To create n:m class mappings, a simple greedy algorithm is used. At first, the class pairs are sorted in a descending order according to their similarity. The algorithm iterates over the pairs and if none of the current pair’s classes is part of a mapping, a new mapping is created. If one class is already part of a mapping and the second is not, the second is added to the mapping the first is already part of. If both classes are part of a mapping, the pair is ignored.
In this section, we want to explain briefly how the QVT transformation script is generated from the alignment. The generation is exemplified on the basis of the reference alignment for the cmt and the confOf ontologies that are part of the conference track. The UML diagram in Fig. 3a shows parts of these ontologies and is arranged in a way so that some mappings of the reference alignment can easily be seen.
Fig. 3b shows the generated QVTr script: Each class mapping corresponds to a top relation, which is a possible entry point for the transformation, e.g. <Person, Person> (line 2). During the transformation, free variables (domains) like person1 are bound to instances of the source class model at first. Accordingly, var email is bound to person1’s data attribute email (l. 5). The enforce keyword directs the transformation to create proper instances in the target data model (if necessary). Once person2 is bound to a (newly created) instance, its attribute hasEmail is bound to var email (l. 8). Variables for object attributes (l. 13, 16) are delegated to other relations to bind free variables (l. 19). The delegation is carried out in when clauses, which are preconditions for the relations. The creation of the script is not trivial, because n:m mappings have to be considered or mapped attributes do not necessarily belong to classes that have (a) Excerpt of cmt and confOf Ontologies } } }; } transformation Cmt_ConfOf(cmt : cmt, confof : confof){ top relation Person_Person{ var_email:String; enforce domain cmt person1 : Person {
email = var_email }; enforce domain confof person2 : Person {
hasEmail = var_email } top relation Author_Author{ enforce domain cmt author1 : Author {
writePaper = var_paper1 : Paper{} }; enforce domain confof author2 : Author {
writes = var_paper2 : PaperFullVersion{} }; when{
Paper_PaperFullVersion(var_paper1, var_paper2); } top relation Paper_PaperFullVersion{ enforce domain cmt paper1 : Paper{}; enforce domain confof paper2 : PaperFullVersion{}; (b) Generated QVTr Script been mapped for their part, etc. For a more detailed description on the script generation and its current limitations, the reader is referred to [9]. 1.4
The SEALS compliant5 RSDLWB 1.1 is available at http://goo.gl/3Uj9gS. The provided alignments are available at http://goo.gl/JLsELe. 2
The RSDLWB results are summarized in Tab. 2. The second column denotes how the values for precision, F-measure, and recall were calculated. The harmonic mean of all test cases is stated for benchmark, conference, and multifarm. The tracks anatomy and library comprise only one test case. Concerning the conference track, the values are calculated according two reference alignments ra1 and ra2. The multifarm track has two kind of tasks: The first kind matches the same ontology in different languages (same) and the second different ontologies in different languages (diff). Relating to largebio, RSDLWB could only complete the test case FMA-NCI within 10 hours. 2.1
benchmark The test cases of the benchmark track are systematically generated from three seed ontologies – biblio, cose, and dog – by modifying or discarding ontology features. The evaluation is conducted in a blind fashion, i.e. neither the participants nor the organizers 5 http://oaei.ontologymatching.org/2014/seals-eval.html benchmark biblio H-Mean benchmark dog
H-Mean anatomy conference conference multifarm multifarm library largebio 00:01:26 04:00:17 00:22:17 00:00:36 00:00:36 Mouse-NCI H-Mean ra1 H-Mean ra2 H-Mean (diff) 00:18:00 H-Mean (same) 00:18:00 TheSoz-STW 09:07:08 .99 .99 .978 .81 .76 .16 .34 .781 .66 .75 .749 .59 .54 .04 .02 .073 .5 .6 .607 .47 .42 .02 .01 .038 .237
FMA-NCI 00:36:57 .956 .38
Table 2: RSDL Workbench Results for OAEI 2014 know the generated test cases in advance. RSDLWB achieved very good results regarding F-measure for the biblio and dog test cases. However, RSDLWB did not produce an alignment for cose.
The Adult Mouse Anatomy and a part of the National Cancer Institute Thesaurus (NCI) describing the human anatomy are matched in the anatomy track. In regard to precision, F-measure, and recall, RSDLWB performs slightly worse than baseline StringEquiv. RSDLWB achieved high precision for the price of low recall compared to other systems.
In the conference track, seven independent ontologies in the domain of organizing conferences are matched pairwise, resulting in 21 test cases. The produced alignments from the participants are evaluated against the reference alignments ra1 and ra2. The reference alignment ra2 is generated as the transitive closure computed on ra1. While ra1 was available to participants, ra2 was not. Regarding F-measure, RSDLWB performs better than baseline StringEquiv, but slightly worse than baseline edna, which means an average performance. Since RSDLWB relies only on string-based techniques the results are similar to the baseline algorithms. 2.4
multifarm The goal of this track is to evaluate the ability of the matcher to deal with ontologies in different languages. The cross-lingual matching scenario is relevant for RSDLWB, but we did not investigate on this scenario yet. The low precision, F-measure, and recall values result from the fact, that labels in different languages share less common tokens. Even with enabled BOs, the matcher does not support other languages than English at the moment.
The task of the library track is to match the STW and the TheSoz thesaurus, which include a huge amount of concepts and additional descriptions. These ontologies define multiple labels per concept in different languages. However, RSDLWB does not support multiple labels per concept yet. Rather, it selects an arbitrary label, so that these labels are possibly in different languages, which leads to the same problems as for multifarm and explains the weak results.
The data set of this track comprises the large biomedical ontologies Foundational Model of Anatomy (FMA), SNOMED CT, and NCI. These ontologies are semantically rich and contain a huge amount of concepts. The input size of the ontologies vary across the six test cases. RSDLWB could only complete the smaller FMA-NCI test case within the given time frame of 10 hours. For this particular test case, RSDLWB achieved significantly lower F-measure than the average of all participants. 3
Several adjustments had been made to enable a participation of the RSDLWB in the
OAEI campaign: (
After the first participation of the RSDLWB in the OAEI campaign, we conclude that the system is not optimal for the OAEI test tracks yet and that there were no improvements in any of the OAEI disciplines. As the results show, the matcher heavily relies on labels and rarely on other ontology features. Furhermore, the system in its current shape is not suitable to match large ontologies. 3.2
RSDLWB depends very much on labels. To overcome this issue, similarity metrics must be introduced that take e.g. structural features of the ontologies into account. Since the importance of the similarity metrics varies between the test tracks and cases, the matcher should be adaptive and adjust the weights for these metrics. RSDLWB failed to complete test tracks with large ontologies in a reasonable time – even without using BOs. To improve the runtime of the matcher, we plan to parallelize the retrieval of hypernyms and the calculation of similarities. When BOs are used, the system often produces false negatives because it uses homonyms for the anchoring in BOs. Therefore, we want to adjust the matcher so that it is aware of the matching task’s domain. Furthermore, we want to address cross-lingual matching by importing multilingual data sets of DBpedia or by integrating a translation service. We are confident that we can improve the system once the BO can be exploited effectively. 4
The first evaluation of RSDL workbench in the OAEI 2014 campaign showed good results for the benchmark track, but average to weaker results for the other tracks. The runtime and the quality of the matching results is improvable compared to other systems. We excluded the usage of background ontologies, because they increase the runtime of the system, but did not improve the matching results on average. As soon as we can effectively exploit BOs, we need to improve the systems’ efficiency, because the retrieval of hypernyms has an extra effect on the runtime.