Simulated annealing-based ontology matching (SANOM) participates for the second time at the ontology alignment evaluation initiative (OAEI) 2019. This paper contains the con guration of SANOM and its results on the anatomy and conference tracks. In comparison to the OAEI 2017, SANOM has improved signi cantly, and its results are competitive with the state-of-the-art systems. In particular, SANOM has the highest recall rate among the participated systems in the conference track, and is competitive with AML, the best performing system, in terms of F-measure. SANOM is also competitive with LogMap on the anatomy track, which is the best performing system in this track with no usage of particular biomedical background knowledge. SANOM has been adapted to the HOBBIT platfrom and is now available for the registered users. abstract environment.
SANOM takes advantages of the well-known simulated annealing (SA) to
discover the shared concepts between two given ontologies [
A tness function should utilize the lexical and structural similarity metrics to estimate the neness of an alignment. The version of SANOM participated this year uses both lexical and structural similarity metrics, which are described in the following. 1.1
The cleaning of strings before the similarity computation is essential to increase the chance of mapping entities. SANOM uses the following pre-processing techniques to this end:
Copyright c 2019 for this paper by its authors. Use permitted under
Creative Commons License Attribution 4.0 International (CC BY 4.0).
{ Tokenization. It is quite common that the The terminology of concepts are
constructed from a bag of words (BoW). The words are often concatenated
by white space, the capital letter of rst letters, and several punctuations
such as " " or " ". Therefore, they need to be broken into individual words
and then the similarity is computed by comparing the bag of words together.
{ Stop word removal. Stop words are the typical words with no particular
meaning. The stop words should be detected by searching the tokens
(identi ed after tokenization) in a table containing all possible stop words. The
Glasgow stop word list is utilized in the current implementation 1 .
{ Stemming. Two entities from the given ontologies might refer to a
similar concept, but they are named di erently due to various verb tense,
plural/singular, and so forth. Therefore, one needs to recover the normal words
so that the similar concepts will have higher similarity. The Porter stemming
method is used for this matter [
After the pre-processing step, the strings of two concepts can be given to a
similarity metric in order to calibrate the degree of similarity between concepts.
The base similarity metric computes the sameness of tokens obtained from each
entity. The current version of SANOM takes advantage of two similarity metrics
and take their maximum as the nal similarity of two given tokens. One of this
similarity metric is for sole comparison of stirngs, and the other one is to guage
the linguistic relation of two given names. These similarity metrics are:
{ Jaro-Winkler metric. The combination of TF-IDF and Jaro-Winkler is
popular and has been sucessful in ontology alignment as well. Similarly,
SANOM uses Jaro-Winkler with the threshold 0:9 as one of the base
similarity metrics.
{ WordNet-based metric. The linguistic heterogeneity is also rampant in
various domains. Therefore, the existence of a similarity metric to measure
the lingual closeness of two entities is absolutely essential. In this study, the
relatedness of two given tokens are computed by the Wu and Palmer measure
[
The preceding string similarity metric gives a high score to the entities which have lexical or linguistic proximity. Another similarity of two entities could be derived from their positions in the given ontologies.
We consider two structural similarity measures for the current implementation of SANOM: { The rst structural similarity is gauged by the subsumption relation of classes. If there are two classes c1 and c2 whose superclasses are s1 and s2 from two given ontologies O1 and O2, then the matching of classes s1 1 http://ir.dcs.gla.ac.uk/resources/linguistic utils/stop words and s2 would increase the similarity of c1 and c2. Let s be a correspondence mapping s1 to s2, then the increased similarity of c1 and c2 is gauged by fstructural(c1; c2) = f (s): (1) { Another structural similarity is derived from the properties of the given ontologies. The alignment of two properties would tell us the fact that their corresponding domain and/or ranges are also identical. Similarly, if two properties have the analogous domain and/or range, then it is likely that they are similar as well.
The names of properties and even their corresponding core concepts are not a reliable meter based on which they are declared a correspondence. A recent study has shown that the mapping of properties solely based on their names would result in high false positive and false negative rates, e.g. there are properties with identical names which are not semantically related while there are semantically relevant properties with totally distinct names. The current implementation treats the object and data properties di erently. For the object properties op1 and op2, their corresponding domains and ranges are computed as the concatenation of their set of ranges and domains, respectively. Then, the tness of the names, domains, and ranges are computed by the Soft TF-IDF. The nal mapping of two properties is the average of top two tness scores obtained by the Soft TF-IDF. For the data properties, the tness is computed as the similarity average of names and their corresponding domain.
On the other ow of alignment, it is possible to derive if two classes are identical based on the properties. Let e1 and e2 be classes, op1 and op2 be the object properties, and R1 and R2 are the corresponding ranges, then the correspondence c = (e1; e2) is evaluated as fstructural(c) = fstring(R1; R2) + fstring(op1; op2) : 2 (2) 2
Results This section contains the results obtained by SANOM on the anatomy and conference track. 2.1
The anatomy track is one of the earliest benchmarks in the OAEI. The task is about aligning the Adult Mouse anatomy and a part of NCI thesaurus containing the anatomy of humans. Each of the ontologies has approximately 3,000 classes, which are designed carefully and are annotated in technical terms.
The best performing systems in this track use a biomedical background
knowledge. Thus, their results are not comparable with SANOM which does
not use any particular background knowledge. Among other systems, LogMap
[
The conference comprises the pairwise alignment of seven ontologies. Table 2
displays the precision, recall, and F-measure of SANOM, LogMap, and AML [
According to Table 2, the recall of SANOM is superior to both LogMap and AML. SANOM's average recall is 7% and 14% more than those of AML and LogMap, respectively, but its precision is 10% less than both of the systems. Overall, the performance of SANOM is quite competitive with the top performing systems in the conference track. 2.3
The conference comprises the pairwise alignment of seven ontologies. Table 3
displays the precision, recall, and F-measure of SANOM, LogMap, and AML [