This paper presents the RE-miner results for data linking in the ontology alignment contest OAEI 2020, Spimbench track. RE-miner discovers all minimal and diverse referring expressions of all instances of a given source knowledge graph. In a second step, it exploits these referring expressions to find the possible links to a target knowledge graph. This is the first participation of REminer in the OAEI campaign and produces the best result in terms of F-measure on the Spimbench dataset.
1.1
RE-miner for data linking consists of 2 main steps. The algorithm has been thoroughly
presented in [
This system focuses on the instance matching problem between the instances of a given class of the source dataset, on which the REs have been discovered, and a target dataset having a non-empty set of mapped properties to the source. In other words, this approach assumes the schemas to have previously been aligned.
Create the source dataset. We first create the dataset on the source KG for the given class C, for which we aim to find the alignments. The dataset is created by keeping all instances that are of type C, and all sub-classes of C if the graph’s schema is not saturated. For instance, in the Spimbench track, the instances of Creative Works class are to be linked. The dataset, contains all instanced belonging to this class and its 3 sub-classes namely NewsItem, BlogPost, and Programme.
Referring Expressions. We discover all minimal and diverse referring expressions of
depth 1 on the source knowledge graph [
Linking and Voting Strategy. These REs are then used to find possible links in the
target dataset. For finding the possible candidate links, mapped properties and strict
equality are used between the atoms of a RE and triples of the target knowledge graph.
Moreover, first consider an instance u of type C in the source dataset and imagine
that k different referring expressions fRE1(u); :::; REk(u)g have been associated to
it. Each of these REs can be linked to zero, one, or more instances of the target, using
the bottom-up approach explained in [
The confidence of each RE is inverse proportional to the number of links it suggests. However, if the unique name assumption (UNA) is fulfilled, only one sameAs link can be found between u and an instance x belonging to the target KG. Thus we propose a voting strategy that assigns a weight to each distinct link. The weight is the sum of the confidence degree of the REs proposing that link. Moreover, the weights are normalized such that they have a value between 0 and 1. Finally, the instance x in the target knowledge graph being linked to u with the highest weight is selected. For the Spimbench dataset, we have set a very strict criterion. We only match two instances if and only if the link with the highest weight has a weight equal to one. This way, we imply that we only link two instances if we are really sure about it.
MELT. Matching EvaLuation Toolkit (MELT) is a framework optimized for OAEI
campaigns, facilitating submissions to the SEALS and HOBBIT evaluation platforms
[
Spimbench is an instance matching track and the only track we have done evaluations
on, in this first year of participation. It consists of two datasets of different sizes: the
SANDBOX dataset with about 380 instances and 10000 triplets, and the MAINBOX
dataset with about 1800 instances and 50000 triplets. We have compared our results
with AML [
RE-miner for data linking has shown satisfactory results in the Spimbench instance matching track. Although the source and target KGs shared almost the same ontology, there were still some properties that would not be mapped together using strict similarity. However, this did not hamper the performance of our system. This is because of the fact that RE-miner usually discovers not just one but many more REs for each instance. This will allow the system to choose the target instance most of the REs pointing to agree on. Moreover, for this dataset, we have been fastidious, only outputting links we really deem correct. As future work, we aim to do modifications, allowing us to participate in more tracks for the next years and focus more on enhancing our system’s run-time. 4
In this paper, we briefly presented the main components of our instance matching system RE-miner for data linking. The evaluation of results on the Spimbench track was presented, and we showed a better Precision and F-measure than other systems taking part in the campaign this year. However, in terms of run-time, more improvement and optimization are to be done.