Links among data elements represent the core of the Semantic Web. The links are built with semi-automatic linking algorithms using a variety of similarity calculus. The data interconnected by these algorithms demands automatic methods and tools to maintain its consistency. These changes occur mainly in datasets that represent knowledge in areas that evolve drastically, like biology and medicine. Even though the constant change of these links is an important task for the evolution of these structured datasets, such changing operations can negatively influence the well-established links, which turns difficult the consistency of the connections over time. In this work, we aim to investigate new methods responsible for fixing and updating links among ontologies in the Linked Open Data context.
The remaining of this paper is organised as follows: Section 2 presents the general and specific goals of the ongoing research; Section 3 shows the related work about link maintenance. Section 4 presents the methodology and the framework, while 5 presents what has been done and achieved so far. Finally, Section 6 explains the next steps and challenges.
We aim to investigate, formalize and implement semi automatic link maintenance actions in order to recognize affected links and turn them up-to-date. These actions will address cases of structurally broken links, which part of the link was removed, and semantically broken links, which the meaning of the resources changed.
Our research will provide the following contributions:
An exhaustive study of the behavior regarding broken links and its correlation with simple and complex changes in Linked Open Data datasets; A framework that outputs the broken links found between two versions of the same dataset, in addition to the fixing, rebuild or removal of these links; Each of the steps mentioned at Section 4 retrieves relevant knowledge of the maintenance of links timeline: In Step 1, we can retrieve a list of changed concepts between the versions of the datasets; In Step 2, a list of links that may be broken; In Step 3, an up-to-date dataset.
Previous investigations addressed the problem of broken links using different methods. Literature has shown that most of the proposals only handle the detection phase of the process and are not concerned with the fixing phase. It has also shown that there is no evidence of a solution that can automatically detect and fix these links without human intervention.
Dealing with broken links in the traditional Web is not a novelty problem.
Existing work attempted to adapt traditional Web techniques of dealing with broken links to
the Semantic Web environment [
Another approach stores the versions based on changes in the dataset and, in the
case of inconsistency detection, it is easier to manually restore previous versions and find
the source of the potentially broken link. One of the limitations in this approach is the
number and size of deltas, which are the mapping of differences between two versions of a
given dataset [
In order to keep to links up-to-date, we are building a framework composed of three main steps, listed as:
Step A: Detect the changes that occurred in a given period of time based on two releases of the involved datasets that follow the LOD principles. These changes can be simple changes (atomic changes like addition or removal actions) or complex changes (non-atomic changes like update action, a change composed by a removal followed by an addition actions) of the knowledge stored in the datasets; Step B: Recognize which of the changes found at Step A turned into an affected link (semantically or structurally broken link). This link could be also created, removed, updated or remained untouched (which can also be an example of a inconsistency, given that the dataset had evolved, but the link remains the same); Step C: Apply corrective actions in the recognized affected and broken links. This action can be a reconnection with an unbroken link (like the children or parents of the outdated link) or the removal of the link.
Figure 2 shows the steps mentioned above and subsections 4.1, 4.2 and 4.3 explains the purpose of each of them.
The initial step (Step A of Figure 2) is responsible for detecting the changes that occurred in a given period of time based on two releases of the involved LOD datasets. These changes can be simple changes (like simple addition or removal of triples) or complex changes (update action) of the knowledge stored in the datasets.
As an example, suppose there is a link connecting two resources in two different datasets. The first dataset has a resource entitled “Dubai”. The second dataset also has a
Figura 2. LODMF Framework [
This step derives some research questions: How to detect changes between two LOD datasets? Which kind of changes most occurs in LOD datasets? Is there any relation between the domain of the dataset and the most frequent changes?
Step B is responsible for categorizing the set of links as inconsistent or consistent, depending on the change it suffered. Following the example presented at subsection 4.1, in Step B the changed link (“Dubai” connected to “Dubai” changed to “Emirates of Dubai” connected to “Dubai”) is evaluated. If the change did not alter the meaning intended by the link’s author, then the change is categorized as valid and the link as unaffected. However, this is not the case of the modification from “Dubai” to “Emirates of Dubai”, which changed the meaning of the link. “Emirates of Dubai” from first dataset can not be linked to “Dubai” resource of the second dataset by the predicate “sameAs”, given that these resources do not represent the same thing in the real world. This semantic inconsistency is detected and the set of affected links proceed to Step C.
The research questions related to these steps are: How to precisely categorize a link as inconsistent? Is there a correlation between the number of affected and unaffected links? The proportion of broken links over the total links is the same between the many versions of the dataset through time?
At Step C we apply maintenance actions on the affected links detected at Step B (subsection 4.2). In our example, given that the link between “Emirates of Dubai” and “Dubai” is inconsistent, it should be repaired at Step C. The repairment can be a reconnection to another resource (connect “Emirates of Dubai” from the first dataset to a synonym in the second dataset), a replacement of the predicate connecting the resources (“sameAs” can be changed to “differentFrom”) or, in the last case, the complete removal of the link.
The derived questions in these steps are as follows: As long as an invalid link is found, which actions to take? How can we assure that the newly created link fits in more than the older one? How to semantically reconnect a link?
Tables 1, 2 and 3 shows the collected results in a conducted study [
No Add 0.41%
Tabela 1. Added Cases
3.85%
Tabela 2. Removed Cases
0%
0.04%
4.41%
Tabela 3. Modified Cases
Table 1 shows that Agrovoc dataset applies the concept of Linked Data, linking 99% of their newly added triples to an external dataset. At Table 2, however, 96.15% of identified removed cases show that if an internal triple is removed, the connecting link remained untouched, generating cases of structurally broken links. Table 3, regarding modification, shows that the fourth sub-case concerns the most frequent one, in which the modification of triples led to unchanged links. This case needs additional studies to further observe to which extend these unchanged links remained semantically inconsistent due to the modifications of the associated RDF triples.
We also performed a literature survey to understand to which extent the link maintenance problem for integrity in Linked Open Data was studied. We discovered that most of the developed techniques focus on the discovery part of the broken links, not in the fixing part. We also discovered that, to the best of our knowledge, there is no evidence of a tool that can discover and fix semantically broken links automatically. This study is not yet published.
We are now focusing on developing novel strategies to address the challenges of
identifying broken links and maintaining them (Steps B and C of Section 4). In addition, we
are evaluating state-of-art methods to detect semantic inconsistencies in the links, such
as the usage of genetic programming [