We add to the widely used AgreementMaker system the capability to visualize the results of matching large ontologies with a user interface that supports navigation and search of the ontologies. The interface also supports user intervention when using a feedback loop strategy where users validate candidate mappings that have been computed automatically by matching algorithms. The interface further displays properties of the concepts to facilitate the decision process.
An ontology provides a vocabulary describing a domain of interest and a
specification of the meaning of terms in that vocabulary. An increasing number of
organizations are using ontologies to organize their knowledge. However,
different ontologies exist for the same knowledge domain. To address this issue,
ontology matching is needed, which is the process of finding the relationships,
called mappings, between concepts (classes or properties) of two different
ontologies, the source and the target ontologies [
A variety of algorithms, which we call matchers, have been developed for
matching. For example algorithms that are based on string similarity of the class
labels or on the structure of the ontologies. Advanced ontology matching systems,
such as AgreementMaker, use combinations of a large variety of algorithms [
The purpose of our work is twofold. First, we want an interactive
visualization method for large ontologies. Second, we want to support visual analytics
in a semi-automatic ontology matching process. We define some of these terms
next. Semi-automatic ontology matching integrates automatic and manual
methods. Those mappings that are believed to be incorrect are presented to users for
1 http://oaei.ontologymatching.org/
validation [
Our focus is on ontology matching visualization, not on ontology visualization, that is, we want to support the visualization of complex relationships between source and target ontology structures. Ontology matching visualization is further complicated when matching large ontologies. For example, the display of an ontology as a tree structure using the JTree class can be very helpful for small and medium size ontologies, but is less helpful for large ontologies because of the amount of scrolling needed to locate the different mappings. To better compare and analyze the matching results, a visual representation that can scale to large or very large ontologies is needed. In this paper, we investigate interactive visualizations that use pie charts, which naturally scale to any ontology size.
Our paper is organized as follows. In Section 2, we cover related ontology matching visualization approaches and in particular those that are intended for large ontologies, be they based on graphs, treemaps, or pie charts. We also cover interactive approaches based on matrices that support visual analytics. In Section 3, we describe our visualization technique, starting with the design criteria. We then describe the pie chart visualization and the comparative visualization of matching algorithms, as well as the user interface organization. In Section 4, we describe the use of our interactive interface that supports the validation of mappings in a feedback loop setting and its integration with AgreementMaker. Finally, in Section 5 we draw conclusions and point to future work. 2
A recent survey of visualization methods for ontology matching [
The cluster representation [
Another graph drawing representation that was developed for the
AgreementMakerLight system [
The Prompt+CogZ tool supports multiple visualizations, including one based
on TreeMaps and another one that displays pie charts [
A matrix visualization where the classes of both ontologies are placed along
the X and Y axes provides a more comprehensive view of the matching
process as compared with the aforementioned methods because it allows for the
whole mapping space to be visualized with equal detail. We know of two such
visualizations: the one provided by iMerge [
In conclusion, none of the above approaches provides both for scalability and for an interactive meaningful display that supports visual analytics for large or very large ontologies.
We have made the following choices and present their rationale: – We allow for ontology navigation, exploration, and searching. – We do not display all the mappings at once. If we did so, it would be difficult to find a visualization whose size does not depend on the number of mappings or on the size of the ontologies involved. – We want to focus on the mappings one level at a time and aggregate the results for the children of the nodes at that level. As long as navigation and searching functions are available, users can easily locate any single ontology node in the whole structure and see the mappings they need. – We choose a visualization based on pie charts. The reason of this choice is that no matter how large the data size is, the pie chart requires always the same modest area. Given this, we can display the results of more than one matching algorithm at a time. – When matching two classes, the most valuable information is the confidence score found by the matching algorithm (between 0% and 100%). The visualization can give priority to those mappings that maximize the confidence score between two nodes. – We want to make apparent the differences between matching algorithms. – We enable user feedback acquisition. 3.2
We start by describing how we visualize the information in terms of pie charts. For each visualization there are two pie charts, one that corresponds to a concept or node in the source ontology graph, S, called the current node, and a pie chart corresponding to a concept or node in the target ontology graph, T . We show the percentage of their children whose confidence score falls in a particular range. Figure 2 shows those ranges, namely 81%-100%, 61%-80%, 41%-60%, and ≤40%. For example, 41% of the children nodes of S have confidence scores in the range 81%-100%. Figure 2 also highlights a mapping between two nodes, A that is a child of S and B that is a child of T .
Figure 3 shows schematically subgraphs of both ontologies with roots S and T , their children, among which there are subclasses A and B, the siblings of A and B, and their children (and grandchildren). To enable navigation along the ontologies, users should be able to traverse the ontology vertically from a node to its children but also horizontally from a node to its siblings.
The next question we address from the interface viewpoint is how to combine both types of navigation. We provide a list that represents the children of a node and a tree view that represents the siblings of the current node. The main panel of the user interface for the initial version of the prototype is shown in Figure 4. In the center area there are the two pie charts previously discussed.
Immediately left of the pie charts there is a list. When users click on a pie chart slice, the list contains the ontology nodes with confidence score within the corresponding range, sorted by the confidence score. Clicking on a node in the list leads to an update of the pie charts, as that node becomes the current node.
The leftmost part of the interface contains the tree view. It shows all the siblings of the current node. When clicking on a node in the tree view, both the pie charts and the lists are updated, reflecting the change of the current node. On the top right there is a search box. Upon entering the name of the node in the search box, the left pie chart displays that source node and the right pie chart displays the target node that matches the source node. Accordingly, the tree and list view on the left are updated as well. In addition, for an easier navigation we provide additional functions such as “go to the top level”, “go to the previous level”, and “switch between class and property”. 3.3
To compare the results of two matching algorithms, we need to visualize their results at the same time. We have therefore upgraded our user interface panel to load multiple results as shown in Figure 5. We are making full use of the containers in JavaFX to manage the visual elements within the available space. We use two tile panels to display two ontology pairs. The upper panel is the main panel (colored green) and the lower one is the sub panel (colored yellow). The tree view shows the siblings of the source node in the main panel and the list view shows the children of that source node. Figure 6 displays the schematic representation of the user interface, including the flow panel. The flow panel shows the algorithms we have loaded into the application. In this example, the main panel shows the results of the algorithm we loaded using the second button of the flow panel. When selected, the second button is colored green, so as to provide a color match with the main panel.
Users can choose any algorithm for the main panel or for the sub panel. The difference between the two panels is that all lists get updated according to the changes to the main panel.
When the users click on a slice of the main pie chart, all features including all pie charts and lists are updated. To change the matching algorithm for the main panel, users only have to select another algorithm from the flow panel. 4
We list here our objectives for an interactive mechanism for matching ontologies
that can assist users in a semi-automatic ontology matching process, where users
provide feedback:
– Show candidate mappings for validation to the users; candidate mappings
are determined automatically using quality measures [
Because automatic matching methods do not always provide complete or correct mappings, the combination of user validation with the automatic methods can lead to better results than the automatic methods alone.
The interactive workflow is shown in Figure 7. It shows a “user feedback loop”
(UFL) strategy [
Charts (PC)
Navigate mappings, compare algorithms and properties
Start UFL
No
AM UFL PC MS Start
Load ontologies Visualize trees and perform matching with AgreementMaker (AM) Finish matching
Yes (Re)Rank candidates for user validation No
Open feedback panel with mappings
Close
No Yes End
Perform
Mapping Selection (MS)
Yes
Make determination Yes
No Navigate mappings, compare algorithms and properties Compare confidence scores across concepts
Key (Re)Assign
user confidence
Save Yes
No interfaces provided by AgreementMaker and the new interface communicate, so as to allow for complementary views. The user interface that is used to perform the mapping selection is an important component, which we describe next. 4.2
The user interface for the mapping selection must display one or more mappings to be validated by the users. When it shows more than one mapping, users are asked to choose among them. Navigation using the main user interface (Figure 5) provides the confidence scores to assist users in making their selection. Figure 8a shows the schematic representation of the interface and Figure 8b shows a snapshot of its implementation.
(a) Schematic representation.
(b) Mapping selection, where users input their
level of confidence (0-30-60-90).
Automatic algorithms match classes according to a variety of lexical, syntactic,
and structural criteria [
Using the AgreementMaker system, the source and target ontologies are visualized side by side using a tree paradigm as shown in Figure 10. The control panel
Fig. 9: Property clustering. (at the bottom of the interface) allows for users to run a variety of matchers. In this example, two automatic matchers have been activated (the Parametric String Matcher and the Vector-based Multi-word Matcher) in addition to manual matching. The picture depicts the display of the two ontologies and the mappings obtained in this way.
Every set of mappings in AgreementMaker is represented by the MatchingTask class. The MatchingTask class contains the following elements: a matcher, its associated parameters (e.g., the confidence score threshold), and the mappings produced by the execution of the matcher. To open our visualization system, users can select the Pie Chart Visualization tab of the drop down menu, as shown in Figure 10. The key point in the integration of the pie chart visualization with AgreementMaker is that we pass all the MatchingTask instances from AgreementMaker to the pie chart visualization. After selecting the Pie Chart Visualization tab, another tab shows up and the pie charts will be initialized automatically (see Figure 5). Users can easily switch between the tree and pie chart visualizations by clicking on the available tabs at the top of the display. 5
We devised a visualization tool for large ontology matching that integrates seamlessly with the widely used AgreementMaker system. It supports advanced navigation, interaction, and analysis and decision making features. In particular, for navigation, our tool supports the visual representation of the source and target ontologies and mappings between classes in those ontologies. It allows for the detailed access to pairs of classes to match, while providing ready access to other parts of the ontologies. Our tool displays confidence scores between classes and provides an overview of the confidence scores for the children of those classes.
For interaction, our tool supports user-driven navigation of classes and properties, the ability to search for a specific class, and to traverse the ontologies vertically (children of a class) and horizontally (siblings of a class).
Finally, for analysis and decision making, our tool displays several possible mappings to the users, so that they can choose among them, as part of a user feedback loop strategy that combines automatic with manual matching methods.
Clearly, there are several directions for future work. The first one is that
we would like to extend the comparison of matching algorithms to more than
two at a time. It may be the case that no single visualization strategy works
separately, especially for very large ontologies. AgreementMaker already provides
several different strategies [
This research was partially supported by NSF Awards IIS-1143926, IIS-1213013, and CCF-1331800. We would like to thank one of the anonymous reviewers, whose questions helped improve the final version of the paper.