Ontologies are increasingly gaining practical usage for semantic data in various ways and across multiple domains. From this growing applicability arises an evergreater need to manage large datasets, reduce analytical complexity and efficiently as well as accurately integrate different heterogeneous ontologies into or within existing systems, all while minimizing data corruption and maintaining existing semantics. In this paper, we present the GeneTegra Alignment Tool (GT-Align), a practical implementation of the ASMOV ontology alignment algorithm within a Web-based interface, focusing on biomedical data and using Unified Medical Language System (UMLS) for the background knowledge. GT-Align allows iterative alignment of multiple ontologies as well as active user involvement throughout the process.
Ontologies have been increasingly acknowledged
as an appropriate abstraction instruments for
representing entities and their relationships within
various domains
In this paper, we present a demonstration of an ontology alignment Web interface called GTAlign, consisting of a server implementation that wraps an ontology alignment algorithm and exposes REST API endpoints which the client side user interface employs to enable iterative alignment of biomedical ontologies.
Our solution to the problem of ontology
alignment is twofold. First, we use an ontology
alignment algorithm to identify shared relationships
between heterogeneous entities and generate a set of
suggested mappings. Second, we allow the user to
engage with the results on each iteration by
accepting, rejecting or clearing (reverting an
acceptance or a rejection) mappings or mapping
groups. The interface also allows the user to
upload a set of equivalence mappings as an input
partial alignment to bootstrap a new alignment
process. All mappings must be positively
accepted; in other words, no mappings are deemed
accepted until positively indicated as such by the
user. Rejection of a mapping is an indication that
such a mapping should never happen. Clearing of
a mapping, on the other hand, indicates that it is
not accepted but still possible. The GT-Align Web
Interface supports manual evaluation of results
through visual inspection, including inspection of
parents and children of elements as well as
inspection of labels and other textual information. The
tool also provides information on the confidence
of a mapping as calculated by the underlying
ASMOV algorithm
The main goal of GT-Align is to enable easier ontology alignment of biomedical data and allow domain experts to validate the results and thus ensure high quality alignments. Put succinctly, GTAlign enables the production of an alignment between any two biomedical ontologies, allowing users to review and revise mappings interactively.
The underlying alignment algorithm for GT-Align
is called ASMOV, which was developed for use
in the integration of data and ontologies in the
biomedical and life sciences domain within the
GeneTegra Information System
(www.genetegra.com). The algorithm makes
use of an iterative approach with similarity
calculations along multiple dimensions coupled with a
process of semantic verification that seeks to
remove mapping inconsistencies. ASMOV uses a
combination of string-, constraint-,
formalresource-, graph-, model-, and instance-based
matching mechanisms. ASMOV has participated
in several rounds of the evaluations performed by
the Ontology Alignment Evaluation Initiative
(OAEI), placing as one of the top three performers
in the benchmark tests of the contests in which it
has participated
Visualization types that have been shown as
most effective at enabling user involvement in an
alignment process are tree and graph structures,
with both having specific benefits to the user
Additionally, the GT-Align user interface is built
on modern web technologies including
JavaScript/HTML/CSS as well as SVGs for data
visualizations, enabling GT-Align to stay on the cutting
edge of UI tools
The following sections focus on the individual visualizations and capabilities of the different views present in the GT-Align Web Interface. 3.1
This view allows the user to upload ontologies into the system, which they can further inspect in the Hierarchical Tree View. The ontology import process uses an extensible set of rules to normalize lexical labels used within it, marking one as the preferred label and others as alternative labels. These labels are then annotated to concepts within the UMLS Metathesaurus. Having normalized labels provides a consistent visual identification scheme that is more easily recognized and thus friendlier to the user. 3.2
This part of the system displays ontologies as a
hierarchical tree of concepts. The ontology tree
visualization serves as the fundamental
visualization in GT-Align. The view is shown in Figure 1.
When concepts are asserted as children of
multiple parents, they are displayed within each parent,
as is standard practice. Metadata about the
ontology, including its ID and any annotations such as
textual descriptions, are displayed on an
information pane. The system also utilizes an
autocomplete search for individual concepts within the
ontology. The indented tree visualization presents
information in a commonly used abstraction
allowing users to explore ontologies without a
specialized knowledge of the visualization itself. This
enables anybody familiar with the concept of an
ontology to get started with the software very
quickly. The usefulness of an a tree visualization
in displaying hierarchical relationships has been
demonstrated by its long-term usage in many
areas. From visualizing file systems or HTML
structures to visualizing ontologies in tools such as
WebProtégé
Through this view, the user can execute an alignment with specific parameters. The user starts by selecting two ontologies that will be aligned. Each of the ontologies can either be selected from the set of ontologies that were previously added or may be uploaded by the system. Additionally, the user can upload a partial alignment as an input parameter to the alignment process. 3.4
This view provides a tabular summary of all the alignments executed in the system. It contains the historical overview the alignment processes ran by the user along with details about each process. An example of this view is shown in Figure 2. Clicking on a specific column heading will sort the table according to the corresponding parameter. The displayed parameters include links to the source and target ontologies used in the alignment, the custom name supplied by the user along with the date and time for when the alignment was created or last modified and how long it took to execute. The view also displays the current execution status which gets updated as an alignment progresses through different stages. Alignments can execute asynchronously within the GT-Align platform, allowing the user to perform tasks in parallel. The extensive computational work is offloaded to the server and doesn’t hinder the user experience. Once an alignment is completed, the user receives a notification of its final status. Selection of an alignment transitions the user to the Alignment Overview. 3.5
After an alignment is obtained, the mappings are presented in an overview pane with a circular graph layout. Due to the structure of ontologies, a graph-based visualization is a natural fit for displaying their alignment. Unlike indented trees, graphs are more suitable to display multiple inheritance without any visual redundancy. This prevents the user of potentially needing to make additional efforts when understanding the data at hand or being confused by concept repetition. Tree visualization is particularly less adequate when displaying large ontologies because the expansion of nodes to greater depths can quickly become overwhelming. Large trees also make it difficult to access the overall structure of an ontology. Using a graph visualization allows us to handle large amounts of data in a way that is more customizable and flexible. The graph visualization is shown in Figure 3.
Concepts from both ontologies are distinguished in the graph by color and positioning. They are separated based on their originating ontology where the concepts from the source ontology are placed on one side and the concepts from the target ontology on another. The user can rotate the graph as they please. Further clustering of concepts is performed based on their hierarchical position in the ontology. The closer a concept is to the root, the closer it is to the center of the graph. Conversely, the outer section of the graph represents the leaf nodes. This design allows the user to easily asses the structure of both ontologies while performing minimal work.
Two concepts connected with a line represent a single mapping. Thickness of this line corresponds to the confidence value, i.e. the level of confidence in the mapping being correct, according to the underlying algorithm. The thicker the line the higher the generated mapping confidence. This lets the user evaluate a section of the alignment by the amount of high- or low- confidence mappings it contains.
The user can click on a specific mapping or group of mappings under a common parent. Selection of a group of mappings transitions to the Mapping Group View, while selection of any individual mappings transitions to the Mapping View. At the top is a toolbar that provides control to filter the mappings by the confidence value. Additionally, the mappings can be filtered by the mapping origin (suggested by algorithm, found by algorithm, provided by partial alignment), mapping state (accepted by user, rejected by user, undefined), and a branch in the ontology. This filtering is automatically reflected in the graph, allowing users to quickly see the alignment overview at different scales. Besides filtering, the toolbar allows for bulk editing of mappings, enabling the user to accept, reject or clear mappings for large sections of the alignment. The toolbar additionally allows the user to export the mappings through the Alignment RDF format and the EDOAL format, as well as a merged OWL ontology. 3.6
This view, shown in Figure 4, displays the suggested mappings for a concept and its children. The main purpose of this view is to allow the user to examine a group of mappings separately from the alignment as a whole. The mappings are shown as a vertical list of concept pairs, giving the user an alternative presentation to the graph that is familiar and straightforward.
Each concept pair contains a rectangular visualization of their mapping confidence on a scale of 0-100. The mapping state is show above each of the confidence visualizations. Control buttons are provided allowing the user to alter the mapping state of the whole group as well as of individual mappings within it. Selection of an individual mapping transitions to the Mapping View section. This multi-pane view, shown in Figure 5, provides an in-depth visualization of a single mapping, allowing the user to review or modify the mapping.
Top left pane: This pane contains information about the currently focused source concept selected by the user, including a preferred label, alternative labels, and annotated UMLS codes if available. All other panes display information in relation to this currently focused concept.
Top right pane: This pane contains a selected target concept for mapping to the currently focused concept. In the center between the top right and left panes are controls allowing the user to accept or reject the mapping between the focused and selected concept, or create a new one if none yet exists. The controls also include an option for the user to swap the focused and selected concepts, causing other panes to adjust accordingly. Three central panes under the controls show one or more mappings in different states.
Bottom center panes: The top pane shows the accepted mapping for the focused source concept, if one exists. The middle central pane shows a list of possible mappings according to ASMOV. The top concept in this list is the mapping suggested by ASMOV, other mappings are alternative possibilities. The bottom central pane shows a list of rejected mappings, if any exist. In all three panes, mappings show their preferred label and mapping confidence value. Clicking on a concept in any of these panes places it on the selected mapping pane, making it the new focused concept.
Bottom side panes: Two bottom panes on each side contain the hierarchical tree view of both ontologies (see Hierarchical Tree View). On the bottom left is the ontology tree corresponding to the focused source concept, and on the bottom right is the ontology tree corresponding to the selected target concept. Each ontology tree highlights the selection of the corresponding concepts. All ancestors of both concepts are shown in the respective tree view along with all siblings of each ancestor, but children of ancestor siblings are initially hidden although the user can explore them if desired. Clicking on a concept within the left ontology tree will set it as the currently focused concept, updating all other panes accordingly. Clicking on any concept from the right ontology tree places it on the selected concept pane, choosing it as the mapping for the source concept. 4
Previous sections highlight the current status and main features of the GT-Align Web Interface. There are several planned features expected to be produced in the future. Subsequent releases will allow users to specify subsumption relationships in addition to equivalence relationships. GT-Align will include the capability of performing an automated evaluation of precision and recall against a reference alignment, and to display the results of such evaluation. Additionally, it will incorporate functionality to dynamically change the set of fixed weights for the various similarity values calculated by ASMOV. It will also present the separate confidence scores for the different measures of similarity generated by ASMOV. GT-Align will support pivot systems such as EDOAL to allow importing of alignments from other systems. Finally, GT-Align will support interactive collaboration by multiple users. 5
In this paper, we presented GT-Align, a versatile implementation of an ontology alignment Web interface, backed by an efficient, robust and fieldtested algorithm called ASMOV. Besides the algorithmic alignment, the interface enables iterative user involvement, allowing domain experts to improve and validate the results thus contributing to the quality of the alignment. We showed features to evaluate relationships between entities in different biomedical ontologies as well as explore ontologies on their own through hierarchical trees. The interface enables users to analyze aligned mappings from a high-level perspective through groups refined by optional user filters and combined with algorithm results. It further provides capabilities for fine grain inspection of individual concepts through their mappings as well as relations to other concepts. An assortment of visualizations provided by the user-interface enables multiple perspectives on the data itself along with the alignment results. Additionally, we presented our plans for future development. In summary, GT-Align is a robust and easy to use solution for ontology alignment of biomedical data.
This work is supported by grant R44GM097851 from the National Institute of General Medical Sciences (NIGMS), part of the U.S. National Institutes of Health (NIH).