Knowledge Graphs have emerged as a core technology to enhance the search experience of users in research and industry alike. Applications are being developed to build and exploit graph structures in data repositories, to improve search and integration of knowledge. However, data is available on the Web in diverse formats and integrating data from di erent sources is still an open research area. In this paper, we propose a method to build a schema for Knowledge Graphs that leverages ontologies by integrating them in a single uni ed graph, enriched by an ontology matching step. We evaluate this method by comparing the di erent stages of graph structures and measuring structural properties of the resulting graphs. We also propose an approach that groups ontology classes by conceptual type, showcasing how an underlying ontology-based schema can enhance a Knowledge Graph. The results demonstrate the viability of this approach, highlighting how the chosen techniques based on edge addition successfully make our ontology-based knowledge graph schema more complete and tightly connected.
The Web provides an infrastructure where knowledge can be stored, shared,
and created, commonly in an unstructured format, using natural language. In
response to this lack of structure, the Semantic Web [
Search engines often nd unstructured information just by crawling the Web.
Recently, however, search engines have been evolving towards a more developed
search experience by leveraging Semantic Web technologies to structure
information. For instance, Google coined the term Knowledge Graph1 (KG) and
described it as a way to nd new information about entities quickly and easily. The
concept of KG has since been adopted by the Semantic Web community to
describe graphs that aggregate knowledge from di erent domains by including
realworld entities and their relationships [
However, creating KGs from information on the Web, even when structured, is not a trivial task. When information sources were not speci cally developed
things-not.html (Retrieved in November 2018)
Copyright 2019 for this paper by its authors. Use permitted under Creative
Commons License Attribution 4.0 International (CC BY 4.0).
with interoperability in mind, data structures do not follow the same standards
or models, and integration becomes a tough challenge. Therefore, ontologies
are fundamental to ensure the consistency between datasets since they formally
represent and precisely de ne concepts and their relationships. When combined
with LOD, they improve the semantic content of the data and link datasets at
a schema level (as shown, for example, in [
Figure 1 illustrates a well-connected network of ontology classes that is used at the schema level to integrate knowledge in a KG. Ontologies are structured as a graph where the concepts/classes are nodes and the axioms/properties are edges. These axioms refer to relationships within the ontology (intra-links) or link to concepts in other ontologies (inter-links). Then, graph enrichment methods facilitate the discovery of new relationships between ontology entities (enrichment links), resulting in a more tightly connected graph. Finally, assigning classes from the schema layer to entities in the data layer (annotation links) allows the inference of new relationships or paths between entities (inferable links).
Intra-links Inter-links Enrichment links Annotation links
Inferable links Classes 1. Organization 2. Company 3. Educational Institution 4. School 5. University 6. Organization 7. Educational Organization 8. Higher Education Organization 9. University 10. Abstract Information 11. Course Properties taught at owl:subClassOf rdf:type O1
D1
D2 O2
Data Layer Dataset 1: University data, e.g. name, number of students, location Dataset 2: Course data, e.g. name, lecturer, hours University_X D1 D2 Mathematics DBO
1
KG: one describing universities and the other describing courses. After
aligning the data with the schema, a connected ontology graph allows for the newly
discovered shorter path (blue path) describing a Course that is taught at an
Organization that, in this case, is a University. Finally, akin to the UMLS
Semantic Types [
In this paper, we propose and evaluate a method that exploits existing ontologies to improve integration of data at the schema level in KGs. For this purpose, we analyse the impact of axioms and ontology matching in the overall structure of the graph. We also demonstrate how to take advantage of an ontology graph as the underlying schema of a KG by proposing a method that assigns semantically relevant groups to ontology entities.
We start the article by discussing related work, followed by a detailed explanation of the methods and evaluation. We then present and analyse the results. Finally, we present our conclusions, limitations of the methods and future work. 2
Related Word
DBpedia [
The problem of knowledge integration addresses interoperability problems
in data. KARMA [
KnowLife [
Methods
The following sections describe our methods for building, enriching, and characterising an ontology graph to be used as the foundation of a KG. Figure 3 shows the stages of the proposed methodology based on the hypothetical example from Figure 2. The following sections describe these stages in detail.
Hierarchical Parsing
Relationship Extraction A graph G is de ned as G = (V; E), where V is a set of vertices or nodes and E is a set of edges that connect the nodes.
In this step, a set of ontologies is loaded and processed into a graph G. The ontology classes are represented as nodes in V, while E translates into the set of hierarchical relationships obtained from the rdfs:subClassOf axiom. In Figure 3 the red lines represent these axioms in the three example ontologies. The graph resulting from this step is named Base Graph or BGraph for short. 3.2
Relationship Extraction When present, equivalent class axioms, property domain/range de nitions, and logical de nitions are added to the ontology graph. Equivalent class axioms (owl:equivalentClass ) are added as bi-directional edges between the equivalent nodes. The property domain/range de nitions are added as an edge between the domain and the range, and the edge is labelled with the property name.
Logical de nitions [
Figure 3 illustrates owl:equivalentClass axioms, in red, connecting DBO's Organization with AKT's equivalent class. Similarly the axioms connect University between the two ontologies. The diagram also shows an hypothetical logical de nition (green line) to connect the intersection between DBO's Organization and University with AKT's Higher Education Organization. The graph resulting from this step is an increment over the Base Graph and was named Axiom Graph or AGraph for short. 3.3
Ontology Matching Ontology matching is the process of nding mappings between overlapping concepts in ontologies to create alignments. We apply this approach as a edge enrichment step, however, these edges do not necessarily adhere to ontology formalisms. Instead, they represent links between related concepts to be leveraged when extracting knowledge from a KG.
AgreementMakerLight (AML) [
We match sets of ontologies with AML's Word Matcher, which uses a bagof-words strategy to nd word overlap between two ontology class labels and scores the mappings with Jaccard similarity. Mappings with a score higher than 0:4 were selected. If con icting mappings were found, only the highest scoring ones were selected for the nal alignment. We performed a baseline evaluation of the chosen ontology matching strategy by using the anatomy dataset from the Anatomy track of the OAEI 20184, which provides two anatomy ontologies and a manually curated reference alignment.
Figure 3 illustrates a possible mapping (red line) between DBO's Educational Institution with AKT's Educational Organization. The graph resulting from this step is an increment over the Axiom Graph and was named Mappings Graph or MGraph for short. 3.4
Characterisation and Evaluation To the best of our knowledge, no benchmark or comparable approach exists to evaluate our methods. However, the graphs obtained from the hierarchy parsing, axiom extraction, and ontology matching (BGraph, AGraph, and MGraph, respectively) are increments of each other, i.e. the AGraph is built from the BGraph and the MGraph is enriched from the AGraph. This incremental building process allows us to assess how the more complex ontology axioms and the ontology matching process a ect the base structure of the ontology graph.
We adopt the BGraph as the baseline for the evaluation since it represents
the minimum set of existing edges between the considered ontologies. We then
evaluate each step as an increment over the previous graph with
characterisation measures that provide an overview of the structure. This evaluation strategy
assesses the impact of each step, giving a better understanding of the graph
evolution. Ideally, the results of this evaluation present graphs that are increasingly
more cohesive and connected, translating into graphs that can relate concepts
more e ciently and e ectively. The metrics used are de ned as follows:
{ Connected Components (CC) refers to sub-graphs where every node has
a path to all other nodes in the same sub-graph.
{ Clustering Coe cient (CCF) [
More speci cally, the number of CC and the AD are directly a ected with the addition/removal of edges. In the ontology graph, these metrics evaluate the
overall connectedness of the graph. Unchanging values between assembly steps mean that the number of edges added had a low impact on the graph structure.
CCF and CV have a range of [
The distribution of shortest paths evaluates the impact of the edges added in each step in terms of e ciency of the graph when querying for similar or related concepts. An increase in the number of shortest paths demonstrates that new edges are creating new paths between the concepts in the graph. 3.5
Semantic Type Assignment Besides providing a standardised structured schema for a KG, is possible to extract knowledge from the ontology graph and later exploit it to enhance the KG. For this purpose we propose a method that assigns a semantic type to ontology classes. The semantic type allows ontologies to be grouped beyond their original domain into clusters of related concepts. These groups can be exploited, for example, in inference tasks for consistency checks or for information summarisation and extraction.
In this step, we associate nodes with a term obtained from speci c
background knowledge. For the purpose we used WordNet synonyms for the type
assignment of general domain ontologies and UMLS Semantic Network [
WordNet semantic type assignment. WordNet is a database of lexical elements that are grouped into sets of cognitive synonyms (synsets). In February 2019, WordNet had 117 659 unique synsets.
We assigned WordNet synsets to general domain concepts by selecting the
rst synset that matches the label of each ontology class. When no match is
found, the label is split into its words and the rst synset matching each word
is added as a semantic type. Our approach was evaluated by removing manually
assigned semantic tags in the SemCor corpus [
UMLS semantic type assignment. The UMLS Semantic Network consists on a set of broad concepts that categorise all concepts in the UMLS Metathesaurus. In November 2018, the UMLS Semantic Network had 127 unique semantic types.
We assigned semantic types using two methods: direct assignment and automatic extraction. When a class in an ontology cross-references a UMLS source term, we directly assign the node in the schema with the same type as the cross-referenced UMLS source term. However, when no semantic type is directly extracted, it is automatically computed as follows: 1. The preferred label, L, is stemmed and stop words are removed. 2. L is matched with terms from the UMLS Metathesaurus by rst nding all UMLS terms that match at least one word of L and then selecting the terms with the highest number of intersecting words, if no perfect match is found. 3. If more than one match is found, they are scored with a cosine similarity function, and the ones with the highest scores are selected.
The approach is evaluated by removing the labels from ontology classes crossreferenced with UMLS types and re-computing them with the described approach, evaluating the original types against the new with precision and recall. 4
Results and Discussion This section details and discusses the results of the assembly, enrichment and enhancement of the ontology graph. 4.1
Ontology loading
We performed experiments over ontologies of diverse domains found in the
Linked Open Vocabularies (LOV) [
We also used a domain speci c ontology set extracted from the Ontology
Lookup Service (OLS) [
These two test sets provide a rich environment for testing with ontologies
with di erent characteristics. Biomedical ontologies are domain-speci c
ontologies and, therefore, are more likely to have equivalent or related terms between
them. These ontologies are also characterised by their well-de ned standards [
Ontology Matching The baseline evaluation ontology matching strategy obtained a precision of 53.6% and a recall of 83.2% when matching the two ground-truth anatomy ontologies. Despite the reasonably good performance with this matching strategy, a precision of 53.6% can lead to a signi cant number of incorrect relations in the graph. However, the OAEI Anatomy track evaluates the ability of an ontology matching system to nd complete equivalences between two ontologies. In our case, we are looking for related concepts that do not need to be equivalent, e.g., the matching approach aligns lateral thyrohyoid ligament with thyrohyoid. These two concepts are not equivalent and, therefore, the OAEI Anatomy reference alignment does not contain this mapping. However, these two concepts are related since lateral thyrohyoid ligament is conceptually related to thyrohyoid. Therefore, in the case of building a KG, it is still desirable to nd this relationship since it can facilitate the connection between two di erent concepts that are not semantically equivalent but have a real-world anatomical relationship.
O General ontologies Biomedical ontologies jOj 314 181 example, the class http://purl.obolibrary.org/obo/HAO_0002311 is a root class with no descendants and, therefore, is isolated in the graph.
Overall, in the BOG, the BGraph and the AGraph are structurally similar since only a small number of edges was added ( 7% more edges). However, between the AGraph and the MGraph, the di erences are more prominent due to an 48:9% increase in the number of edges. The BGraph and AGraph have 94% of the nodes in the LCC. In the MGraph, 99% of the nodes are connected in the LCC. Most of the single elements connected components are due to ontologies that do not follow common standards and formalisms to build an ontology or reuse classes. For example, the ontology Flora Phenotype Ontology 5 was created only with classes and contains no intra or interlinks between classes.
Figure 4a and Figure 4b show the distribution of shortest paths for the ontology graphs. Due to the large size of the BOG, the distance histogram was computed for a sample of 60% of the total number of nodes of this graph. ×107 (i)
In both gures, plot (i) corresponds to the overlap of the distance histogram of the BGraph with the AGraph, and plot (ii) corresponds to the overlap of AGraph and MGraph. Overall, both axiom and mapping edges have a signi cant impact in the number of shortest paths that connect the nodes, which is directly related with the increase of the number of elements connected in the LCC, i.e., more nodes connected, new shortest paths created. Despite the lower impact of the edges added to AGraph in terms of structural properties in comparison to MGraph, the distribution of shortest paths shows a comparable increase in the number of shortest paths.
These new paths facilitate the discovery of relationships between concepts in the graph, but also help shorten distances between them. More concretely, the changes in the distribution of shortest paths can be illustrated, for example, by the changes in the shortest path between the node with the label cancer (DOID:162) and malignant cell (CL:0001064). In BGraph, the shortest path between the two nodes is 10, in AGraph is 4, and in MGraph the concepts nodes are directly connected, i.e. a shortest path length of 1.
In summary, the characterisation of the proposed ontology graph demonstrates the contribution of each of the building steps to the overall structure of the graph. The addition of axioms and mappings improves the connectedness of
the graph and creates new paths between nodes. The ontology matching step proved to be signi cant, which shows that, in the context of building a foundation for a KG, ontology matching is an approach that can connect previously disconnected parts of the graph facilitating the process of nding related concepts. The new edges allow concepts in the graph to be more easily reachable between each other, facilitating the discovery of new relationships. 4.4
Semantic types The evaluation of the methodology for WordNet type assignment achieved a precision of 63:2% and a recall of 77:1%. Considering the reasonable performance of the naive approach, we proceeded to the assignment of types to the ontology classes in the GOG. The approach identi ed at least one type for 73.6% of the nodes. The most common types were item.n.01 with 1 562 nodes, assay.n.01 with 120, and biological.a.01 with 99 nodes. The most common types are broad terms due to the assignment of all synsets that match each word of the label to the ontology class. A manual inspection revealed that this implementation detail was also responsible for most of the wrong type assignments.
For example, the synset person.n.01 is de ned in WordNet as \a human being". The automatic matcher assigned 29 ontology classes to this synset. A manual analysis showed that 75.8% of the classes assigned to this synset closely matched the WordNet concept of person or were related to it. For example, http://xmlns.com/foaf/0.1/Person, http://d-nb.info/ standards/elementset/gnd#DifferentiatedPerson, and http://purl.org/ goodrelations/v1#Individual are three classes that are grouped under the synset person. Five of the explicitly seven wrong assignments occur due to the methodology of matching each word separately if no synset is found. For example, Individual epitope immunization in vivo (http://purl.obolibrary.org/ obo/OBI_0001176) matched due to the keyword \individual".
The UMLS evaluation obtained a precision of 85:1% and a recall of 99%. Both strategies obtained a reasonably high performance, with the approach developed for UMLS performing better than the WordNet method. In relation to the semantic type assignment in the biomedical ontology graph, 3:3% nodes had a direct semantic type assigned. The automatic approach was able to nd a type match for 76:9% of the remaining nodes. The most frequent types were Eukaryote (T204) with 724 749 nodes, Bacterium (T007) with 525 529 nodes, and Plant (T002) with 247 825.
Overall, the goal of the semantic type assignment was to provide an example of an approach to extract information from the ontology graph as a KG schema. In this context, this semantic grouping system clusters concepts into topics and, considering that both WordNet and UMLS have their own graph structures, these groups can be further exploited to extract more relationships. For example, the semantic type Body Part, Organ, or Organ Component (T023) groups every relevant concept under the same topic. In a search engine scenario, this semantic type can be used to retrieve all the information of the KG about body parts. 5
Conclusion This work proposes methods to create an ontology graph that can be exploited to enhance the knowledge contained in a KG. We analyse these methods by comparing the evolution of the graph at di erent stages and assess the impact of each incremental step in the overall structure of the graph. We add edges to the ontology graph in three steps: owl:subClassOf axioms, remaining class axioms, and ontology alignments. Each step yields a more tightly connected and cohesive graph than the step before. The addition of the extra class axioms over the basic owl:subClassOf hierarchy shows that complex axioms have a signi cant impact on the overall graph structure, specially the number of paths between nodes. In the context of a KG schema, taking advantage of these axioms provides more relationships between concepts that contribute to the global usefulness of a KG.
We also propose to enrich the original ontology by performing pairwise ontology matching over all the ontologies in the graph. This approach showed to have a signi cant impact in the structure of the graph, especially by connecting previously disconnected portions of the graph, enabling the discovery of new relationships and paths between concepts.
Finally, we present a method to exploit the proposed schema by assigning semantic types to the ontology classes. We use background knowledge extracted from WordNet and the UMLS Semantic Network to group ontology classes into their more speci c domains. This approach showcases the potential to extract information from the ontology schema that can be leveraged in di erent scenarios, such as nding new relationships between sets of concepts. 6
Limitations and Future Work In this paper, we proposed and discussed methods that can be leveraged in di erent scenarios, however, these results also highlighted several limitations to the approach that will be improved in the future.
We currently only consider object properties if they have domain and range assigned. In these cases, the property is added as the label of the edge between its domain and range. In the future, we will analyse how to leverage ontology properties has a new assembly step in the graph building process.
Currently, logical de nitions are added as two separate edges, when in reality
these de nitions represent an intersection between three concepts. Therefore,
the integration of logical de nitions as edges will be improved to more closely
represent this relationship. In the future, we will explore how logical de nitions
can be fully explored, including assessing the potential of more complex ontology
matching to discover new relationships involving three concepts [
Despite the optimisation of the ontology matching step, ontology matching is a complex subject which was simpli ed for the purpose of evaluating its applicability in this case. In the future, we will examine how to optimise this process further, including developing an heuristic for large scale ontology matching. A robust ontology matching step can also be exploited to make a more compact graph by merging equivalent concepts in a single node.
In relation to the semantic type assignment, currently we propose naive approaches that achieve reasonably good results. However, a fully automated and reliable process would be ideal for semantic type assignment. Further applications for querying and relationship prediction will be developed to extend the usefulness of the ontology graph schema. Overall, the approaches proposed and the analysis show the potential of a uni ed ontology graph as KG schema to knowledge extraction and relationship discovery.
Acknowledgements This work has been partly funded by Science Foundation Ireland (SFI) under Grant Number SFI/12/RC/2289, Insight Centre for Data Analytics.