In ontological theories, is-a hierarchy must represent the essential property of things and hence should be single-inheritance, since the essential property of things cannot exist in multiple. However, we cannot avoid multiperspective issues when we build an ontology because users often want to understand things from their own viewpoints. Especially, in the Semantic Web, the variety of user issues in capturing target domains. In order to tackle this multi-perspective issue, we should adopt an approach of dynamically generating is-a hierarchies according to the viewpoints of users from an ontology using single-inheritance. This article discusses a framework for dynamic is-a hierarchy generation and its implementation as an extended function of Hozo. Through the function, users can understand an ontology from their own viewpoints.
Ontologies are designed to provide systematized knowledge and machine readable vocabularies of domains for Semantic Web applications. The competences of semantic technologies strongly depend on the ontology which they use. Ontology is defined as “An explicit specification of conceptualization” [1], and it clearly represents how the target world is captured by people and systems.
Semantics of concepts (classes) are defined clearly through the description of their relationships between other concepts in an ontology. In particular, the most important relationship is an is-a (sub-class-of) relationship which represents a relation between a generalized concept and a specialized concept. Class hierarchies according to is-a relationships are called is-a hierarchies, and they form the foundation of ontologies. That is, is-a hierarchies in an ontology reflect how the ontology captures the essential conceptual structure of the target world.
Therefore, in ontological theories, an is-a hierarchy should be single-inheritance because the essential property of things cannot exist in multiple. Imagine that objects, processes, attributes, etc. have their own unique and essential properties. The use of multiple-inheritance for organizing things necessarily blurs what the essential property of things is. This observation is strongly supported by the fact that both of the well-known upper ontologies, DOLCE and BFO, use single-inheritance hierarchies.
Nicola Guarino criticizes the careless usage of is-a relationships without enough ontological consideration as is-a overloading [2] and proposes an ontology development methodology, called OntoClean, which defines concepts based on metaproperties such as rigidity and anti-rigidity. DOLCE is developed based on the OntoClean methodology using single-inheritance is-a hierarchy. BFO is the upper ontology used by the OBO Foundry1 which aims to create a suite of orthogonal interoperable reference ontologies in the biomedical domain. BFO also uses singleinheritance hierarchy, and it is recommended in the guidelines of OBO Foundry to avoid careless usage of multiple-inheritance.
However, we cannot avoid multi-perspective issues when we build an ontology across multiple domains. It is because domain experts often want to understand the target world from their own domain-specific viewpoints. In many cases, their interests are different even if they are experts in the same domain. In some domains, there are many ways to categorize the same kinds of concepts, and different taxonomies are used depending on the purpose and situation.
For example, in the medical domain, a disease is interpreted from various viewpoints. Consider diabetes as an example. Clinician may pay attention to body parts with the abnormalities and classify diabetes as an abnormal blood sugar level. On the other hand, certain specialists may pay attention to the main condition and may classify diabetes as an abnormality in metabolism, and other specialists may classify diabetes as a lifestyle disease. Staffs administering medical care implicitly understand which is-a hierarchy should be used for disease interpretation in correlation with their respective interpretations. This suggests that one is-a hierarchy of diseases cannot cope with such a diversity of viewpoints since a single-inheritance hierarchy necessarily represents one viewpoint.
Many efforts are under taken to solve these multi-perspective issues. The OBO Foundry proposes a guideline for ontology development stating that we should build only one ontology in each domain [3]. This is an effort to exclude the multiperspective nature of domains from ontologies. Ontology mapping is used as an approach to acceptance of multiple ontologies based on the different perspectives in a domain. It aims to make clear the relationships between different ontologies. Someone may consider that multiple-inheritance is an easy way to solve these multiperspective issues. Because multiple-inheritance causes some ontological problems as mentioned above, our ontology development tool, named Hozo2, allows the user to use a multiple-inheritance only when he/she represents clearly from which upperconcepts the essential property is inherited3. However, if we define every possible is-a hierarchy using multiple-inheritances, they would be very verbose and the user’s viewpoints would become implicit.
In order to tackle these multi-perspective issues, the authors take a user-centric
approach based on ontological viewpoint management. It dynamically generates is-a
hierarchies according to the viewpoint of users from an ontology using
single1 http://www.obofoundry.org/
2 http://www.hozo.jp
3 It is represented by two kinds of is-a relationships: (essential) is-a and (non-essential) IS-A.
inheritance. The main strategy is composed of: (
The rest of this paper is organized as follows: In section 2, we introduce dynamic is-a hierarchy generation according to viewpoints. In section 3, we discuss implementation of the framework as an additional function of Hozo. In section 4, we shows its application to a medical ontology for dynamic is-a hierarchy generation of disease. In section 5, we discuss related work. Finally, we present concluding remarks with future work. 2
We implement the dynamic is-a hierarchy generation system as an additional function of Hozo [4]. Fig.1 shows an example of ontology defined using Hozo. Ontologies are represented by nodes, slots and links. The nodes represent concepts (classes), is-a links represent is-a (subclass-of) relations, and slots represents part-of (denoted by “p/o”) or attribute-of (denoted by “a/o”) relations. A slot consists of its kind (“p/o” or “a/o”), role concept, class restriction, cardinality. Roughly speaking, a slot corresponds to property in OWL and its role name represent name of property. Its
Class restriction (=owl:someValuesFrom ) cardinality (=owl:cardinality) p/o slot represents a part‐of relation (=rdf:Property) a/o slot represents an attribute‐of relation (=rdf:Property)
Role concept (≒property name )
Fig.1 An example of ontology defined using Hozo. class restriction and cardinality correspond to owl:someValuesFrom and owl:cardinality respectively. Their restrictions refer to other concepts which are defined elsewhere. However, semantics of Hozo’s ontology includes some concepts related to role which are not supported in OWL because it is designed based on an ontological theory of role [5]. While we have designed three levels of role representation model in OWL to capture the semantics level-wise [6], we use the simplest model described above in this paper.
In the target ontologies, concepts (classes) are defined by several slots which represent properties and restrictions for them. These definitions are inherited from super-concepts (super-classes) to their sub-concepts (sub-classes) along with is-a links. Furthermore, in some sub-concepts, some inherited definitions are specialized according to is-a hierarchies of concepts which are referred by their restrictions. For example, bicycle in Fig.1 inherits front-wheel from Two-wheeled vehicle and its classrestriction could be specialized from Wheel to Bicycle-wheel. This research focuses on these characteristics of is-a hierarchies and considers an approach to reorganize isa hierarchies of concepts based on is-a hierarchies of concepts referred to by their definitions.
The user selects something as an aspect from the definition of the target concept for
dynamic is-a hierarchy generation (see. Fig.2(
The user selects a base hierarchy from hierarchies of concepts which the aspect is
referring to (see. Fig.2(
The system defines new sub-concepts of the top concept of target concepts by
specializing the definition of the top concept according to the class restriction
selected as an aspect and base hierarchy (see. Fig.2(
The scope of a transcription of the base hierarchy could be managed by specifying the number of the target layers rather than to use all concepts of the base hierarchy for transcription.
The system reorganizes the is-a hierarchy by comparing the original is-a hierarchy and the transcriptional hierarchy generated in step 3. The system compares the subconcepts of the target concept (we call them existing sub-concepts) with the concepts on the transcriptional hierarchy (we call them generated sub-concepts) according to the aspect and the base hierarchy. When an existing sub-concept’s definition specified by the aspect subsumes the definition of a generated sub-concept, the existing subconcept is classified into sub-concepts of the generated sub-concept. If an existing concept is classified into sub-concepts of multiple generated sub-concepts, the existing concept is classified into the lowest sub-concepts. As a result, all existing concepts are classified into sub-concepts of the generated concepts on the transcriptional hierarchy according to the aspect and the base hierarchy4. Through the above four steps, the system can dynamically generate is-a hierarchies by reorganizing existing sub-concepts according to the transcriptional hierarchies of base hierarchies. 4 The result of reorganization corresponds to the result of classification using DL-reasoner while it is implemented by procedural ways in Hozo.
Original is-a hierarchy of “disease”
The aspect (
disease
Refer to
(
abnormal state The base hierarchy (is-a hierarchy of “abnormal states”) vessel abnormality blood abnormality stenosis infarction hyperglycemia
Angina Myocardial Stroke diabetes
Classificationin fianrcttoio snub‐concepts (
Although DL-reasoners can classify classes (concepts) automatically by reasoning,
the result of classification is only an is-a hierarchy which is determined uniquely
according to the definitions of the classes. Therefore, it is different from our dynamic
reorganization according to the users’ viewpoints. DL-reasoners can generate a
different is-a hierarchy only when class definitions in the ontology have changed.
§１ In the Case of that an Is-a Hierarchy is Selected as a Base Hierarchy
As an example, we consider a dynamic is-a generation of diseases which is defined in
terms of several slots such as “main pathological state”, “abnormal object” and so on
(see. Fig.3). Here, we suppose the user selects the class-restriction of “main
pathological state” as an aspect (Fig.3(
First, sub-concepts of “disease” such as “disease with vessel abnormality as main
pathological state” and “disease with blood abnormality as main pathological state”
are dynamically generated by specializing the definition of “disease” according to the
class restriction selected as the aspect and the base hierarchy. After repetitions of
generations of sub-concepts, the transcriptional hierarchy of “disease” is obtained
(Fig.3(
diabetes
relationships between them. In the case shown in Fig.3(
§２ In the Case of that an p-is-a Hierarchy is Selected as a Base Hierarchy
In the next example, we suppose the user selects the class-restriction of “abnormal
object” as the aspect and the p-is-a hierarchy of “human body” as the base hierarchy
for a dynamic is-a generation of disease in the same ontology with the previous
example (Fig.4(
In the property inheritance mechanism of ordinary is-a relationship, when a superclass and its sub-class have the same slot, the class restriction of the sub-class’s slot must be a sub-class of the super-class’s one as well. However, in some case, the class restriction of the sub-class’s slot must be a part of the super-class’s. For example, when <disease of a pulmonary valve is-a disease of a heart>, both “disease of a pulmonary valve” and “disease of a heart” have a slot of “site of the disease” and the class restriction of the former must be a part of the latter, that is <pulmonary valve part-of heart>.
To cope with such cases, on the basis of our latest theory of roles, we introduced “p-” operator in Hozo which automatically We write “p-heart” generates a generic concept instead of “heart”． representing all the parts of the entity to which the operator is attached. The operator enables parts to be inherited by ordinary
Fig.5 An example of usage of p-operator. property inheritance mechanism. In the case of Fig.5, for example, we write “p-heart” instead of “heart”, then the slot of its subclass inherits not subclass of “heart” but its parts. When p-X is used, Hozo automatically generates a generic concept representing all of the defined parts of X including all parts which have X as their ancestor. This is valid because each part is a subclass of “X’s parts class” which coincides with p-X. According to mereology, the theory of parts, p-X includes itself which is not the very X as an entity but X as its part.
On the basis of this theory, Hozo automatically generates is-a relationships between p-X such as <p-pulmonary valve is-a p-heart>. As a result, an is-a hierarchy of p-X is generated according to part-of hierarchy of X. The is-a hierarchy of p-X is called p-is-a hierarchy5 and could be selected as a base hierarchy for a dynamic is-a generation.
In the case of Fig.4, since the class restriction of “abnormal object” is “p-human
body”, we can select it as an aspect and p-is-a hierarchy as a base hierarchy for
dynamic is-a generation. Then, sub-concepts of “disease” such as “disease with
pnervous system as abnormal object” and “disease with p-circulatory system as
abnormal object” are dynamically generated according to the aspect and the base
hierarchy. As a result, the transcriptional hierarchy of “disease” based on p-is-a
hierarchy of “p-human body” is obtained (Fig.4(
In addition to these examples, we can select is-a hierarchies which are generated using the proposed method as a base hierarchy to generate another is-a hierarchies. That is, our dynamic is-a generation could be executed recursively.
The dynamic is-a generation is applicable to reorganizations of a portion of an is-a hierarchy of “disease”. For example, we can select a middle-level concept (e.g. “disease of heart” as the target concept for the dynamic is-a generation.
In these ways, we can dynamically generate is-a hierarchies of diseases according to the selected aspects and base is-a hierarchies from various viewpoints.
We implemented a prototype of dynamic is-a hierarchy generation system as an extended function of Hozo. The system was developed using HozoCore, which is Java API for ontologies built using Hozo, and Hozo OAT (Ontology Application Toolkit), which is Java library for GUI of ontology-based applications developed using HozoCore.
The new function consists of three modules: is-a hierarchy viewer, viewpoint setting dialog, and dynamic is-a generation module. The is-a hierarchy viewer shows an is-a hierarchy of an ontology in a tree representation (Fig.6). The user selects a target concept on the is-a hierarchy for a dynamic is-a generation. The definition of the selected target concept is shown on the viewpoint setting dialog. In the dialog, the user selects a viewpoint for the dynamic is-a generation by choosing an aspect, a kind 5 To deal with p-is-a hierarchies in OWL, we can represent them by some design pattern of ontologies such as SEP triple proposed by Udo Hahn and his group [8].
H o z o o n t o l o g y O W L o n t o l o g y
H o z o ‐ o n t o l o g y e d i t o r
H o z o C o r e
Is‐a hierarchy viewer
Viewpoint setting dialog
Dynamic is‐a hierarchy generation module of base hierarchy and the number of target layers of a transcriptional hierarchy according to his/her interests. The dynamic is-a generation module generates an is-a hierarchy according to the specified viewpoint. The generated is-a hierarchy is shown on the is-a hierarchy viewer and could be saved as an ontology file.
While the target of the system is an ontology in Hozo’s format, it also can support an ontology in OWL because Hozo can import/export OWL ontologies. When the generated is-a hierarchy is exported in the OWL format, its generated sub-concepts in the transcriptional hierarchy are represented by owl:equivalentClass which have restriction on properties selected as the aspect. The user can reorganize the is-a hierarchy of the exported OWL ontology based on the transcriptional hierarchy by reasoning using a DL reasoner.
We applied dynamic is-a hierarchy generation system to a medical ontology which we are developing in a project supported by Japanese government [7, 9]. In our medical ontology, diseases are defined by specifying typical disorder roles, such as pathological condition, symptom, played by abnormal state. Fig.7(a) shows the framework to define diseases. Its disorder roles are represented as slots with classrestrictions for constraining slot values. These slots are used as aspects for dynamic generation of is-a hierarchies of diseases.
For example, when we select the pathological condition of disease as an aspect and the is-a hierarchy of abnormal state as the base hierarchy, the is-a hierarchy of disease is generated (Fig.7(c)). In the generated is-a hierarchy, concepts which have names represented by “disease which has X as pathological condition” (e.g. disease which has abnormality in the structure as pathological condition) are sub-concepts generated through the dynamic is-a hierarchy generation. Their concept names are automatically determined by the system using a template. Exiting sub-concepts are reorganized as sub-concepts of them. For instance, acute cardiac infarction is classified into a sub-concept of disease which has cardiac infarction as pathological condition. From the generated is-a hierarchy, we can understand diseases according to the classification of pathological conditions. (a) The framework to define diseases.
(b) The class-restriction selected as an aspect.
The original is-a hierarchy of “disease”
The generated is-a hierarchy (c) The generated is-a hierarchy.
On the other hand, when we select the object of pathological condition as an aspect and p-is-a hierarchy of the human body as a base hierarchy, the system generates the is-a hierarchy of disease which is similar to the part-whole hierarchy of the human body. For instance, acute cardiac infarction is classified into a sub-concepts of disease which has a pathological condition in the myocardium.
Moreover, we have developed a medical information system to consider how the dynamic is-a hierarchy generation function can be used in other systems [10]. It is used as an index for semantic navigation in the system. We also performed an informal evaluation of the implemented system in a workshop 6 and received 6 The number of participants was about 25. It includes not only the members of the medical ontology development but also others who work in the medical domain. favorable comments from medical experts. They especially liked the dynamic is-a hierarchy reorganization, which is the first solution to the multi-perspective issues of medical knowledge in the world.
In order to avoid multiple-inheritance, some researchers took an approach that they developed ontologies using single-inheritance and reorganized them by reasoning using a DL-reasoner [11]. It corresponds to reorganization of is-a hierarchy based on a transcriptional hierarchy in step 4 of the proposed method. However, the approach needs that the transcriptional hierarchy is developed in advance while it is dynamically generated by the system in the case of the proposed method.
Faceted Classification is used to represent classifications from multipleperspectives. In the Semantic Web, some researchers proposed Faceted Search for semantic portals [12, 13]. They use Faceted Classification according to the user’s choose of facets from the definition of ontologies to provide user-centric semantic search. In order to formalize the Faceted Classification, Bene Rodriguez-Castro proposed an ontology design pattern to represent Faceted Classification in OWL [14]. Although the proposed method use a similar technique to Faceted Classification for transcription of a hierarchical structure, it is different from Faceted Classification since we focus on considerations of ontological meaning of generated is-a hierarchies. Introduction of a p-is-a hierarchy is one of the results of the ontological investigations.
However, there are some rooms to ontological investigate on a method of dynamic is-a hierarchy generation. For instance, we need to investigate is-a hierarchies of roleconcepts and role-holders [5] while this paper concentrated on is-a hierarchies of basic concept (normal type). Dynamic is-a hierarchy generation based on more complicated viewpoints is also important subject to be considered. For example, we are considering viewpoints to cope with a new disease model based on an ontological consideration of causal chains [9]. Because the latest version of our medical ontology based on the new disease model has more rich definitions than previous one, it would support more complicated viewpoints for dynamic is-a hierarchy generation based on causal chains in diseases. We believe these ontological considerations would clarify the feature of the proposed method.
In this paper, we discussed multi-perspective issues of is-a hierarchy construction in ontologies and proposed a method of dynamic generation of is-a hierarchies. The main idea is reorganization of is-a hierarchies from the original ontology according to viewpoints of users. The proposed method was implemented as a new function of Hozo, and we applied it to our medical ontology for a preliminary evaluation. As a result, we confirmed that it could generate several kinds of is-a hierarchies from the medical ontology according to the user’s viewpoints. As of April 12, 2011, 6051 diseases have been defined in the medical ontology by 12 clinicians, and these definitions are currently being refined. The demonstration of the dynamic is-a hierarchy generation is available at http://www.hozo.jp/demo/. The function is also supported by the latest version of Hozo.
Currently, we are improving the dynamic is-a hierarchy generation function in order to support more detailed disease model which we proposed in [9]. We are also developing a dynamic is-a hierarchy generation system for OWL ontologies using OWL-API while it is partly available through OWL import/export function of Hozo. Future work includes ontological investigations of the proposed method and evaluation of the system through application of several ontologies.
A part of this research is supported by the Ministry of Health, Labour and Welfare, Japan, the Japan Society for the Promotion of Science (JSPS) through its “Funding Program for World-Leading Innovative R&D on Science and Technology (FIRST Program)”, and Grant-in-Aid for Young Scientists (A) 20680009.