Representations of domain knowledge range from those that are ontologically formal, semantically rich to those that are ontologically informal and semantically weak. Representations of knowledge are important in many tasks, one of which is the support of travel around information spaces through the identification and linking of concepts in a field. In this paper we explore how representations of ontologically informal, semantically weak domain knowledge as captured by the Simple Knowledge Organisation System (SKOS) can enable a system to take advantage of the large number of existing ontological representations to support semantic linking of Web based information and thus facilitate information travel.
Ontology Language (OWL)5. RDF provides a base vocabulary for describing resources and ad-hoc relationships between them. OWL is an extension of RDF and provides a vocabulary for building ontologies. Ontologies are used to encode knowledge about a particular domain in the form of the entities and the relationships between them. An ontology language like OWL has well defined semantics, this facilitates computational interpretation of statements expressed in OWL. It is hoped ontologies will provide the content to annotate documents on the web, this Semantic Markup provides a mechanism for computers to interpret a document’s contents.
Whilst ontologies offer great promise, the library and information sciences
have a history of using knowledge artefacts, known as Knowledge Organisation
Systems (KOS), to classify and index documents [
The Sealife project7 seeks to develop a series of browsers in the context of
the Semantic Web and Semantic Grid [
The work presented here relates specifically to gathering the background knowledge, in the form of ontologies and controlled vocabularies, to support document navigation. We investigated the use of a KOS over a formal ontology in providing background knowledge to support document navigation in a Semantic Web browser like Sealife. The Conceptual Open Hypermedia SErvice (COHSE) application is an implementation of a Sealife browser. We extended COHSE to support SKOS style representation of knowledge; COHSE had previously worked solely with OWL ontologies. We discuss the advantages of using SKOS over OWL
in this application scenario and demonstrate the application in travelling the web for resources on infectious disease using a KOS developed in SKOS. 2
Navigation via hypertext is still the mainstay of the current Web [
Consider this simple example; A document on the Web contains information about Polio and the term Polio is marked up as a hypertext-link to another document on the Web. This link is hard coded and is unary to the other document, the target of this link is chosen and controlled by the owner of these documents. If the document had some Semantic Markup, from a Knowledge Base (KB) about human diseases, a Semantic Web browser could interpret its content and offer services (or links) relating to Polio. By having background knowledge about Polio, the browser also identifies Polio as having related terms such as Polio disease, Polio virus, Polio symptoms, Polio immunisation, and Polio treatment, again it could offer appropriate service based on these terms. Having this background knowledge available to the Web browser offers potentially new and beneficial avenues for exploration of information relating to Polio.
This example is just one of many potential uses of Semantic Web
technologies. Despite its potential, uptake of semantic technologies on the web has been
slow due to the large cost associated with developing KBs [
browser, Sealife has added new functionalities to both GoPubMed and COHSE and investigated the use of Semantic Web browsers with use cases from the life sciences. 3
Conceptual Open Hypermedia supports the construction of hypertext link
structures built using information encoded in ontologies [
The COHSE architecture has been demonstrated in several fields, including
the GOHSE [
Much work has already been published on the requirements for a Semantic Web
browser in the context of document navigation on the web [
Ontologies, like those represented in OWL, have been used to meet the requirements of navigation in a Semantic Web browser. However, by restricting these applications to ontologies, we are unable to take advantage of semantically weaker structures, like thesauri, that also contain knowledge in structure that have proven useful for navigation in the library sciences. On deeper inspection we believe that semantically weaker structures actually provide a better “fit” for the kind of navigation we want to support with COHSE. This is not to say that we reject the use of formal ontologies or OWL within the COHSE system. Rather, we have introduced an additional level of abstraction into the knowledge model that allows us to implement the underlying knowledge structure using whichever formalism we require (OWL, SKOS, etc.) while presenting a unified interface to the user.
OWL is currently a standard recommendation by the W3C for representing
ontologies on the web. OWL provides a language for defining formlised
conceptual models. It comes with well defined semantics that enable precise
interpretations of these models by logic based reasoners – reasoning can then be used
to help build and maintain consistent ontologies. These features are key for the
development of a Semantic Web that supports automated machine-to-machine
processing of data. OWL enables the data to be expressed explicitly and with
little or no representational ambiguity. However, such an approach comes at a
cost. OWL ontologies are complicated to build and maintain and require various
levels of expertise in not only the ontology’s subject matter, but also certain
aspects of philosophy and logic [
Given the fact that Semantic Web browsers are intended for human interaction with the data, the relationships between concepts in the KB need not be as strict as those represented in OWL. For example, consider navigating documents following the specialisation of some concept. Whilst the strict sub-super class relationships in OWL can support this navigation, the semantics of this relationship are not necessary for the task. Restricting these applications to OWL ontologies alone means that modeling specialisations of concepts, that we as humans may find completely intuitive, becomes difficult or even impossible in OWL. In OWL, a subclass relationship means each instance of the subclass is also an instance of the superclass and as such, is a strong statement about two classes of instances. Specialisation and generalisation relationships need not be describing this particular relationship between instances.
The examples given for polio above are a good example of this point. For a
further demonstration of this point we can look at a terminology that is
commonly used in medicine – the Medical Subject Headings (MeSH) [
The contrast in semantics means that conversions from MeSH into OWL (with the broader/narrower relationships represented as subclass axioms) are not possible without easily misinterpreting the intended semantics. From a navigation point of view, these weak semantics are acceptable and mean the background knowledge can contain relationships between terms that may be unpalattable to represent ontologically, but perfectly reasonable when we make the semantics of the relationship similar to that of broader/narrower. The looser notions of broader/narrower as found in vocabularies or thesauri provide the user with weaker statements amounting to “this entity has something to do with another more specialised/generalised entity”. This enables us to create knowledge bases that contain the kinds of relationships between terms, that we as humans find more intuitive in a context of navigation and retrieval.
One could argue at this point that a possible solution would be to simply introduce new vocabulary into OWL that represents the broader/narrower relationships, but this introduces a requirement that the application machinery be aware of these, resulting in a non-generic solution. SKOS does provide a recognised standard for representing these relationships often found in KOS (broader/narrower/related). Although these relationships may not have the precise semantics that comes with OWL’s relationships, in this context, the looser interpretation is more appropriate to the task in hand. By enabling COHSE to work with both OWL and SKOS style representations, we lose nothing and benefit from being able to introduce a wider range of knowledge bases that do not convert into OWL ontologies in a straight-forward manner. 4.1
The Simple Knowledge Organisation System (SKOS) is a proposed standard for representing and publishing classification schemes, thesauri, taxonomies and subject heading systems on the web. It is currently under development as part of the W3C Semantic Web Deployment Group (SWDWG) and a SKOS specification has been published as a W3C last call 12.
The SKOS model can be used to structure and represent knowledge artefacts that contain statements about concepts and the relationships between them. Thesauri, taxonomies, classification systems, subject heading and other similar artefacts are considered to be different types of concept schemes, and often share many features in common. These features are primarily in the form of a lexical resource along with some semantic relationships between each resource. The semantic relationships between resources are typified by synonym, hypernym, hyponym, antonym, broader, narrower, related and so on.
One of the main goals of SKOS is to provide a simple and extensible model that can be used to express these kinds of relationships between resources. SKOS is designed to be extensible and modular. Central to SKOS is the core vocabulary deemed sufficient to represent the common features found in concept schemes. A concept can be considered any unit of cognitive thought, each concept has a set of properties which include the lexical forms that describe it and its relationship to other concepts. The class skos:Concept has instances which are the resources being described in the concept scheme. Each resource will typically have some preferred label (skos:prefLabel), some alternative labels which are considered synonyms (skos:altLabel) and a definition (skos:definition). Concepts are organised into hierarchies using broader-narrower relationships, or linked via associative relationships.
Returning to the requirements for background knowledge in a navigation system, we find that SKOS fulfills all three. SKOS has support for representing a rich set of varied lexical information, which is useful for mapping term in documents to concepts from the SKOS. The core semantic relationships in SKOS allow us to express the informal relationships held between concepts that are 12 http://www.w3.org/TR/swbp-skos-core-guide/ deemed suitable for navigating a knowledge space. As SKOS provides a minimal set of features to support the task in COHSE, it can be used as a representation for knowledge bases that do not convert readily to OWL. 5
For a Sealife browser such as COHSE to be useful across such a diverse subject
as biology we need a system to rapidly collect the current available ontology-like
resources together, and represent them using standard Semantic Web
technologies. The biomedical domain already has a rich collection of vocabularies and
ontologies such as MesH [
Converting existing terminologies into representation suitable for Semantic Web applications can be difficult. Before SKOS, OWL was deemed the standard for representing knowledge artefacts on the Web. Conversions from the MeSH, GALEN and OBO formats into OWL exist, but this is non-trivial as many of these formats do not have a precisely defined semantics like OWL. Care must be taken when converting to OWL from formats where the semantics are weaker than those found in OWL. In contrast, the weak semantics of SKOS make it possible to convert a wide range of artefacts into SKOS, making the artefacts available to SKOS aware applications. The conversion is potentially lossy, especially when converting from OWL to SKOS, but where it is simply the structure that is necessary for the task and not the semantics that can drive sophisticated inference, such loss is acceptable. In addition, of course, with such a conversion, the original artefact still exists in its original representation. Providing both SKOS and OWL representations of knowledge bases is perfectly reasonable and should be encouraged.
Conversions into SKOS already exist for the MeSH thesaurus [
Many of the ontologies and controlled vocabularies being developed in the life
sciences are done through the OBO foundry [
Here we present the outlines for a conversion of the OBO format into SKOS. When converting OBO ontologies into SKOS we can take the relationships defined in the RO and map them to the semantically weaker notions found in SKOS to assert broader, narrower and related relationships between SKOS concepts. Here is an example of the conversion one might make when mapping ontological properties to SKOS properties.
– OBO REL:part-of -> skos:broader (e.g. finger part of hand) – OBO REL:contains -> skos:narrower (e.g. skull contains brain—ignoring the cabity for now) – OBO REL:adjacent-to -> skos:related (e.g. nuclear membrane adjacent-to cytoplasm)
Another advantage when converting properties from ontologies to SKOS is the ability to assert the inverse. Consider an ontology where Nucleus partOf cell, from an ontological point of view this implies that each and every Nucleus is partOf some Cell. The inverse, however, is not true, every Cell does not hasPart Nucleus. When converting to a SKOS model we can assert the inverse using the broader property to say that Nucleus has a broader term called Cell, which is reasonable. When navigating around documents about cells, the system could then also provide links to documents about nuclei—users interested in cells are often also interested in nuclei. 6
One example application for the COHSE Sealife browser is to provide dynamic
hyper-linking of resources from the National electronic Library of Infection (NeLI)
[
The ability to identify user groups is important. Users can range from
members of the public, molecular biologists, to clinicians and family doctors. Each
group has a different perspective on the bio-medical domain, and is therefore
interested in different kinds of information. By providing alternative vocabularies
for different users, the system can identify link sources relevant to that user and
also provide multiple targets to relevant web resources. Table 1 shows four
different user groups, some questions they might want answering and the different
kinds of target sites a Sealife browser would offer them based on user type [
Question
Targets Sites Family Doctor Tuberculosis drugs and side ef- British National Formulary (BNF) (GP) fects? Clinicians Tuberculosis treatments Public Health Observatories guidelines? (PHO) Molecular Biolo- Drug resistant tuberculosis PubMed gists species? General Public What is tuberculosis?
The system is demonstrated with a simple use case involving a news site linking to NeLI. News sites are often the first to report on disease outbreaks via news feeds. Consider the scenario where a traveller is planning a trip to Namibia, only to find an article on the BBC website16 about a recent outbreak of Polio. COHSE can provide links to relevant resources that had not been included by the original author. Such resources could include information about the polio virus, its effect on humans, vaccination information and also geographical information about the local area. A family doctor, in contrast, might use a vocabulary skewed to the their interests to link through to sites on drugs, details of symptoms and clinical presentations, treatment and local hospital facilities etc..
Figure 2 shows the COHSE system in action. The first image shows the original BBC article, this page has the links provided by the original author of 16 http://www.bbc.co.uk the page, these include links to related news stories and also a link to the World Health Organisation. Whilst these links may be useful to a user, it is easy to imagine a wide range of related information one might want to get to from this page. In this scenario the user wants information on polio vaccination. The BBC article contains no direct link so the user would be forced to use a search engine. Querying polio vaccination against a search engine would typically bring a wide range of documents that contain the word polio and/or vaccination, the user must manually filter through these documents to find the information they require. The search engine will not discriminate between news article, fact sheets or clinical guidelines, neither will it provide any indication as to who are the world authorities on vaccination information. All of this must be processed by the user which takes time and effort. There is also the problem that the search engine will miss documents that do not mention polio directly. The documents might mention poliomyelitis, which is the clinical term for the disease caused by polio virus. These documents may be vital to the user but missed because of the language used in the query.
The COHSE system can help the user in this scenario by identifying the
key concepts in the documents and offering relevant links from the current site
based on related terms and synonyms. When viewing the BBC article through
the COHSE portal the document is first processed by the COHSE DLS (see
Section 3). In this scenario the COHSE Knowledge Server (KS) has been loaded
with a SKOS vocabulary that contains terms relating to infectious disease (the
vocabulary was developed at NeLI and is represented in SKOS17 [
When the polio link is selected, the term is dynamically queried against the COHSE Resource Manager (RM). In this scenario the RM has been loaded with five web-services. These web services have been selected by experts from NeLI as being authoritative sites on information about infectious diseases. The term polio along with broader, narrower and related terms from the concept scheme are queried against the RM with the most relevant results returned as link targets to the user. In this example we see that Polio Vaccination is related to the term Polio via the narrower relationship. Polio Vaccination when queried against the RM returned links to a document from NeLI with guidelines about polio immunisation. The second screenshot (Figure 2) shows the links box offered to the user with links relating not only to polio, but to all the related terms found in the concept scheme. As the user continues to navigate around the web, the COHSE system continues to identify concepts in web pages and offer appropriate hyperlinks.
One important aspect to guiding navigation in this way is that the modality
is kept, having to switch between windows and search engine results breaks the
modality and can interfere with the task at hand [
The infectious disease vocabulary used as background knowledge for the NeLI
use case has been developed in SKOS [
The advent of the Semantic Web/Grid has brought large computational resources to bear upon knowledge intensive sciences such as biology. Despite this, user facing tools that support document orientated tasks are not yet available within this new paradigm for computing. Semantic technologies can be applied to this problem to overcome some of the limitations of the current Web’s navigational structure. We have argued that a representation with weaker semantics, such as SKOS, will enable us to exploit the wealth of ontologies, thesauri and other types of knowledge organisation schemes already existing within the biomedical domain.
Representations with stricter semantics are not always suitable for representing artefacts such as MeSH. KOS such as MeSH were created, not to make ontological descriptions of biological reality, but to aid navigation and retrieval of information about biomedicine.
The nature of formal ontologies can sometimes make it difficult to express
relationships between concepts that experts from the domain would expect to find
under some circumstances, such as in information retrieval tasks [
Funding by the Sealife project (IST-2006-027269) for Simon Jupp is kindly acknowledged. Yeliz Yesilada was supported by Sun Microsystems Laboratories.