Methodology: The paper is structured into two main parts. “Semantic Web Technology” provides a high-level description, with examples, of the main standards and concepts: extensible markup language (XML), Resource Description Framework (RDF), RDF Schema (RDFS), ontologies, and their utility in information retrieval, concluding with mention of more advanced SW languages and their characteristics. “Semantic Web Applications and Research Projects in the Biomedical Field” is a brief review of the Unified Medical Language System (UMLS), Generalised Architecture for Languages, Encyclopedias and Nomenclatures in Medicine (GALEN), HealthCyberMap, LinkBase, and the thesaurus of the National Cancer Institute (NCI). The paper also mentions other benefits and by-products of the SW, citing projects related to them.

The rise of extensible markup language (XML) and metadata standards such as Dublin Core (DC), along with the initiatives of the World Wide Web Consortium (W3C) to create a “Semantic Web,” point toward a new world of Web-based information, a world in which information will be machine understandable and machine readable. In the vision of the Semantic Web (SW) outlined by Berners-Lee et al. in a milestone article [1], intelligent search programs (also called “software agents”) are able to draw sophisticated inferences from metadata attached to Web-based information. This vision (and its underlying technology) shows great promise, though its impact has yet to be fully realized. To this end, much research work is currently being carried out and significant funds are being allocated, both in Europe and the United States.

The main goal of the SW is to add logic to the current Web, in other words, express the meaning of data, the properties of objects, and the complex relationships existing between them by a series of formal rules, which would make information accessible to machines. Machine accessibility should be understood as representing information in such a way that it is possible to make queries based on the meaning (i.e., semantics) of the data, independent of the form in which the information is presented.

Languages such as XML and other standards derived from it achieve this aim, but only in a limited way. Consider the following three XML tags:

1. <article name=“Duodenal ulcer”>

<author>Smith</author>

</article>

2. <author name=“Smith”>

<article>Duodenal ulcer</article>

</author>

3. <bibRecord>

<author>Smith</author>

<article>Duodenal ulcer</article>

</bibRecord>

All the above tags express the same meaning: an article called “Duodenal ulcer” authored by Smith. However, for a machine, the three representations are different, because the order of encapsulation (nesting) of the tags is different. The above representation has very little semantics.

XML is, however, sufficient for many applications, where structures are relatively simple and the enforcement of a strict syntax or format can be assured. Numerous examples of such representations exist. The most familiar perhaps is the display of PubMed records in XML format, but the fact is that unless the exact structure of the XML is respected, namely the document type description (DTD) specified by the National Library of Medicine (NLM), the code is not validated.

The SW approach is to add another layer on top of the XML standard, adding more meaning to the encoded information. This extra layer is represented by standards such as Resource Description Framework (RDF)/RDF Schema (RDFS) and Web Ontology Language (OWL).

The RDF standard is based on several fundamental concepts. The first such concept is “resources.” Resources are the basic objects or “things” that are to be described in the domain (e.g., articles, books, authors, electronic resources, etc.). All resources are identified through a unique, global identifier called the universal resource identifier (URI). In most applications, the URI is the uniform resource locater (URL) of a Web page, a part of a Web page (e.g., anchor URL), or a link to a document available on a Web server. However, the URI is a more general concept than a Web link—the only condition is that it uniquely identifies a resource. In a library setting, the MARC catalog record for a book, for example, could be considered a URI.

Another essential concept is “classes,” which are basically collections of objects and things with common properties. It is very important to distinguish between a class and an instance of a class or an object. For example, “article” is the class, while “Duodenal ulcer” is an object or instance in the “article” class. Classes can be further divided into subclasses; for example, the “article” class is a subclass of “Printed Resource,” which in turn is a subclass of the more general class “Resource” (which may contain both printed and electronic resources). This is similar to the concept of hierarchical classification in information science.

A third important concept is “properties.” Properties describe relations between other classes or resources. For example, the “article” class may have the properties “has_author,” “appears_in_journal,” “has_publication_date,” and so on. To some degree, this relationship can be seen as the facets of facetted classification, though the analogy should not be taken too literally.

In other words, an ontology is the attempt to formulate an exhaustive and rigorous conceptual schema (i.e., a map of concepts and their relationships) in a given domain. The most commonly cited definition has been given by Gruber: “an explicit specification of conceptualization” [5]. Sabou has provided a clearer one, with the same terms: “a shared conceptualization of a domain” [6]. The emphasis here is on the term “shared,” because all parties using the information have to agree to use the same representation. For example, when different people classify articles under some system of categories and subcategories, the categories are known and agreed on by everyone. Figure 1 provides an example of a possible ontology for the library domain.

In Figure 1, two types of elements appear: classes, represented as rectangles, and properties of classes, represented as ovals. Note that there are two types of arrows as well, corresponding to the two main types of relations allowed by RDF/RDFS. The thicker, closed arrows are “subclass of” (also called “is_a”) relationships. For example, any instance of the class “book” is a “Printed Resource” and therefore is also a “Resource.” This corresponds to a hierarchical structure between classes.

The second type of relationship is the “Domain-Property-Range” relationship, followed through the thinner, open arrows (e.g., resources have an author or have, as publication year, some instance of year). Note that properties defined for larger classes also apply to all their subclasses. For example, because the authors have specified in Figure 1 that all resources can have an author, this means that “Printed Resources” and “Electronic Resources” also have one and, hence, books and articles. However, the property “hasJournal” only applies to the subclass of journal articles and does not apply (at least in this ontology) to books or electronic resources.

First, recall that the RDFS is used to present structure, while the RDF is used to input actual information for such a structure. Consider the ontology structure from Figure 1. Definition in RDFS would include tags such as the following. The topmost tag defines the XML domain of the tags used.

<rdf:RDF xmlns:rdf=“http://www.w3c.org/…/rdf”

xmlns:rdfs=“http://www.w3c.org/…/rdfs”>

<rdfs:Class rdf:about=“Article”>

<rdfs:subClassOf rdf:resource=“PrintedResource”>

</rdfs:Class>

<rdf:Property rdf:about=“hasAuthor”>

<rdfs:domain rdf:resource=“Resource”>

<rdfs:range rdf:resource=“Author”>

</rdf:Property>

…. </rdf:RDF>

The first tag (i.e., “BibliographicResource”) may refer to an unpublished clinical report. For simplicity, suppose users already know that the author and date are at the second level of nesting in XML. A query can now be made on the above structure to retrieve all bibliographic resources that have the author Smith and appeared after 2000. This would return only the first record, the only one explicitly marked as a Bibliographic Resource, and not the second one.

For human understanding, it is obvious that journal articles are also bibliographic sources of information. However, a machine does not know this, unless its underlying representation (or ontology) contains this information, hence the benefits of more semantically meaningful representations. Furthermore, the benefits of using more semantic representations increase as the size and complexity of the information to be represented increases. The ontologies used in the biomedical domain, with hundreds of classes and properties, often entail such complexity.

It should be mentioned that RDF has limitations regarding the type of structures that can be represented by it. More powerful representation languages for the SW include: Defense Advanced Research Projects Agency (DARPA) Agent Mark-Up Language (DAML), developed with DARPA's support in the United States, and Ontology Inference Layer (OIL), developed in Europe, at the Artificial Intelligence Department of the Free University (VU) of Amsterdam. Considerably more powerful than RDFS, these two languages were later united and a “lightweight” version was proposed as a possible standard to the W3C, called Web Ontology Language (OWL). Programs also exist to “write” in ontology languages, such as Protégé, developed by Stanford Medical Informatics, and a national resource for biomedical ontologies and knowledgebases supported by NLM. Examples of the type of representations that cannot be specified in RDF/RDFS but are possible in OWL are:

• Cardinality restrictions: For example, users might want to say that a printed resource may have more than one author but only one publication year.
• Disjointedness of classes: For example, users may wish to enforce that a resource is classified either as printed or electronic.
• Other special characteristics of properties, such as “uniqueness,” a book can only have one International Standard Book Number (ISBN); “inverse properties,” for example, “is published by” applied to a book is the inverse of the property “publisher of” applied to a publishing house; or “local scope,” restricting the domain of the property to certain classes.”

Traditional SW approaches generally aim to achieve 100% accuracy, by using formal logic to capture the meaning of information beyond a point where it cannot escape correct computer interpretation. However, providing information already fully structured according to an agreed ontology can be a very labor-intensive task. In fact, most existing information comes in unstructured (e.g., free text, HTML) or semi-structured formats (e.g., MeSH). Some projects, such as NLM's Semantic Knowledge Representation Project, aim to bridge this gap by constructing programs that build semantic representations automatically based on biomedical free text, making use of existing resources.

Another problem encountered by SW proponents is the fact that ontologies used to represent information in open environments such as the Web do not generally coincide (e.g., the field “date” in one ontology may be represented by “year,” “month,” “day” in another). This problem is addressed though a line of work called ontology merging.

For example, faceted classification (originally proposed by Ranganathan in the 1930s) uses concepts such as classes and facets to allow multiple, flexible ways to label and retrieve information sources. This approach resembles the multiple classes and properties used in RDF/RDFS (though the latter protocols have the clear advantage of being grounded in formal logic). There is much revived interest in the faceted approach, and its benefits are being explored in the context of the Web. The recently reported “concise medical taxonomy” of the American Medical Association [33] explored the capacity of faceted indexing to create more concise, Web-friendly displays. However, as already known in the library community, faceted classification never really took off in real life. One of the main reasons it did not take off was that it appealed much less to human [sic!] logic than a hierarchical approach. Most of the taxonomies used today are hierarchical and, therefore, conceptually simpler. Arguably, hierarchies provide less expressivity and flexibility but have the major advantage that standards are much easier to understand and apply uniformly by all parties.

Librarians and library activities can contribute in many ways to the SW on a practical level [7, 34, 35]:

• get well acquainted with metadata standards such as DC and the various ways in which they may be expressed (HTML, XML, RDF)
• learn more about the concept of ontologies and about information representations using XML or RDF
• participate in projects that aim to use SW technology; librarian expertise will be highly valued

The authors argue that both the medical library and SW communities have a lot to learn from each other. On the one hand, the SW approach provides novel ways to represent and retrieve information. These approaches are grounded in formal artificial intelligence, which assures that the representations used are logically consistent and their meaning is also accessible to machines.

Nevertheless, the importance of medical librarians in this process is not only unlikely to decrease, but their role will become even more crucial. Developing the various ontologies and medical taxonomies cannot lead to any useful real-life applications without major input from librarians. As Soergel [36] points out in relation to WordNet, the largest and best known existing ontology, “a wonderful system whose construction would have profited from applying experience in thesaurus construction … and standard methods for classification.” He also shows that: “large and useful systems are being built with more effort than necessary,” because often the existing experience from classification science fails to be used. The authors argue that the two directions of work, so far largely parallel, are probably converging. The number of research projects under way confirms the authors' supposition.