There can be no argument that the wealth of information now available on-line has created numerous problems for information retrieval. One consequent of the information age is the promise of Digital Libraries, with on-line access not just to library catalogs, but to the documents themselves. Furthermore, progress in the text encoding community offers the potential for significant amounts of these documents to exist in some marked-up form (see, for example the TEI [30], the CES [22], the MEP [8], and the WWP [13]).

Having information about the content of a document readily available in a machine-readable format may seem to imply the ability to do document retrieval based on this content information; the community, however, has yet to deliver tools that exploit the markup in a manner that justifies the effort required to create the documents [22]. We have been performing ontological analysis in the domain of library catalog systems as part of a project endeavoring to address this problem.

Many aspects of the library ontology presented here would not seem very novel to anyone experienced with conceptual modeling and ontology building (in fact it is based on Tom Gruber’s bibliographic ontology [16]). It is, however, a radical departure from the current paradigm in library catalog systems whose foundation is quite clearly the need to place books on shelves in an orderly way. Modern library scientists may find this an over-simplified and out-dated criticism of current library access methods, but we hope to show in this paper that it is in fact still quite relevant.

The most interesting scholarly work that has been ongoing for many years in this effort is a deep and thorough analysis of the ontology of subjects [38]. Subject classification can be viewed from a variety of perspectives ranging from the purely pragmatic to the purely philosophical. The notion of a subject is so deeply ingrained in all aspects of our culture and language that this part of the work has been particularly fruitful and challenging. The bulk of this paper, therefore, will deal with the issues of the representation of subjects, classification by subjects, and subject-enhanced retrieval.

The paper begins with a discussion of the goals of the project, and the goals of the ontology work, then presents the new ontology for library card catalog systems. The remainder of the paper will be devoted to the ontological nature of subjects. Elements of this paper have been published previously in [40].

The Axiom project seeks to combine four different aspects of research into a single digital library: text encoding of non-technical (i.e. humanities) documents [8], linguistic text analysis [20], knowledge representation [41], and user interfaces [6].

The center of the project is the Model Editions Partnership (MEP), maintainers of a large repository of fully marked-up electronic versions of historical documents from the U.S. Civil War period [8]. The documents include letters, manuscripts, newspapers, and diaries, as well as the more traditional books and magazines. The markup of the electronic documents is quite deep; in addition to structural components (paragraphs, line breaks, font changes, etc.) and basic header tags (title, author, date), the body of the documents include: identification of names, dates, cross-references to other documents, events, etc. The markup is done mostly by hand, and the development of automated tools for assisting these encoders is a topic of current research.

The existence of the data in this structured form opens the door, at least intuitively, to the possibility of dramatically more expressive queries than library users are familiar with today, because it enables retrieval based on the contents of the documents. Web-based keyword searches also permit this, yet they have no (or very little, at best) semantics, and can not make the simplest differentiations between word sense usage. The markup of names is of particular interest in this regard, because in addition to being tagged as names, the tags also differentiate person names from place names, organization names, etc. In other words, we expect to find markup such as:

The city of <name type=city>Washington</name> was named after the first American President, <name type=person><firstname>George</firstname> <lastname>Washington</lastname></name>

Clearly this encoded information allows queries to be expressed that differentiate between Washington the place, and Washington the person, a differentiation that keyword search does not allow.

We wish to use the richness of the encoded data to enable multiple modes of retrieval of the documents. We feel that one key ingredient to accomplish this is expanding the "card catalog" representation beyond books to support far more expressive queries, such as "Publications written by people at Vassar College," or "Works of fiction reviewed by someone at Vassar who is interested in AI," or even, "Papers on AI published at a conference sponsored by ACM." The increased expressiveness should serve to enhance the utility of the card catalog for three basic kinds of users:

1. Users with vague notions about what the object of the search is. Library users frequently don’t know the title or author of the book they are looking for, or even if it exists. For example, someone may be interested in reading about "Formal Ontology." Such a user has no idea what the relevant titles or authors are. During a search for a publication on that subject, if none were found, the user could generalize the search to, perhaps, "Knowledge Representation." The key to generalizing (or specializing) a search in this manner is access to the subject taxonomy, which is normally not the case in existing card catalog systems. Furthermore, if the user was aware that interesting research in formal ontologies was taking place at a particular institution, the user might search for "Publications on knowledge representation written by someone at Vassar College." Again, since we are dealing with a library containing potentially billions of publications, the ability to refine a search will be critical in making the information accessible [39].

2. Users doing scholarly research. We are working with two text-encoding groups who are in the business of marking up historical manuscripts and making them available in electronic form. Their primary customers are scholars who do not have direct access to the manuscripts (because of proximity, because the manuscript is fragile, etc.), and having the data in electronic form opens up vast new possibilities. Etymologists (people who study words) would like to make queries such as "what is the date of the first manuscript that uses the word ‘cleave;’" other researchers might ask for "Books by women authors in the 18th century containing mythological characters." It would be impossible to exaggerate the implications of this technology for scholarly research in the humanities .

3. Users with vague recollections of the object of the search. It is not uncommon for a library user to be searching for something they read or heard about at some time, but don’t remember exactly what it was. The more information about each publication that can be provided, the better the chance that something the user remembers is actually recorded. For example, "The paper on KR written in the early ‘80s by someone at BBN," or, "A paper on planning at an AI conference in Florida." This category was, in fact, the initial motivation for this research [38].

Drawing from these user types, we have developed a new paradigm for library search and retrieval, one that tightly integrates browsing and querying of an information space that represents more than simply books, but the people, places, events, etc., that are useful background information for supporting the retrieval process.

Every library card catalog system operates on the basic assumption that the object of a query is a document. When one types "FIND SUBJECT ONTOLOGY" in a library card catalog system, that system understands this to mean, "find all books whose subject is ontology." Our new paradigm does not change the abstract interaction; the result of any search will still be a book (or, more generally, a document). The basic revelation of this new paradigm is the recognition that in a digital library, the object of a specific query may not be a document, but a person, place, event, or other kind of object. These other objects, which represent what we term the background knowledge, extend the information space defined by the catalog and add an important step to the normal interaction between a user and the catalog system: browsing. We expect users to engage in an iterative process of querying for information, using the query results as a starting point for browsing, and using their browsing results to guide them in making new queries. This process repeats until the desired documents are found. An example is provided in Section 2.4 below.

This new paradigm requires a representation that allows for the background knowledge to be found, either through queries or browsing, and we have developed a simple ontology for supporting this new style of interaction. We present the ontology here only briefly; a more complete description can be found in [38] and [41].

Figure 2-1: Some of the concepts in the digital library ontology.

Our ontology consists of five principal classes: documents, document modalities, objects, events, and subjects. The objects include people, places, companies, organizations, etc. The events include conferences, meetings, wars, battles, etc. A subset of the taxonomy for this ontology is shown in Figure 2-1. The subjects are described later.

Individuals of the non-subject classes have numerous relationships through which they can be connected. A Person can be the author-of any document, and the mere existence of such a link implies that the person is also an author. Organizations (such as "AAAI") can sponsor events (such as "AAAI-97"), people can be members of organizations ("Lincoln was a member of the U.S. Government"), and can participate in events ("Sherman participated in the burning of Atlanta"). These links or relationships provide the basic structure for browsing the information space. In our interface, when looking at information about an object (e.g. "Lincoln"), clicking on a link (e.g. member-of) will move you to the object linked to (e.g. "U.S. Government").

It should be noted that we do not wish to create an encyclopedia here, only enough information and background material to enhance the on-line browsing process and make the sheer magnitude of digital library information accessible, and useful.

Instances of documents, events, and objects may be classified by subject. For a person, being classified by a subject indicates interest in that subject. For an organization, it means an official association with the subject, not necessarily simply the propagation of the interest of its employees or members (Chris is interested in cars, but Vassar College is not). Subjects should be propagated along certain relationships, but that will be described later.

Finally, we also represent the tags in the markup, and relationships between them from a semantic (not syntactic) basis. We augment the existing tagset with other "virtual tags" (tags that are not present in the markup but can be derived automatically) in the representation. This allows us to process a variety of background information. We have split our background information ontologies, which are fairly simple objects and rules regarding the types of texts that are encoded, into different subject areas as well.

We believe the principal advantage our approach offers is the improvement of searching and browsing for information retrieval. The added knowledge, resulting from both a principled approach to ontology development and the background information that ties the meanings of the tags together, makes it possible to more precisely specify a query.

Imagine, for example, a scholar who is motivated by recent events to research how commonplace it was during the civil war period in American history for government officials to mention government business in their personal correspondence.

The historian would begin by finding a digital library that includes marked-up versions of historical documents, such as the MEPs Civil War era documents [8]. He would then enter the query, "Find all government officials during the years 1860-1865," and be presented with a list of all the government officials the system knows about. The system knows about government officials because the markup includes tags that identify government documents, and in one of our historical ontologies we have a rule that says, "The author of a government document is a government official." There are similar rules for names that appear in tags that signify senders, recipients, signatures, etc. The system can also infer dates of service for these officials from the dates of the documents.

The power here comes from the fact that it is fairly easy to specify such rules, and then capitalize on the data already in the marked-up documents. Thus, rather than enumerating all the government officials during the Civil War, we specify the rules and let the system gather that information for us. We have found an enormous wealth of information in these marked-up documents that can be exploited in this way, such that the information itself is more accessible. Note that, of course, this information may be incomplete, e.g. if the library does not include documents in which a particular government official is mentioned in the right context (i.e. in an author, recipient, etc., tag), the system will not know that person is a government official.

The user's next step is to pick a person from the list of government officials, and display the information the system knows about that person. This will reveal to the user the kind of information the system keeps about individual people, or in this case about government officials. The user will find links called "author-of, recipient-of, sender-of, member-of," and many others. From these links, this particular historian would check the "author-of" link, and be presented with a list of all the documents this person was the author of. The documents are listed along with their most specialized parent (i.e. the parent concept lowest in the concept taxonomy), so the user will see a list of document titles and document types, for example:

The user now selects LETTER-32, and sees all the information that the system keeps about such a document, including links like "has-author" and "has-recipient," and all the parent concepts (the list view above reveals only one parent, but most instances will have many parent concepts inherited down through the taxonomy from all its immediate parents). LETTER-32 is an instance of a LETTER, PERSONAL-DOCUMENT, HISTORICAL-DOCUMENT, DOCUMENT, MEP-DOCUMENT, and several other concepts. The user notices the existence of the PERSONAL-DOCUMENT concept, and selects it.

All concepts in the ontology have brief natural language descriptions, so the user will see something like what is below. This information is pruned a bit from the actual output:

The user now has learned all the information he needs to form a more specific query, which is essentially, "Find all the PERSONAL-DOCUMENTs written by GOVERNMENT-OFFICIALs in the years 1860-1865 whose recipients were not GOVERNMENT-OFFICIALS." The result of this query may be quite large, but in fact is precisely what the user is looking for, and finishing the research will involve reading through these documents.

Eventually the scenario above would be augmented by a powerful interface that assists users with the query language. Until that time, and probably even after it, we expect that trained librarians who understand the representation and the query language will assist users with this system. Note that, of course, modern communication technology, from phones to chat rooms to 3D web spaces, imply that the human assistance does not require proximity, and the goal of making these new digital libraries accessible to the world is still achieved.

There are many example user scenarios that better detail the extraordinary advantages we believe can be gained by applying knowledge representation and ontology techniques to digital libraries of marked-up texts in [41].

Perhaps the most interesting part of the scholarly aspect of this research has been a study of the history of classification, and in particular the use of subject classification in libraries. One could easily spend a lifetime trying to catch up on everything that has been written on the idea of classification. The history of this area, as with all histories, is ripe with colorful characters, personality conflicts, love triangles, deep political intrigue, valiant and selfless acts for the advancement of knowledge, and many other factors which, in addition to being interesting, are also relevant to understanding the meaning of any piece of work. This particular part of the research was always done with clear project goals in mind, and frequently stopped well short of a even a substantial survey.

One of the most useful tools in ontology building is a dictionary, and when researching the ontology of common terms, it is often helpful to study the etymology of the terms themselves.

The English word subject derives through Latin from Ancient Greek, where it is the combination of the words "to throw" and "under." The OED lists over 20 definitions of the word, which range from the idea of a vassal or minion (i.e. "a subject of the crown"), to the metaphysical definition of substance, to the "subject matter" of art and science (i.e. "Mona Lisa was the subject of this painting"), to the grammatical role of words in a sentence, to the use in Logic of an object which is predicated, to the theme of a literary composition. [31]

The English word topic has a similar etymology, deriving also through Latin from Ancient Greek. The original Greek meaning of the word is a place, or a locality (hence the use in "topical ointment").

Finally, the word about, which is normally used to relate books to their subjects, derives also from a spatial etymology meaning "near" or "around" (such as "scars about the face").

The sense we are interested in for these words is believed to have been first brought into usage by Aristotle, who envisioned topics as "places in which a rhetorician might look for suggestions treating his theme" [31]. Subjects were things that were placed under (thrown under) these topics. They would be the very things the rhetorician might be looking for. This implies that subjects are not the same as topics; one is a place and the other is something located at that place.

It is pure conjecture, but nevertheless reasonable to infer, that the very thing Aristotle was thinking of here was a library, and that therefore the importance of organizing knowledge by subject dates back to him.

Although undergoing steady development throughout the ages, arguably the next most significant event in the history of classification after its invention by Aristotle was Francis Bacon’s attempts to organize and represent all the knowledge gained by human science [3]. Bacon’s framework was hierarchical (a "tree of knowledge"), and drew on efforts in individual sciences to classify the natural world. Bacon’s work was probably not notable in itself other than being considered the first real attempt to perform this kind of analysis.

A more impressive effort, while clearly based on and inspired by Bacon, was the well-known Encyclopedia by Diderot and D’Alembert in the mid eighteenth century [11]. Although built on work done previously, it was the first encyclopedia. In order to organize the knowledge in this work, Diderot (primarily, it is believed) and D’Alembert created a much more detailed and sophisticated tree of knowledge than Bacon’s.

The Encyclopedia’s tree of knowledge is significant in many ways. We are interested specifically in two. First, it represents an entirely practical effort. The tree was clearly meant as a contribution to the development of the philosophy of science and the classification of knowledge; however, the effort to devise the tree was guided by the need to organize the Encyclopedia. Many deeper philosophical issues were passed over when they were not relevant to the contents of the work itself.

Second, the work of Diderot was well known by an American Francophile, who used it to guide the organization of his own extensive personal library. This American, Thomas Jefferson, was the third president of the United States, and the founder of the U.S. Library of Congress (LOC), whose initial contents were precisely those of Jefferson’s library.

This is by no means a claim that the LOC was the first to organize material by subject. Indeed, evidence points to attempts at such arrangements throughout history. The first surviving record of a subject catalog was in 1483, in Melk, Austria [37]. Various other personal and public libraries were organized by subject since then. The turn of the 19th century, which is roughly when the LOC was created, is about the time that subject classification of books began to be an accepted practice in libraries around the world. Until these changes were made, books were cataloged alphabetically by author, by title, or by the first lines of text.

Libraries began to be more dependent on accurate subject classification, and while all early classifiers were students of the philosophy of science, and therefore used the work of Bacon, Diderot, D’Alembert and their scholarly descendants to guide them, by the mid-nineteenth century library classification had become an art of its own, which for practical reasons had to separate itself from philosophy. The main reason for this separation was that the subject catalog was driven by a single goal, the placement of books, whereas the philosophy of science clearly had other desires. "Classification," said John Cutter, developer of the Expansive Classification system that was to be the primary influence on the American LOC standard, "is the process of equating subjects with the topic contents of books, rather than with the logically derived elements of the general classification of knowledge." [29]

To further drive a wedge between philosophy and the new library science, many scholars objected to the use of a subject catalog. W. Stanley Jevons believed that such an organization, as opposed to an author-arranged index, supported students who didn’t know their literature well enough [24]. He claimed that students should first be educated to know who was doing the important work in a field of inquiry. The well known philosopher and logician Augustus DeMorgan also strongly objected to the idea of a "classed catalog," believing it to be too much, "one man’s theory of the subdivision of knowledge, and the chances are against it suiting any other man." [9]

Library classification ensued with its overriding goal still to place books on shelves. Whether intentionally or not, this clearly links back to the etymology credited to Aristotle. Ainsworth Spofford, director of the LOC during the mid-19th century, and a major historical figure in the development of library subject classification, wrote, "Places in the library are subjects themselves." [29]

Further study of classification revealed that the most important part of the process was capturing rules to guide the classifier. The earliest evidence of recorded rules for subject classification is attributed to Adrien Baillet, who developed classification rules for indexing the private library of Chretian-Francois de Lamoignon in the late 17th century [37]. Accuracy was not considered as important as consistency. Spofford’s system was not even intended to be exploited by library users, but by trained library assistants. [29]

Probably the most familiar and extensive work on library subject classification were those of Dewey (the decimal classification system) [10], Hartwig (the Halle Schema) [28] and Cutter. Dewey and Cutter each added two important points to their systems: the ability to add and remove categories, and the use of a relative index. The former was deemed critical as new fields of science and inquiry come and go, and the latter was important because the growth of library collections over time meant they could not guarantee that a certain subject would always be able to fit in a certain place. This gave libraries the ability to move the whole collection without requiring a recomputation (by humans) of the whole index. These two factors, plus the incredible care with which Dewey and Cutter approached the actual contents of their subject hierarchies, are the main reasons why both systems are still in use today, over a hundred years later.

Library classification is a field, a profession, requiring extensive training and experience. It would not be unreasonable to claim that library scientists have created the first formal ontologies to be used in systems (these systems began as printed indexes, then cards, and now databases). These subject catalogs represent enormous amounts of human labor spread out over two centuries. The accomplishment is nothing short of impressive, representing an essential blend of philosophy, analysis, and practical compromises. We examine here, in slightly more detail, the LOC Classification system (LCC) to gain further insight into the difficulties of representing and using subjects.

To begin with, the LCC is a hierarchical system. Subject categories have sub-categories, and sub-categories may have sub-sub-categories, etc. This hierarchy, perhaps due to its direct correlation with the physical location of books (but perhaps not as discussed below), was considered a mereological hierarchy, "Subcategories are more like parts of their parent categories than subsets." [28]

The design of the system, initially, was not so much meant for retrieval of information, but a way to locate books on similar topics near each other. Again, we have a throwback to the etymology of locality. This organization supported physical browsing better than an author or title-based approach, something which surely any library user has experienced. If a book’s subject and location were, in fact, the same thing, and given that a book can not physically be in more than one place at a time, it is no surprise that all library classification began with a, "one book, one subject," paradigm.

This created obvious problems, not just for books that contained information on more than one subject, but for books whose single subject seemed to be an amalgamation of two existing categories. Early classifiers noted the difference between intrinsic sub-category relationships, such as, "zoology falls under natural history," or, "crops fall under agriculture," and extrinsic relationships, such as, "politics of agriculture," being under either agriculture, politics, both, or something else.

Eventually, Charles Martel of the LOC devised a strategy for dealing with extrinsic relationships between subject categories, finding that there were a few (politics, history, education), that cropped up far more often in combination with others. His strategy involved the use of aspects or facets of a subject. Each subject category, and potentially each subdivision thereof, had seven facets: serials, periodicals, dictionaries, etc.; philosophy and theory; history; general works; law and regulations; study and teaching; and the subdivisions. Thus, a book on the history of Chemistry would be classified under the history facet of the chemistry subject. A book on the theory of Physical Chemistry would be classified under the theory facet of the subject Physical Chemistry, itself a category in the subdivision facet of Chemistry.

Other qualities of books that affected their classification were the time of publication, the geographical origin, and the language. Specific sequences were created for works relating to specific persons, and for materials related to particular works (such as books about the works of Shakespeare).

Despite the addition of facets, combination topics still presented a significant problem to classifiers. Although Horne, working in the British National Library, had proposed classification of books under multiple subjects in 1825 [24], by the beginning of the 20th century libraries were still stuck on shelving as the principle factor in book classification, thus every book needed a primary subject class. Martel observed that "countless subject combinations [were] constantly found in new variety," and that transfers of books from one category to another were quite common, as better choices for the primary subject became clear from use.

The scheme in use always favored primary classification by the orientation of the author, which while helpful in resolving many classification problems, propagated the secular nature of scholarship. This factor was also clearly heuristic, and could point evidence in entirely the wrong direction at times.

Biographies were handled specially and normally placed in a special area of the history facet of the subject that the person the book was about was associated with. This was not usually done uniformly and depended on the existing structure of the history facet. Some subjects had very limited history facets because they were new.

Literature, fiction and poetry in general, was classified by author, requiring new categories be created every day.

Letters, collections, and essays were considered the most difficult to classify, and each was done in a manner that seemed appropriate at the time.

Most of these principles are still in practise today. The biggest paradigm shift has been towards retrieval and access and away from locating books on shelves, and the COLON classification system is probably the most modern approach [1]. The COLON system emphasizes facets more directly, with a richer set of facet categories than the LCC, and was developed by Ranganathan to address the rapid growth of human knowledge, and the rather natural observation that multiple subject headings for a book was the rule, not the exception. Ranganathan did not believe any book should have a primary subject, but eventually relented in this requirement of his system, admitting that the books had to be placed somewhere [34].

The COLON system also added the notion of relatedness in subjects, and this idea along with facets has been adopted by the thesaurus community [7]. One term is related to another if they are normally associated with each other, and one is not more general than the other. For example, "airplanes" is related to "air mail service." Thesaurus structures do not specifically represent subjects; the goal is to provide relationships between terms used in any setting, but there is significant overlap in the problems of thesaurus representation.

Thesauri typically have four kinds of relationships between terms: synonym, broader, narrower, and related. E.g. "plane" is a synonym with "airplane," broader than "jetliner," narrower than "vehicle," and related to "air mail."

The addition of facets as a relationship between terms is fairly new, and has been driven by the same problems described here: specifically that it is frequently the case that one term may appear prima facie to be narrower than another, but further analysis reveals that’s not quite right. For example, it is not truly the case that "cloning" is a narrower term than "biology;" in some ways it might be consiered broader: with moral, political, religious, economic, and other aspects. With facets, however, we can keep cloning a subtopic of biology and add facets to biology (and cloning) for politics, ethics, etc.

In the course of this research, we have investigated the notion of subject hierarchies from many different angles. The main motivation has been the belief that digital libraries imply a profound shift in certain assumptions that have guided library subject classification since its inception:

• Digital documents do not need to be shelved, and can appear in many "places" at once. Modern library classifiers are trained to classify books under as many subjects in the catalog as appropriate. It can not be denied, however, that the book must have a primary classification so that it can be shelved, and that the basic classification scheme itself was designed for the purposes of shelving, not retrieval.

• In the card catalog ontology outlined in Section 2.3 above, the object of a search does not need to be a document. We want to be able to classify people, organizations, conferences, etc. by subject as well. Library classifiers, as mentioned earlier, actually do use information they know about individual authors to help determine the primary subject classification of a book, however this information (about the author) is not recorded. Furthermore, we want to be able to employ rules of inference about the classifications that are made explicitly in order to propagate the classifications to other types of objects.

It should be noted that the purpose of this work is to provide an extensive card catalog system, that supports browsing and retrieval as described in Section 2.3 and Section 2.4. Our goal is not to create an encyclopedia, in fact we hope to be able to stop a long way short of Cyc [27], while still providing a significant advance in usefulness over existing card catalog systems.

This work in subject classification has been ongoing since 1994. The following sections outline other approaches to modeling subjects we have investigated.

An essential ingredient in all systems with subject classification, from Bacon, through Diderot, LCC, Dewey, COLON, and even Yahoo!, is a taxonomy or hierarchy of subjects. This notion seems inexorably tied to subject classification. Our work actually began with the ontology for digital library card catalogs [38] implemented in a description logic [5], and because description logics are particularly good at representing taxonomies, adding a subject taxonomy seemed a trivial and useful additional feature. Five years later, the card catalog ontology has been stable for four years and subjects are an open area of research.

Our main reason to use a subject taxonomy is to narrow a search. This has proven particularly effective in web catalogs such as Yahoo!, but it brings up an important point: library subject catalogs, as mentioned in Section 3.3, were developed as a cataloging and shelving method, not for retrieval. Many of the problems we faced, and solutions we experimented with, took a different turn since shelving was not of concern to us.

An initial and recurring obstacle we encountered when trying to add a subject taxonomy to the card-catalog system was in trying to capture the power of description logics to do subsumption over the subject taxonomy. A more complete discussion of description logics can be found in [5], a formal treatment in [2], and a comparison to first-order logic in [4]. A brief overview of the nature of subsumption reasoning in description logics follows, but the reader is strongly encouraged to review the referenced literature.

A description logic concept describes a set in such a way that membership can either be stated explicitly or computed. For example, if we define the concept BIRD to be any animal that has feathers, and define an individual Harry to be an animal that has feathers, then a description logic will derive that Harry is a member of the set BIRD. Like many representation systems, description logics make a syntactic distinction between concepts and individuals. In addition, unlike many representation systems, description logics are capable of reasoning at the terminological level. For example, if we additionally create a concept called EAGLE to be any animal with feathers and a sharp beak, a description logic will derive that EAGLE is subsumed by BIRD. While taxonomic relationships are a fairly common element of most representation systems, description logics permit describing why one concept is more specific than another. Finally, most description logic implementations deal with taxonomies very efficiently by caching all subsumption relationships; as a result, retrieving the individuals of any concept (i.e. the members of any set) is a simple memory access and requires no computation.

Given a representation and reasoning tool like a description logic, creating a subject taxonomy seems prima facie simple. We encounter, however, three main problems when trying to integrate a subject taxonomy with our card-catalog ontology: