Using BIBFRAME for bibliographic description

Bibliographic description is an essential process of librarianship. In the distant past this process took the form of simple inventories. In the last century we saw bibliographic description evolve from the catalog card to the MARC record. With the advent of globally networked computers and the hypertext transfer protocol, we are seeing the emergence of a new form of description called BIBFRAME which is based on the principles of RDF (Resource Description Framework). This essay describes, illustrates, and demonstrates how BIBFRAME can be used to fulfill the promise and purpose of bibliographic description.†

Librarianship as collections & services

Philadelphia FlowersLibraries are about a number of things. Some of those things surround the collection and preservation of materials, most commonly books. Some of those things surround services, most commonly the lending of books.†† But it is asserted here that collections are not really about books nor any other physical medium because those things are merely the manifestation of the real things of libraries: data, information, and knowledge. It is left to another essay as to the degree libraries are about wisdom. Similarly, the primary services of libraries are not really about the lending of materials, but instead the services surround learning and intellectual growth. Librarians cannot say they have lent somebody a book and conclude they have done their job. No, more generally, libraries provide services enabling the reader to use & understand the content of acquired materials. In short, it is asserted that libraries are about the collection, organization, preservation, dissemination, and sometimes evaluation of data, information, knowledge, and sometimes wisdom.

With the advent of the Internet the above definition of librarianship is even more plausible since the materials of libraries can now be digitized, duplicated (almost) exactly, and distributed without diminishing access to the whole. There is no need to limit the collection to physical items, provide access to the materials through surrogates, nor lend the materials. Because these limitations have been (mostly) removed, it is necessary for libraries to think differently their collections and services. To the author’s mind, librarianship has not shifted fast enough nor far enough. As a long standing and venerable profession, and as an institution complete with its own set of governance, diversity, and shear size, change & evolution happen very slowly. The evolution of bibliographic description is a perfect example.

Bibliographic description: an informal history

Bibliographic description happens in between the collections and services of libraries, and the nature of bibliographic description has evolved with technology. Think of the oldest libraries. Think clay tablets and papyrus scrolls. Think of the size of library collections. If a library’s collection was larger than a few hundred items, then the library was considered large. Still, the collections were so small that an inventory was relatively easy for sets of people (librarians) to keep in mind.

Think medieval scriptoriums and the development of the codex. Consider the time, skill, and labor required to duplicate an item from the collection. Consequently, books were very expensive but now had a much longer shelf life. (All puns are intended.) This increased the size of collections, but remembering everything in a collection was becoming more and more difficult. This, coupled with the desire to share the inventory with the outside world, created the demand for written inventories. Initially, these inventories were merely accession lists — a list of things owned by a library and organized by the date they were acquired.

With the advent of the printing press, even more books were available but at a much lower cost. Thus, the size of library collections grew. As it grew it became necessary to organize materials not necessarily by their acquisition date nor physical characteristics but rather by various intellectual qualities — their subject matter and usefulness. This required the librarian to literally articulate and manifest things of quality, and thus the profession begins to formalize the process of analytics as well as supplement their inventory lists with this new (which is not really new) information.

Consider some of the things beginning in the 18th and 19th centuries: the idea of the “commons”, the idea of the informed public, the idea of the “free” library, and the size of library collections numbering 10’s of thousands of books. These things eventually paved the way in the 20th century to open stacks and the card catalog — the most recent incarnation of the inventory list written in its own library short-hand and complete with its ever-evolving controlled vocabulary and authority lists — becoming available to the general public. Computers eventually happen and so does the MARC record. Thus, the process of bibliographic description (cataloging) literally becomes codified. The result is library jargon solidified in an obscure data structure. Moreover, in an attempt to make the surrogates of library collections more meaningful, the information of bibliographic description bloats to fill much more than the traditional three to five catalog cards of the past. With the advent of the Internet comes less of a need for centralized authorities. Self-service and connivence become the norm. When was the last time you used a travel agent to book airfare or reserve a hotel room?

Librarianship is now suffering from a great amount of reader dissatisfaction. True, most people believe libraries are “good things”, but most people also find libraries difficult to use and not meeting their expectations. People search the Internet (Google) for items of interest, and then use library catalogs to search for known items. There is then a strong desire to actually get the item, if it is found. After all, “Everything in on the ‘Net”. Right? To this author’s mind, the solution is two-fold: 1) digitize everthing and put the result on the Web, and 2) employ a newer type of bibliographic description, namely RDF. The former is something for another time. The later is elaborated upon below.

Resource Description Framework

Resource Description Framework (RDF) is essentially relational database technology for the Internet. It is comprised of three parts: keys, relationships, and values. In the case of RDF and akin to relational databases, keys are unique identifiers and usually in the form of URIs (now called “IRIs” — Internationalized Resource Identifiers — but think “URL”). Relationships take the form of ontologies or vocabularies used to describe things. These ontologies are very loosely analogous to the fields in a relational database table, and there are ontologies for many different sets of things, including the things of a library. Finally, the values of RDF can also be URIs but are ultimately distilled down to textual and numeric information.

RDF is a conceptual model — a sort of cosmology for the universe of knowledge. RDF is made real through the use of “triples”, a simple “sentence” with three distinct parts: 1) a subject, 2) a predicate, and 3) an object. Each of these three parts correspond to the keys, relationships, and values outlined above. To extend the analogy of the sentence further, think of subjects and objects as if they were nouns, and think of predicates as if they were verbs. And here is a very important distinction between RDF and relational databases. In relational databases there is the idea of a “record” where an identifier is associated with a set of values. Think of a book that is denoted by a key, and the key points to a set of values for titles, authors, publishers, dates, notes, subjects, and added entries. In RDF there is no such thing as the record. Instead there are only sets of literally interlinked assertions — the triples.

Triples (sometimes called “statements”) are often illustrated as arced graphs where subjects and objects are nodes and predicates are lines connecting the nodes:

[ subject ] --- predicate ---> [ object ]

The “linking” in RDF statements happens when sets of triples share common URIs. By doing so, the subjects of statements end up having many characteristics, and the objects of URIs point to other subjects in other RDF statements. This linking process transforms independent sets of RDF statements into a literal web of interconnections, and this is where the Semantic Web gets its name. For example, below is a simple web of interconnecting triples:

              / --- a predicate ---------> [ an object ]
[ subject ] - | --- another predicate ---> [ another object ]
              \ --- a third predicate ---> [ a third object ]
                                          yet another predicate
                                                  \ /

                                         [ yet another object ]

An example is in order. Suppose there is a thing called Rome, and it will be represented with the following URI: We can now begin to describe Rome using triples:

subjects                 predicates         objects
-----------------------  -----------------  -------------------------  has name           "Rome"  has founding date  "1000 BC"  has description    "A long long time ago,..."  is a type of  is a sub-part of

The corresponding arced graph would look like this:

                               / --- has name ------------> [ "Rome" ]
                              |  --- has description -----> [ "A long time ago..." ]
[ ] - |  --- has founding date ---> [ "1000 BC" ]
                              |  --- is a sub-part of  ---> [ ]
                               \ --- is a type of --------> [ ]

In turn, the URI might have a number of relationships asserted against it also:

subjects                  predicates         objects
------------------------  -----------------  -------------------------  has name           "Italy"  has founding date  "1923 AD"  is a type of  is a sub-part of

Now suppose there were things called Paris, London, and New York. They can be represented in RDF as well:

subjects                    predicates          objects
--------------------------  -----------------   -------------------------    has name            "Paris"    has founding date   "100 BC"    has description     "You see, there's this tower..."    is a type of    is a sub-part of   has name            "London"   has description     "They drink warm beer here."   has founding date   "100 BC"   is a type of   is a sub-part of  has founding date   "1640 AD"  has name            "New York"  has description     "It is a place that never sleeps."  is a type of  is a sub-part of

Furthermore, each of “countries” can be have relationships denoted against them:

subjects                         predicates         objects
-------------------------------  -----------------  -------------------------  has name           "United States"  has founding date  "1776 AD"  is a type of  is a sub-part of       has name           "England"       has founding date  "1066 AD"       is a type of       is a sub-part of        has name           "France"        has founding date  "900 AD"        is a type of        is a sub-part of

The resulting arced graph of all these triples might look like this:


From this graph, new information can be inferred as long as one is able to trace connections from one node to another node through one or more arcs. For example, using the arced graph above, questions such as the following can be asked and answered:

  • What things are denoted as types of cities, and what are their names?

  • What is the oldest city?

  • What cities were founded after the year 1 AD?

  • What countries are sub-parts of Europe?

  • How would you describe Rome?

In summary, RDF is data model — a method for organizing discrete facts into a coherent information system, and to this author, this sounds a whole lot like a generalized form of bibliographic description and a purpose of library catalogs. The model is built on the idea of triples whose parts are URIs or literals. Through the liberal reuse of URIs in and between sets of triples, questions surrounding the information can be answered and new information can be inferred. RDF is the what of the Semantic Web. Everything else (ontologies & vocabularies, URIs, RDF “serializations” like RDF/XML, triple stores, SPARQL, etc.) are the how’s. None of them will make any sense unless the reader understands that RDF is about establishing relationships between data for the purposes of sharing information and increasing the “sphere of knowledge”.

Linked data

Linked data is RDF manifested. It is a process of codifying triples and systematically making them available on the Web. It first involves selecting, creating (“minting”), and maintaining sets of URIs denoting the things to be described. When it comes to libraries, there are many places where authoritative URIs can be gotten including: OCLC’s Worldcat, the Library of Congress’s linked data services, Wikipedia, institutional repositories, or even licensed indexes/databases.

Second, manifesting RDF as linked data involves selecting, creating, and maintaining one or more ontologies used to posit relationships. Like URIs, there are many existing bibliographic ontologies for the many different types of cultural heritage institutions: libraries, archives, and museums. Example ontologies include but are by no means limited to: BIBFRAME,, the work of the (aged) LOCAH project, EAC-CPF, and CIDOC CRM.

The third step to implementing RDF as linked data is to actually create and maintain sets of triples. This is usually done through the use of a “triple store” which is akin to a relational database. But remember, there is no such thing as a record when it comes to RDF! There are a number of not a huge number of toolkits and applications implementing triple stores. 4store is (or was) a popular open source triple store implementation. Virtuoso is another popular implementation that comes in both open sources as well as commercial versions.

The forth step in the linked data process is the publishing (making freely available on the Web) of RDF. This is done in a combination of two ways. The first is to write a report against the triple store resulting in a set of “serializations” saved at the other end of a URL. Serializations are textual manifestations of RDF triples. In the “old days”, the serialization of one or more triples was manifested as XML, and might have looked something like this to describe the Declaration of Independence and using the Dublin Core and FOAF (Friend of a friend) ontologies:

<?xml version="1.0"?>
<rdf:RDF xmlns:rdf=""
xmlns:dcterms="" xmlns:foaf="">
<rdf:Description rdf:about="">
	<foaf:Person rdf:about="">

Many people think the XML serialization is too verbose and thus difficult to read. Consequently other serializations have been invented. Here is the same small set of triples serialized as N-Triples:

@prefix foaf: <>.
@prefix rdf: <>.
@prefix dcterms: <>.
<> dcterms:creator <>.
<> a foaf:Person;
  foaf:gender "male".

Here is yet another example, but this time serialized as JSON, a data structure first implemented as a part of the Javascript language:

"": {
  "": [
	  "type": "uri", 
	  "value": ""
 "": {
   "": [
	   "type": "literal", 
	   "value": "male"
   "": [
	   "type": "uri", 
	   "value": ""

RDF has even been serialized in HTML files by embedding triples into attributes. This is called RDFa, and a snippet of RDFa might look like this:

<div xmlns=""
<div typeof="rdfs:Resource" about="">
  <div rel="dcterms:creator">
    <div typeof="foaf:Person" about="">
      <div property="foaf:gender" content="male"></div>

Once the RDF is serialized and put on the Web, it is intended to be harvested by Internet spiders and robots. They cache the data locally, read it, and update their local triples stores. This data is then intended to be analyzed, indexed, and used to find or discover new relationships or knowledge.

The second way of publishing linked data is through a “SPARQL endpoint”. SPARQL is a query language very similar to the query language of relational databases (SQL). SPARQL endpoints are usually Web-accesible interfaces allowing the reader to search the underlying triple store. The result is usually a stream of XML. Admitted, SPARQL is obtuse at the very least.

Just like the published RDF, the output of SPARQL queries can be serialized in many different forms. And just like relational databases, triple stores and SPARQL queries are not intended to be used directly by the reader. Instead, something more friendly (but ultimately less powerful and less flexible) is always intended.

So what does this have to do with libraries and specifically bibliographic description? The answer is not that complicated. The what of librarianship has not really changed over the millenium. Librarianship is still about processes of collection, organization, preservation, dissemination, and sometimes evaluation. On the other hand, with the evolution of technology and cultural expectations, the how’s of librarianship have changed dramatically. Considering the current environment, it is time to evolve, yet again. The next evolution is the employment of RDF and linked data as the means of bibliographic description. By doing so the data, information, and knowledge contained in libraries will be more accessible and more useful to the wider community. As time has gone on, the data and metadata of libraries has become less and less librarian-centric. By taking the leap to RDF and linked data, this will only become more true, and this is a good thing for both libraries and the people they serve.


Enter BIBFRAME, an ontology designed for libraries and their collections. It is not the only ontology intended to describe libraries and their collections. There are other examples as well, notably,, FRBR for RDF, MODS and MADS for RDF, and to some extent, Dublin Core. Debates rage on mailing lists regarding the inherent advantages & disadvantages of each of these ontologies. For the most part, the debates seem to be between BIBFRAME,, and FRBR for RDF. BIBFRAME is sponsored by the Library of Congress and supported by a company called Zepheira. At its very core are the ideas of a work and its instance. In other words, BIBFRAME boils the things of libraries down to two entities. is a subset of, an ontology endorsed by the major Internet search engines (Google, Bing, and Yahoo). And since is designed to enable the description of just about anything, the implementation of is seen as a means of reaching the widest possible audience. On the other hand, is not always seen as being as complete as BIBFRAME. The third contender is FRBR for RDF. Personally, the author has not seen very many examples of its use, but it purports to better serve the needs/desires of the reader through the concepts of WEMI (Work, Expression, Manifestation, and Item).

That said, it is in this author’s opinion, that the difference between the various ontologies is akin to debating the differences between vanilla and chocolate ice cream. It is a matter of opinion, and the flavors are not what is important, but rather it is the ice cream itself. Few people outside libraries really care which ontology is used. Besides, each ontology includes predicates for the things everybody expects: titles, authors, publishers, dates, notes, subjects/keywords, added entries, and locations. Moreover, in this time of transition, it is not feasible to come up with the perfect solution. Instead, this evolution is an iterative process. Give something a go. Try it for a limited period of time. Evaluate. And repeat. We also live in a world of digital data and information. This data and information is, by its very nature, mutable. There is no reason why one ontology over another needs to be debated ad nauseum. Databases (triple stores) support the function of find/replace with ease. If one ontology does not seem to be meeting the desired needs, then (simply) change to another one.††† In short, BIBFRAME may not be the “best” ontology, but right now, it is good enough.


Now that the fundamentals have been outlined and elaborated upon, a workflow can be articulated. At the risk of mixing too many metaphors, here is a “recipe” for doing bibliographic description using BIBFRAME (or just about any other bibliographic ontology):

  1. Answer the questions, “What is bibliographic description, and how does it help facilitate the goals of librarianship?”
  2. Understand the concepts of RDF and linked data.
  3. Embrace & understand the strengths & weaknesses of BIBFRAME as a model for bibliographic description.
  4. Design or identify and then install a system for creating, storing, and editing your bibliographic data. This will be some sort of database application whether it be based on SQL, non-SQL, XML, or a triple store. It might even be your existing integrated library system.
  5. Using the database system, create, store, import/edit your bibliographic descriptions. For example, you might simply use your existing integrated library for these purposes, or you might transform your MARC data into BIBFRAME and pour the result into a triple store, like this:
    1. Dump MARC records
    2. Transform MARC into BIBFRAME
    3. Pour the result into a triple-store
    4. Sort the triples according to the frequency of literal values
    5. Find/replace the most frequently found literals with URIs††††
    6. Go to Step #D until tired
    7. Use the triple-store to create & maintain ongoing bibliographic description
    8. Go to Step #D
  6. Expose your bibliographic description as linked data by writing a report against the database system. This might be as simple as configuring your triple store, or as complicated as converting MARC/AACR2 from your integrated library system to BIBFRAME.
  7. Facilitate the discovery process, ideally through the use of linked data publishing and SPARQL, or directly against the integrated library system.
  8. Go to Step #5 on a daily basis.
  9. Go to Step #1 on an annual basis.

If the profession continues to use its existing integrated library systems for maintaining bibliographic data (Step #4), then the hard problem to solve is transforming and exposing the bibliographic data as linked data in the form of the given ontology. If the profession designs a storage and maintenance system rooted in the given ontology to begin with, then the problem is accurately converting existing data into the ontology and then designing mechanisms for creating/editing the data. The later option may be “better”, but the former option seems less painful and requires less retooling. This author advocates the “better” solution.

After a while, such a system may enable a library to meet the expressed needs/desires of its constituents, but it may present the library with a different set of problems. On one hand, the use of RDF as the root of a discovery system almost literally facilitates a “Web of knowledge”. But on the other hand, to what degree can it be used to do (more mundane) tasks such as circulation and acquisitions? One of the original purposes of bibliographic description was to create a catalog — an inventory list. Acquisitions adds to the list, and circulation modifies the list. To what degree can the triple store be used to facilitate these functions? If the answer is “none”, then there will need to be some sort of outside application interfacing with the triple store. If the answer is “a lot”, then the triple store will need to include an ontology to facilitate acquisitions and circulation.

Prototypical implementation

In the spirit of putting the money where the mouth is, the author has created the most prototypical and toy implementations possible. It is merely a triple store filled with a tiny set of automatically transformed MARC records and made publicly accessible via SPARQL. The triple store was built using a set of Perl modules called Redland. The system supports initialization of a triple store, the adding of items to the store via files saved on a local file system, rudimentary command-line search, a way to dump the contents of the triple store in the form of RDF/XML, and a SPARQL endpoint. [1] Thus, Step #4 from the recipe above has been satisfied.

To facilitate Step #5 a MARC to BIBFRAME transformation tool was employed [2]. The transformed MARC data was very small, and the resulting serialized RDF was valid. [3, 4] The RDF was imported into the triple store and resulted in the storage of 5,382 triples. Remember, there is no such thing as a record in the world of RDF! Using the SPARQL endpoint, it is now possible to query the triple store. [5] For example, the entire store can be dumped with this (dangerous) query:

# dump of everything
SELECT ?s ?p ?o 
WHERE { ?s ?p ?o }

To see what types of things are described one can list only the objects (classes) of the store:

# only the objects
WHERE { ?s a ?o }

To get a list of all the store’s properties (types of relationships), this query is in order:

# only the predicates
WHERE { ?s ?p ?o }

BIBFRAME denotes the existence of “Works”, and to get a list of all the works in the store, the following query can be executed:

# a list of all BIBFRAME Works
WHERE { ?s a <> }

This query will enumerate and tabulate all of the topics in the triple store. Thus providing the reader with an overview of the breadth and depth of the collection in terms of subjects. The output is ordered by frequency:

# a breadth and depth of subject analsysis
SELECT ( COUNT( ?l ) AS ?c ) ?l
  ?s a <> . 
  ?s <> ?l

All of the information about a specific topic in this particular triple store can be listed in this manner:

# about a specific topic
SELECT ?p ?o 
WHERE { <> ?p ?o }

The following query will create the simplest of title catalogs:

# simple title catalog
SELECT ?t ?w ?c ?l ?a
  ?w a <>           .
  ?w <>    ?wt .
  ?wt <>  ?t  .
  ?w <>      ?ci .
  ?ci <>       ?c  .
  ?w <>      ?s  .
  ?s <>        ?l  .
  ?s <> ?a

The following query is akin to a phrase search. It looks for all the triples (not records) containing a specific key word (catholic):

# phrase search
SELECT ?s ?p ?o
  ?s ?p ?o
  FILTER REGEX ( ?o, 'catholic', 'i' )

Automatically transformed MARC data into BIBFRAME RDF will contain a preponderance of literal values when URIs are really desired. The following query will find all of the literals and sort them by the number of their individual occurrences:

# find all literals
SELECT ?p ?o ( COUNT ( ?o ) as ?c )
WHERE { ?s ?p ?o FILTER ( isLiteral ( ?o ) ) }

It behooves the cataloger to identify URIs for these literal values and replace the literals (or supplement) the triples accordingly (Step #5E in the recipe, above). This can be accomplished both programmatically as well as manually by first creating a list of appropriate URIs and then executing a set of INSERT or UPDATE commands against the triple store.

“Blank nodes” (URIs that point to nothing) are just about as bad as literal values. The following query will list all of the blank nodes in a triple store:

# find all blank nodes
SELECT ?s ?p ?o WHERE { ?s ?p ?o FILTER ( isBlank( ?s ) ) }

And the data associated with a particular blank node can be queried in this way:

# learn about a specific blank node
SELECT distinct ?p WHERE { _:r1456957120r7483r1 ?p ?o } ORDER BY ?p

In the case of blank nodes, the cataloger will then want to “mint” new URIs and perform an additional set of INSERT or UPDATE operations against the underlying triple store. This is a continuation of Step #5E.

These SPARQL queries applied against this prototypical implementation have tried to illustrate how RDF can fulfill the needs and requirements of bibliographic description. One can now begin to see how an RDF triple store employing a bibliographic ontology can be used to fulfill some of the fundamental goals of a library catalog.


This essay defined librarianship as a set of interlocking collections and services. Bibliographic description was outlined in an historical context, with the point being that the process of bibliographic description has evolved with technology and cultural expectations. The principles of RDF and linked data were then described, and the inherent advantages & disadvantages of leading bibliographic RDF ontologies were touched upon. The essay then asserted the need for faster evolution regarding bibliographic description and advocated the use of RDF and BIBFRAME for this purpose. Finally, the essay tried to demonstrate how RDF and BIBFRAME can be used to satisfy the functionality of the library catalog. It did this through the use of a triple store and a SPARQL endpoint. In the end, it is hoped the reader understands that there is no be-all end-all solution for bibliographic description, but the use of RDF technology is the wave of the future, and BIBFRAME is good enough when it comes to the ontology. Moving to the use of RDF for bibliographic description will be painful for the profession, but not moving to RDF will be detrimental.


† This presentation ought to be also be available as a one-page handout in the form of a PDF document.

†† Moreover, collections and services go hand-in-hand because collections without services are useless, and services without collections are empty. As a buddhist monk once said, “Collections without services is the sound of one hand clapping.” Librarianship requires a healthy balance of both.

††† That said, no matter what a person does, things always get lost in translation. This is true of human language just as much as it is true for the language (data/information) of computers. Yes, data & information will get lost when moving from one data model to another, but still I contend the fundamental and most useful elements will remain.

†††† This process (Step #5E) was coined by Roy Tennant and his colleagues at OCLC as “entification”.


[1] toy implementation –
[3] sample MARC data –
[4] sample RDF data –
[5] SPARQL endpoint –

Published by

Eric Lease Morgan

Artist- and Librarian-At-Large

2 thoughts on “Using BIBFRAME for bibliographic description”

  1. I would argue that there is, in fact, such a thing as a record in RDF. A named graph (as specified in SPARQL 1.1) is a bag of triples that can be managed (queries, updated, deleted) as a unit, and hence corresponds quite closely to the idea of a “record”, though of course the content of a named graph is not constrained, in the way traditional records are, to some preset structure.

    The close correspondence between records and named graphs (as units of data management) makes named graphs (especially via the SPARQL 1.1 Graph Store protocol) a very convenient and useful technology for bridging legacy metadata stores to the Linked Data world.

  2. Thanks for posting this. A great primer. One question. I wonder if you meant “convenience” rather than “connivence” in this sentence, “Self-service and connivence become the norm.”

Comments are closed.