Metadata Semantics Shared Across Languages

The Dublin Core™ is a set of fifteen basic categories (such as creator, title, subject, and publisher) for describing information resources (see http://dublincore.org/). When embedded in documents on the World-Wide Web, these Core descriptions can be extracted by global indexing services for use like a sort of library catalog. Some first adopters of the Dublin Core™ appreciate the generic simplicity of these fifteen basic categories and use them "as is" -- let's call them the "minimalists". Others -- let's call them the "structuralists" -- use qualifiers to narrow the semantics of Core categories for specialized uses -- for example, to specify that a particular "creator" is a "composer" as opposed to an "author" or "photographer". The success of the Dublin Core™ as a standard will depend on its ability to satisfy the structuralists' need for this kind of specificity without compromising the semantic integrity of the fifteen broad categories expected by the minimalists.

Until now, the Dublin Core™ has been defined and implemented only in English. Yet the meanings of these fifteen basic categories in their "minimalist" sense could just as well be explained in French, German, Japanese, or Thai. The Dublin Core™ could become that which has long eluded internationalist-minded librarians: a simple description model consistent across many languages and scientific disciplines.

But if the Dublin Core™ is to meet the need of specialists across many languages, it will also have to allow users in those languages to define their own qualifiers. And since the Dublin Core™ has so far been used only in English and on the World-Wide Web, where indexing tags must use only the English alphabet, its designers have not yet had to cope with multiple sets of tags and qualifiers in multiple languages and alphabets. (Note that the language of the Dublin Core™ itself is independent of the language of documents being described; records created with the Thai Dublin Core™ could refer to articles in Japanese.)

These issues were discussed most recently at the Fourth Dublin Core™ Workshop in Canberra, Australia, 3-5 March 1997 ( http://www.dstc.edu.au/DC4/). This paper reports some conclusions reached there and the consensus of a break-out group on how we might move towards a manageable multiplicity of Dublin Cores in languages other than English.

For the purposes of indexing and searching, equivalencies between parallel Dublin Cores in multiple languages could be assured by an agreed set of fifteen machine-readable indexing tags -- let's call them, more generically, tokens -- standing for the fifteen Core categories. These tokens, embedded with HTML in documents (or wherever else metadata may be placed), would flag the Core elements for automatic Web indexers. For better or for worse, the tokens currently used for this are English words and will most likely remain so for now, just as HTML tags for Web documents and the function words of most computer programming languages are English-like.

But the problem is more complex when we consider qualifiers. Qualifiers guarantee interoperability of more precise semantics to the extent that they are shared by communities of users. Similarly, they would guarantee interoperability across languages to the extent that they were shared between Dublin Cores in different languages. A few qualifiers, such as "author", will be useful all over the world and will likely have universal tokens. One could therefore search for an author regardless of whether the search form labelled it "author" or "putang" ("author" in Thai).

But as the Web increasingly becomes used for local and regional purposes, it seems likely that qualifiers will proliferate for which one will not need the global interoperability afforded by such widely shared concepts. As the Dublin Core™ is adopted more widely, Web-based registries will evolve to document these qualifiers, both local and universal, along with their tokens and definitions. One can picture a Thai registry for Dublin Core™ that listed fifty or so qualifiers shared with other Dublin Cores alongside qualifiers specific to the Thai language and Thai cataloging practice.

For a Dublin Core™ in Thai, one would ideally like to have a framework for defining two parallel sets of tokens for qualifiers: a set of local tokens expressed in the Thai language and alphabet, and a set of matching universal tokens for those qualifiers that were shared with other Dublin Cores. Local tokens with no universal equivalents, while useful locally, would simply be ignored by the global crawler services. Records created in Thailand, using local qualifiers, and indexed by a global crawler service that ignored those qualifiers, could still be retrieved in France via a "minimalist" search over the fifteen Core categories.

For several reasons, this ideal cannot readily be realised with today's Web technology. The current HTML format for metadata (the META tag) does not provide any standard way to distinguish between global and local tokens. And within HTML, machine-readable tokens -- whether universal or local -- must be expressed in 7-bit ASCII (the English alphabet from A to Z, plus numbers).

Just how serious these obstacles are for deploying Dublin Core™ in languages other than English depends in part on how metadata will be created in the future. If we were to assume that metadata will be typed in largely by hand -- and the most popular program for creating Web pages today is Microsoft's Notepad, a simple text editor -- then the limitation of ASCII will be a real problem. However, it is much more likely that users will type their descriptions into pop-up forms, perhaps with help menus and validation procedures. Software will take care of formatting the metadata properly and with the appropriate tokens. In such controlled environments, one might get around the limitations of ASCII by using transliteration: a user would type a qualifier in Thai letters and software would perform the necessary conversion into ASCII. Of course, the raw results might no longer be entirely comprehensible or editable by native speakers with plain text editors, so one could object to this workaround on grounds of principle.

Fortunately, it seems likely that these limitations of ASCII and HTML will be transcended by new technologies over the next year or two. The character-set limitation on tag names will fade as 7-bit ASCII is replaced by 16-bit Unicode, a code table that encompasses all the characters of the world's most common scripts. And the limitations on metadata tagging will be transcended by two new Web formats: PICS-NG and Web Collections. PICS was originally designed to support systems for rating Internet content (for example, to allow users to block access to pornography), though it is evolving into a general system in which labels and local tokens in many languages can be mapped onto generic metadata structures (see http://www.w3.org/pub/WWW/PICS/). Web Collections is a recent initiative for designing a general way to define sets of metadata (see http://www-ee.technion.ac.il/W3C/WebCollection.html).

Both PICS-NG and Web Collections will implement the basic elements needed for multiple Dublin Cores: global tokens, local tokens, and local descriptions. Moreover, both of these metadata frameworks are being developed by heterogenous communities of experts in resource description, annotations, digital signatures, digital cash, and resource description, and they have the support of the biggest software companies. Indeed, these initiatives are shaping the next big version of HTML itself. It is quite possible that both of these proposals will be sanctioned by the World-Wide Web Consortium (W3C) in June of 1997, after which it will hopefully take but a few months to a year for stable browsers, servers, and tools to come to market. To the participants in the Canberra workshop interested in creating Dublin Cores in languages other than English, it seemed wiser to anticipate these new solutions than to invest much energy in extending today's HTML.

While waiting for this deployment, communities can work on translating the Dublin Core™ into various languages. These translations will need to be discussed and reworked until they really make sense to native speakers. The resulting descriptions will need to reflect as precisely as possible the intent and semantic scope of the fifteen basic Core elements. Beyond that, local needs will determine the choice of qualifiers and their correspondence or non-correspondence to qualifiers in the English-language model.

As a simple first step, it seems desirable to make the descriptions of Dublin Cores in languages other than English available on Web pages, perhaps with lists of qualifiers, explanatory material, and examples of usage. These Web sites could be linked among themselves; indeed, the Thai description of, say, the Subject element might usefully be linked to the descriptions of that element in English and German. Help pages might describe the advantages of using universal qualifiers. Perhaps these interlinked Web servers could provide a platform for future, more advanced registry services, such as automated lookups of element values across languages, which could assist retrieval in ways we cannot yet clearly imagine. In such an infrastructure of peer servers, no one model would dominate (in a logical sense), as the English-language Dublin Core™ does now.

Researchers from several countries have indicated interest in establishing Dublin Cores, and projects are already underway in Berlin, at the Humboldt University in collaboration with the Max Planck Institute, and in Bangkok, at the Technical Information Access Center of the National Science and Technology Development Agency. A mailing list has been established to discuss Dublin Cores in multiple languages, and further workshops are foreseen. For further info, please contact Tom Baker at [email protected]

Originally posted 21 April 1997. Change in text, 27 August 1997: replaced "specialists" with "structuralists".