Currently under development at the Institute for Advanced Technology in the Humanities at the University of Virginia, the Rossetti Hypermedia Research Archive is an electronic database of the complete works of the Victorian poet and painter Dante Gabriel Rossetti. It combines digital images (of manuscripts, printed pages, and artwork) and tagged ASCII texts, all fully cross-referenced via hypertext links within and among files. In development since the early 1990s, the Rossetti Archive has constituted the leading edge of electronic scholarly editions, having first confronted many of the issues and problems involved in such projects, and now approaching the completion of its first installment, to be issued in electronic form by the University of Michigan Press in 1999.

Jerome McGann, the general editor and creator of the Rossetti Archive, has written several excellent pieces on its background and purposes. His "The Rossetti Archive and Image-Based Electronic Editing", "Imagining What You Don't Know: The Theoretical Goals of the Rossetti Archive", the general "Introduction" to the Archive, and "The Rationale of Hypertext" are readily available on the World Wide Web. Further, it seems that the basic editorial implications and radical functionality of electronic archives no longer need to be asserted or explained, given the amount of ink (digital and otherwise) spilled on such topics in recent years.[1] Instead of covering that ground yet once more, I will share some of the lessons we learned in developing the Archive, as illustrated by some of the specific documentary instances that allowed us to learn those lessons.

Almost all electronic editions to date, including the Rossetti Archive, use some form of Standard Generalized Markup Language (SGML) to describe and present their materials. Projects that are currently being published on the World Wide Web use HyperText Markup Language (HTML), a subset of SGML that has become the standard for Web browsers. Many scholars and scholarly collectives have found in the Web a fast and inexpensive way to publish editions, and have thus adopted HTML as their tagging language. The problem with that approach is that HTML is a rendering language only: It tells the computer how to display the text on the screen, but it does not allow the editor to describe the logical, structural, or material features of the text. As a result, a great deal of documentary evidence must be edited out of those Web editions, or preserved in ASCII commentary alone. SGML, on the other hand, is a descriptive language, to which rendering protocols can be mapped when desired. From the start, the Rossetti Archive was conceived as an SGML-based project. Each document in the Archive was marked with a large set of descriptive tags that self-consciously preserve its unique structures and appearances.

Another subset of SGML, based on the language protocols developed by the Text Encoding Initiative (TEI), was rejected because it made scant provision for describing the physical features of documents, and because it assumed certain structural and presentational regularities that poets like Rossetti make it a rule to violate. We have developed a unique version (or "flavor") of SGML, designed to account in the most complete and faithful way for the varying features of Rossetti's works. The Archive's search engine will process these richly-described SGML files, and the results will be presented in HTML format via the World Wide Web. Accompanying JPG or GIF images of all pages and paintings will also be provided. In this article, I will focus on some of the features of the language the Rossetti Archive uses to describe its written and printed documents, leaving aside the pictorial works, which present their own set of issues. In particular, I want to discuss some of the particular struggles between SGML and our documents, and offer some insight into the evolution of our thinking about textual encoding, description, and rendering of the Rossetti Archive documents.

Can Logic Contain Art?

As the blacksmith and prophet Los begins his "terrible labors" in William Blake's Jerusalem, he announces, "I must Create a System, or be enslavd by another Mans."[2] So, too, the creators of the Rossetti Archive avoided preexisting versions of SGML in hopes of escaping servitude to a logical system that was not sensitive to the documentary situations at hand. For example, Rossetti was extremely attentive to the physical appearance of his printed books, so our version of SGML had to include tags for describing material features such as bindings and endpapers. Furthermore, Rossetti revised frequently and produced many intermediate (or alternate) versions and texts — many of which have survived — so our markup language had to be able to account for complex relations among manuscripts, sets of mixed proof pages, one-shot trial books, and "double works," which exist as both paintings and poems, among other things. Suffice it to say that the TEI protocols required extensive modification and expansion, and the result has been the evolution of a unique set of logically related descriptive tags.

In SGML, the master file that contains such a set of tags with their relations defined and made explicit is called a Document Type Description, or DTD. We have modified the Rossetti Archive DTD many times as new materials have challenged our previous assumptions about the kinds of logical and physical structures a document might contain. As a result of this often ad hoc process of evolution, our DTD may not be as lean as those for which all documentary features and relations can be accurately predicted in advance. Yet it seems that for large and complex collections of materials, particularly those not previously edited by 20th-century scholars (as Rossetti had not been), such complete predictions are impossible. So, for the Rossetti Archive, most often we have had to devise ways to escape not enslavement by another man's system, but the constraining implications of our own previous incarnations of the DTD.

That process of continual adaptation revealed one underlying limitation of SGML as a means of describing large sets of documents: SGML is by definition a mechanism of logical regularization. No matter how extensively you adapt the DTD to fit a group of texts, you always create a system of rules that defines what is possible. Each time you add a new document to the database, it has a potential to disrupt this system — a potential that increases with the great attention to detail required of scholarly editions. For small-scale archives, or archives in which the material is very regular and predictable, this situation may not be problematic, since the DTD may require no serious modifications, and will become stable quickly. However, for large, complex, and expanding collections of documents (that is, for most scholarly editions), the textual evidence will produce an interminable series of challenges to the assumptions of regularity that SGML requires.

An editor of an electronic archive can respond to such challenges in one of two ways. The first is Procrustean: He may use the existing tagging logic to describe the new documentary situation, overriding or altering the textual features to make them fit. At first, that may sound completely unacceptable, but in fact every creator of a scholarly edition in print or electronic form inevitably makes such compromises with the medium of transmission. Not everything can be preserved, because we have no way of imagining everything a text can be, much less a logical way of describing it. All editing, with the possible exception of strict facsimile reproduction, involves more or less intrusive processes of normalization and regularization. However, in the Rossetti Archive and most other electronic editions, the user may compare the tagged text to digital facsimile images of the page in question, in order to resolve small questions of rendering. For example, our DTD does not allow us to specify precisely where on the page a given bibliographic signature falls; and as we discovered, the locations of these do vary among documents. But since that information can be ascertained by calling up the facsimiles, we have deemed such granularity in our tagging unnecessary.

The second response to challenges to one's rules of description is simply to alter the DTD to account for the new situation: In that case, the tail wags the dog. One instance of an unforeseen textual feature causes a change in the basic structuring logic of the archive — a change that can have implications for many of the other documents in the archive. The result is continual updating of files as the DTD changes. For example, in some of the proofs for his 1870 Poems, Rossetti inserted into the sequence of the documents single revised pages that he had had printed for that purpose. In order to account for those pages as such, we introduced a new tag (revise) within the last year, and brought all of the affected files up to date. The change was important for those wishing to trace the development of Rossetti's 1870 volume, and was certainly worth making.

Even the simplest of alterations in thinking about the way a text should be described can have daunting implications. In the early stages of the project, we tagged manuscript corrections in the following manner: If Rossetti struck out a word, phrase, or line and wrote in a substitution, we marked the old material with del tags (indicating a deletion) and let the new reading stand. For example, in one of the proofs for the 1870 Poems, Rossetti deleted the word "course" in line 133 of "The Blessed Damozel," and substituted the word "paths." Under the old dispensation, that was tagged

Within the last year, however, we realized that that system was inadequate because it did not make clear which word was added by hand. Put another way, it occludes the distinction between the printed words of the proof page and Rossetti's autograph script. We decided that <add> tags needed to be placed around all such additions, so that the text in question is now tagged like this:

It was a simple change, but one that meant we had to reexamine and alter every document containing manuscript alterations by Rossetti (an inveterate reviser), comparing page images to the SGML markup and making the required changes. Episodes like that can consume many hours and seem to be some of the thousand natural shocks that electronic editions are heir to.

The Rossetti Archive has always been a kind of research laboratory, where revisions and refinements occur out of a directed process of trial and error. Some of the constant revisiting of files that characterizes it can be attributed to the necessarily collaborative environments in which they are executed. Over the course of the edition, tagging assistants (often graduate students) come and go, and as a result, many hands and minds have shaped the myriad files in the archive. Working alone frequently and making decisions constantly, each person executes the minute particulars of tagging slightly differently, even if he or she conforms to the same DTD. Because of that, local variation is commonplace and that can have disruptive consequences, particularly if the DTD changes. Developing a comprehensive style sheet will reduce such heterogeneity, but keeping it up-to-date and staying true to it could quickly paralyze the project. The potential immensity of electronic archives makes strict regularity among the files something to be longed for, but perhaps never seen.

In the world of SGML, the best way to limit unexpected compromises or alterations is to invest heavily up front in the research and design of a robust DTD appropriate to the needs of the project. As we complete the first installment of the Rossetti Archive, we have arrived at a richly textured and powerful DTD with which to describe and render texts. Yet we came to this by way of some paths that later editors and builders of electronic archives should not have to tread.

Several shortcuts can aid the large-scale adaptation of files to a newly revised system of tagging. Computers are very good at searching for patterns and replacing them with other patterns. Therefore, as long as the changes you need to make are uniform, little problem exists. A few keystrokes in most file-editing programs, such as EMACS or its JOVE subset, will perform such changes in a particular file in seconds.

For large groups of files, where such repeated opening and processing would be time-consuming, we have recently developed a series of scripts that will make global alterations, such as changing misconceived work codes to their proper forms, and placing quotation marks around all line numbers. Such scripts make certain that no instances are missed, and can enact the changes instantly. However, you have to be sure that global replacements will not garble or steamroll the text, and frequently the kinds of updating required cannot be performed automatically.

Finally, an archive might make use of entity references that call up a separate datafile that can be altered as needed. Let us say, for example, that you have attribution information from a number of libraries and museums that will need to be included in many files. If you begin by simply pasting data — the name of the museum, the director's name, a permission statement — into each file separately, you will have to perform similar cutting and pasting if you decide to include more or very different information later. However, if you set up an entity reference for each museum or library, you would need only to alter the file to which it referred, should any changes need to be made. For example, if you want to present the permissions data from the British Library, you could insert the string "& BritLib;" into the marked-up text; it tells the computer to call up the "BritLib" file and insert it into the text at that point. Thus changes made to the "BritLib" file will be replicated automatically in every file.

None of those solutions is particularly helpful when alterations are based on a reconceptualization of textual relations or rendering, which typically require thorough reexamination of all files and involve new decisions, new observations, and new tags. It may be that second- or third-generation projects of that kind will be able to avoid many of the revisions that the Rossetti Archive has experienced. But ultimately, it may be that we should consider inventing grammars of description more forgiving and flexible than SGML. Alternatively, we might acknowledge that SGML cannot easily preserve all the kinds of information that scholarly editions have valued throughout their bookish history. Such an admission might well free us to create specific electronic toolsets and databases whose radical functionality would exempt them from being "editions" as such. While the future of scholarly editions certainly will involve electronic frameworks and digitalization, such editions may come to depend on protocols other than those of SGML.

Rendering Conceptual Units: The Need for Double-Tagging

For the textual editor, another major problem with SGML involves the competing claims of formal structures and typographic rendering. As mentioned above, in order for a computer to display an SGML document, it must be told how to render the text marked by a given tag. For example, should it indent at every paragraph tag? Should it place a set of quotations marks at every quote tag? Should all titles of books tagged as such be rendered in italics? SGML is designed to operate most efficiently if the answer to all of these questions is yes. However, anyone who has worked on a scholarly edition will recognize the problem with such assumptions of regularity. Sets of documents rarely conform to such rigid schema, particularly when they include texts drawn from books, periodicals, and manuscripts from across many decades.

At the Rossetti Archive, we tabled that issue during the initial stages of the project, under two working assumptions: one, that the yet-to-be-created browser that would process our SGML files would somehow sort out questions of rendering, and two, that the facsimile images would provide ample witness to typographic specifics. Only recently, as we have approached the conclusion of the first installment, have we recognized the need to detach rendering from some of our conceptual tags, specifically <quote>, <title>, and <foreign>. For example, an italicized appearance of the title of Rossetti's famous sonnet sequence would be tagged as follows:

whereas the same title rendered in plain-text inside quotation marks would be tagged

The <xref> tags mark the text as a link to another file, in this case to the top-level document called a RAW (Rossetti Archive Work), which introduces and itself contains links to all documentary instances of that work. In the first instance, the <hi rend="i"> tag indicates that the enclosed text is rendered in italics, whereas in the second, the "& ldquor;" and "& rdquor;" are entity references that represent left and right double quotation marks. Note that in this case, the rendering information is nested outside of the title tags, because the title itself does not include any particular rendering.

In earlier versions of our files, the <quote>, <title>, and <foreign> tags carried with them assumptions of typographic alteration. Currently, we use them as strictly conceptual markers, and include other tags (indicating italics, for example) and/or characters (such as single or double quotation marks) to accomplish the idiosyncratic rendering of the text. Clearly, our desire to describe texts both physically and logically cannot be completely accommodated by a single set of SGML tags to which rendering decisions are mapped. In many cases, double-tagging is necessary to achieve the level of specificity that the scholarly edition requires.

The Question of White Space

SGML is good for describing strings of characters (indicating how they function in the document and how they appear on the page), but it is difficult to use it to provide information about white space. The reasons for that seem to be twofold. First, the editorial enterprise is bound by a bookish history that SGML was designed to transcend. SGML files are typically displayed on a monitor rather than a series of pages, so that the page itself becomes a kind of ghost in the machine, a remnant of the previous technology that the current one can't quite grasp. For example, the SGML display will be affected by various screen sizes and browser window sizes that make the formatting of the electronic archive a more fluid process than it was in the page-bound documents of the archive. Second, as students of literary texts, we have only begun to develop effective ways to talk about the meanings of gaps and white spaces in literary texts. That aporia in our critical vocabulary has left us unable to imagine how such spaces might be usefully and faithfully described.

Most browsers will not recognize more than two spaces — and some only one — between characters unless a special SGML tag is inserted. As a result, either every space beyond the first must be tagged, or the rendering can be approximated and reliance for greater specificity must again be placed on the facsimile images. In the Rossetti Archive, we have opted for the latter path in regard to indentation. Thus, a typical stanza, this one from a poem by Rossetti entitled "The Stream's Secret," is marked as follows:

Here, the lines are numbered consecutively, and the various numbers within the indent tags are used to indicate relative degrees of indentation, rather than specific quantities of white space. The browser will be told to indent the lines in proportion to the numbers in the tags, with "1" being the smallest indentation and "4" being the largest. The result is:

Furthermore, those relational indentations are not reset for each poem, so that a line with an indentation of "1" in "The Stream's Secret" will be rendered flush with a line indented "1" in another poem, although they may not be indented an equal number of absolute spaces. The resulting SGML text will contain an approximation of the document's spacing, which can be checked against the page image for greater specificity. This system is admittedly a compromise, arrived at as a way to preserve relational spacing without having to count every space in every document; and later versions of the Archive may well evolve greater levels of fidelity in this regard.

Searching Mechanisms

Every one of Rossetti's discrete works, whether written or pictorial, has its own unique work code, so that a user interested in finding all appearances of or references to that work can do so. In a sense, the deployment of those work codes within the SGML files establishes the central nervous system of the hypertextual archive. At the most basic level, a volume of Rossetti's poems exists as a discrete file with a work code, and each poem (and/or division) within that volume also has its own work code, which is provided in the header tags of that poem (and/or division). For example, the poem "Aspecta Medusa" (work code =1-1865) would be tagged as follows for its appearance in Rossetti's 1870 Poems (work code =1-1870):

Each new work within a document is a division, arranged hierarchically by means of enumerated div tags. By embedding the work code in the div tag as an attribute, we have allowed the computer to find this poem automatically. Users will be able to gather all available documentary instances of or references to "Aspecta Medusa" (or any work by Rossetti, despite name changes or radical revisions that might disable a string search) by asking the computer to search for its defining code.

However, as we moved beyond the primary documents and began tagging works of biography and criticism, that system revealed a simple but fundamental limitation: how do we identify quotations of Rossetti's work as such? Typically, those are not separate divisions of the document in question, and thus the div tag cannot be used. Particularly when they occur in the midst of paragraphs, without the mention of a title, such quotations were in danger of remaining invisible to various kinds of searches. Here again was an example of how, as we expanded the scope of the Archive, we found current practices unsuitable to a new class of documents.

Our solution in this case was an entirely new tag: workunit. Its purpose is to identify and describe quotations of Rossetti's work wherever they appear in the Archive. In practice, the workunit tag operates as follows, in a passage from William Michael Rossetti's Family Letters and a Memoir of Dante Gabriel Rossetti:

Here William Michael quotes two lines from his brother's translation of a poem within the Vita Nuova. In addition to the work code and type of quoted material, the attributes of workunit allow us to specify whether this is a block or an in-line quotation and whether it quotes the work in full or in part. Thus the workunit tag becomes an identifying marker of any quotation of a work by Rossetti, providing for its inclusion in searches for instances of that work, despite its incomplete, untitled, and unpredictable location within another document. Note, too, that the place of the quoted lines within the poem as a whole is preserved through the use of the "r" attribute in the line tag. While the "n" value represents the actual number of the line within the line group on the page, the "r" value refers to the line numbers (here, 69 and 70) of an arbitrarily selected version, or reference text, of the poem, to which all other instances are compared for purposes of collation. That feature will fully automate the often-difficult process of collating and comparing various documentary instances of a given line or set of lines. All versions of lines 69-70 in the Archive can be compared, no matter what their location or numerical value within a given document.

Conclusion

Such examples of the adaptive nature of the Archive's methodologies and practices could easily be multiplied. In fact, the Archive has moved forward through a continual series of confrontations between our methodologies and our praxis, between design and implementation — ultimately, between our assumptions of regularity and the intractable variability of documents. Because we have chosen to design our own system of SGML, theoretically independent of any specific hardware or software platform, we have been bound at the highest level only by the rules of set logic and the principles of bibliographical description.[3] However, our work has been continually shaped by the minute particulars of execution and application. Certainly building that kind of scholarly engine requires one to imagine the questions it will be able to process, and the ways that SGML can make those processes produce accurate information. But the real work of developing the Rossetti Archive and projects like it occurs in the repeated reconfiguring of those imaginings as we enact their implications.

Andrew Stauffer is an assistant professor of English at California State University, Los Angeles. He received his Ph.D. in 1997 from the University of Virginia, where he worked as a research assistant on the Rossetti Archive. His professional interests include Romantic and Victorian poetry, textual studies, and information technology and the humanities. You can find him on the web at http://www.calstatela.edu/faculty/astauff2/ [formerly http://www.calstatela.edu/faculty/astauff2/]. He can be reached via e-mail at astauff2@calstatela.edu.

Notes

1. In addition to the essays collected in the Finneran volume, see Lavagnino, 1996; Landow and Delany, eds., 1993; Shillingsburg, 1986; and Bender, 1976. For examples of other electronic archives under development, see the "Research Reports" [formerly http://jefferson.village.virginia.edu/reports.html], of the projects at the University of Virginia's Institute for Advanced Technology in the Humanities, http://jefferson.village.virginia.edu/. In addition, helpful recent discussions of electronic textuality can be found at the Web site for Bruce Graver's panel, "Electronic Texts and Textuality," at the 1997 NASSR conference, http://www.rc.umd.edu/features/conferences/archive/nassr96/graver-panel.html.

2. Erdman, ed., 1970, pp. 149, 151. The idiosyncratic punctuation here is Blake's own, rendering nicely the burden of Los's statement.

3. Of course, certain computer protocols and technical realities have determined some of the contours of the project, just as the availability of Rossetti documents (and publication permissions) have set certain limits on its current contents.

Bibliography (Printed Materials)

Bender, T.K. 1976. Literary Texts in Electronic Storage: The Editorial Potential. Computers and the Humanities 10:193-99. [doi: 10.1007/BF02426307]

Deegan, M., and Sutherland, K., eds. 1998. Electronic Text: Investigations in Method and Theory. Oxford: Oxford University Press, forthcoming.

Erdman, D.V., ed. 1970. The Poetry and Prose of William Blake. New York: Doubleday.

Finneran, R. J., ed. 1996. The Literary Text in the Digital Age. Ann Arbor: Univ. of Michigan Press.

Landow, G., and Delany, P., eds. 1993. The Digital Word: Text-Based Computing in the Humanities. Cambridge: MIT Press.

Lavagnino, J. 1995: Reading, Scholarship, and Hypertext Editions. Text 8:109-24.

Shillingsburg, P. 1986. Scholarly Editing in the Computer Age: Theory and Practice. Rev. ed. Athens: Univ. of Georgia Press.