Creating an Infrastructure for Scholarly Publication On Line
Anne Mahoney


The Stoa Consortium aims to promote collaborative scholarship, published on line and freely available to other scholars and general readers. As a publisher, we must provide the mechanisms for making these scholarly works available. As an on-line publisher, we intend to provide more than what a good print publisher can provide: electronic texts can be explicitly connected as books cannot.

A digital library project generally produces and controls its own texts and images. This makes it relatively easy to interconnect those resources. A digital publisher, on the other hand, does not produce texts itself, and indeed the works published under its auspices need not all live on the same computer system. Interconnection in this environment requires more explict co-operation by the authors and editors of the text: specifically, use of the same markup rules.

Markup for us means SGML or XML, conforming to the TEI DTD and following Stoa conventions for which features to mark, which values to use for certain attributes, and which meta-data to include in the document header. We have found that most scholars wishing to publish texts with us do not know SGML in general, or this DTD in particular. Although they quickly see the benefits of structured markup, they must learn the language and the local idiom. The first piece of publishing infrastructure, then, is support for markup: documentation, editing software, and validators.

Many authors also want detailed control over the appearance of their work. They expect, sometimes unconsciously, that the technology of the Web as it exists today is and will be the final, best way to publish their work. While we prefer to take a longer-term, more general view, we recognize that on-line publication today does in fact mean the Web. The second piece of infrastructure is display formatting, whether for the Web, for printing, or for some future delivery medium.

The third piece of infrastructure is interconnection, which is what makes on-line publication fundamentally different from print. Interconnection means that references to other objects, outside the text, can automatically be hyperlinked to those objects. The typical digital library model is that objects within the library can be hyperlinked to each other, while objects outside are not linked. Since the digital publishing house does not "contain" its texts, on the other hand, there is no reason for it to restrict these reference connections to only works it has published itself. The Stoa has developed a reference database that allows any work published by the Stoa to refer to any resource elsewhere on the Internet. Hyperlinks are generated when the text is displayed, which means they are always as current as the content of the database.

The Stoa's tool set will be made available as open-source software, and we hope that other on-line publishers and digital libraries may wish to share reference information.


The Stoa Consortium was founded in 1998 to promote collaborative scholarship in the humanities, particularly the classics. The Stoa also intends to promote technical standards, for text encoding as well as for image capture, geographic data recording, and other processes. The best way to promote a standard is to provide an application that uses it. For the text encoding standards the Stoa wishes to use, that application is the Perseus text-processing system. The Perseus system works with texts encoded in SGML1 , using the DTD of the Text Encoding Initiative. Those texts can be presented on the Web, converted to PDF, searched, interlinked, and indexed, based on structural and semantic information in the markup.

The Perseus text system is an infrastructure, but it requires additional support, both technical and social. Technical support includes the ancillary applications that facilitate structured markup -- editors, syntax validators, documentation. It also includes applications relevant to the other parts of the system, for example support for image captioning and curatorial cataloguing, but those applications are outside the focus of the present paper.

Social support involves convincing the community that structured markup is desirable and feasible. Many potential Stoa collaborators are concerned that structured markup may be difficult to learn; they would rather use a familiar tool, for example a word processor like Microsoft Word or a typesetting language like TeX. Others note that works published on the Web as it exists today must be converted into HTML, and wonder why it's necessary to write in a different language, then convert the text; they'd rather write in HTML in the first place. Still other scholars expect, not unreasonably, that the technical details are the publisher's problem; just as they do not typeset their own books and articles for publication in print, they do not see why they are expected to use an apparently complex markup scheme for electronic publication. While all these objections are sensible enough, they do not take into account key reasons for using structured markup. I will discuss this further in section two.

The Perseus text system includes a variety of ways to change the display of a text, some under the reader's control and some controlled by the author or editor. Stoa collaborators who expect detailed control over the presentation of their work can have it, up to the limits imposed by HTML of course. In the Perseus Digital Library, each collection has its own logo and color scheme, but most of them have a common appearance, indicating that they are all part of a single digital library. This is a choice made by the Perseus designers, not an intrinsic feature of the software system, and it is in fact straightforward to change the appearance of a collection of texts or even of an individual document. I will go into more details in section three.

The most important feature of the Perseus system is the array of sophisticated features it provides for interconnecting texts. Interconnection -- hyperlinking -- is what makes on-line publication fundamentally different from print. Originally, I expected to implement tools for the Stoa, and last year produced a rudimentary linking system for certain kinds of references. As I thought about the tools we were producing for the Perseus Digital Library, however, I realized it was silly to re-implement something similar, when what I really wanted for the Stoa was the full power of the Perseus toolset. The various automatically generated hypertexts, implicit "information-push" searches, and dynamically generated reversible citations make this digital library toolset one of the richest environments currently available for on-line publication. The Stoa has therefore adopted the Perseus toolset. This has also prompted the Perseus team to generalize the toolset and make it more widely available. Perseus has another collaborating project that hopes to install and use this application, and we expect that in the next year or so we will be able to make the text processing system available to other projects as open source software.

In what follows, I will discuss the Stoa's technical and social support for structured markup; how Stoa projects might exploit the display mechanisms of the Perseus text system; and how structured markup facilitates interconnections among texts, not just within a single collection but in a web of co-operating systems.

Markup Support

Technical support for structured markup is easy: the basic tools are widely available. Any number of editors, ranging from highly specialized SGML systems to general-purpose, configurable utilities, and at prices ranging from free to tens of thousands of dollars, make it possible to edit structured text. None of these editors is more difficult to learn than a modern word processor.

It is also important to be able to verify the correctness of SGML markup, and for this one uses a validator. Some editors have built-in validation capabilities, and some can even ensure that the text being edited is never invalid. It's also possible to use a separate validation program. A compromise between integral and separate validation is an editor that can call upon the validation program when you ask it to. However one invokes it, the SGML validator identifies any syntactic problems with the markup of the text, from mis-spelled tags to missing structural levels.2

Intermediate between technical support and what I'm calling social support is documentation. Good software should always have good documentation, and good standards necessarily have good documentation. The documentation for the TEI DTD is excellent, if a bit daunting in its magnitude. The documentation for the DTD itself, however, is not enough: it's also necessary to explain the standards of the particular project, and the rules imposed by the software that will process the SGML documents. I have written an explanation of the Stoa's particular guidelines, in tutorial form; it's available on the Stoa's web server, where it has been accessed over 6,700 times in the past year. Technical information about the implementation of the text system is available in a series of research articles by members of the Perseus technical staff. Additional documentation is always desirable, however. In this case, particular desiderata include a detailed listing of the conventions for attribute values, the choices of features to be tagged, and other project-specific and application-specific guidelines beyond the use of the TEI, as well as detailed documentation of the major data structures and processing flow in the Perseus text system.

But social support extends beyond documentation. While it's straightforward to explain how to create SGML documents, it is also important to explain why. As I outlined above, there are several different reasons why Stoa collaborators do not immediately embrace structured markup: its presumed difficulty, the apparent silliness of writing in one language only to convert at once to another, and the principle that it should be the publisher's responsibility to do the publishing of a text. I will consider each of these in turn.

Structured markup is not in itself difficult, nor is it a form of computer programming. Whenever we begin a new scholarly article by sketching out a title, an abstract, and section headings, we are making a structured text. When we format the abstract in a smaller type face from the rest of the text, or make the section headings larger and bold, we are applying a markup that reflects the structure. All that SGML does is provide a formalism for expressing this kind of structure in a way that any reader or any software can exploit. When we read a text, if we see occasional short lines of larger, boldface letters, perhaps starting with a number, we recognize these lines as section headings. If we're skimming the text quickly, perhaps the section headings are the only things we read. We know these lines are section headings because we have a great deal of experience reading scholarly articles, and we know how they are ordinarily laid out. We can recognize the section headings even if the article is written in a language we can't read.

With SGML markup, the section headings are recognizable because they are labelled. Whereas in ordinary typographical "markup" the features are recognizable by their appearance, and interpretable in the light of human experience, in structured markup the features are recognizable and interpretable by their names: no human experience is required. In other words, a computer program can find the section headings, for example to produce a table of contents.

In the case of our section headings, it's pretty easy to recognize them unambiguously: nothing else in the text is likely to be a short line of larger, boldface letters. Consider a harder case, a word or phrase in italics. The human reader easily recognizes whether an italicized phrase is a title, a foreign phrase, a reported thought, or an emphasized phrase; these are conventional uses of italics, and context usually makes clear which is which. A program scanning a file, however, has no context, because it cannot actually read the text.3 If you want to list the works cited in the article, or make a glossary of the foreign phrases, you'll need to inspect all the italicized phrases and decide why each one is in italics. With SGML markup, instead of italicizing several different kinds of phrases for several different reasons, we mark each one with what it is: a title, a phrase in a particular language (which we specify), a reported thought, an emphasized phrase, or whatever else it might be. Display processing may indeed render all of these as italics, but the additional information is available to be exploited by indexing tools, automatic dictionary lookups, and so on.

Structured markup, then, allows authors to do the same kinds of things we do with our word processors, but also allows us to store more and clearer information.

But if we are simply going to turn those titles, foreign words, and so on back into italics, why did we bother marking them up? This is the second objection: why write SGML if we are only displaying HTML?

If all you are going to do with your SGML documents is convert them to HTML, then this objection is correct: you have no need of SGML. The point of a digital library system, however, is that you can do much more with structured documents. Every important feature of the Perseus text processing system is based on the fact that the documents are encoded in SGML, not HTML; marked for their structure, not their appearance. That they are displayed on the Web in HTML form is an ephemeral fact, not an intrinsic fact about the texts; the same SGML files can be converted to other formats for other purposes. As new display technologies are developed, it will be possible to use them to render SGML files, based on the semantic information in their markup, without changing the SGML files at all. HTML files, on the other hand, will not necessarily be automatically compatible with any other format at all. The information extraction and visualization tools of the Perseus text system also exploit the structural and semantic information of the SGML files.

The third objection to SGML raised by some Stoa collaborators is that markup should be the publisher's problem, just as it is in print. Print publishers accept text from authors, then convert it to whatever form their in-house system requires. Even if the publisher chooses to use SGML, the author might submit the text on paper, to be re-keyed by the publisher's staff.

This is, indeed, how many publishers do business now, but not all. In fact, some publishers of printed books and journals ask for submissions in SGML, according to a DTD they specify. This is a great advantage for them, because the texts come from their authors already in the correct form for further processing. It is also an advantage for the authors, because they can unambiguously specify the semantics of their text. Good SGML markup makes clear which sections are major and which are sub-sections or sub-sub-sections or sub-sub-sub-sections; this is difficult to do with typography, and error prone with outline numbering. SGML markup can also identify words in other languages, to facilitate spell-checking; can give regularized forms for words in non-standard dialects4 ; or can ensure that footnotes sit next to the text they comment on, rather than on some other page. An author who marks up a manuscript with SGML is indicating exactly what each element of the text means, not just what it might look like. In asking for submissions in SGML form, the Stoa is being progressive.


Regular users of the Perseus Digital Library are accustomed to the look of its texts. In most of the Perseus collections, text is displayed in a single window, with global tools in a top bar and collection-specific tools in a side bar. The background color of these tool bars identifies the collection. All Perseus text pages look much alike.

The designers of the Perseus Digital Library web site chose to retain a common style throughout, but the appearance of the pages is highly configurable. In fact, one of the collections, the Tufts University History collection, has a distinctly different appearance and somewhat different behavior.

The Perseus text processing system allows specification of a display template for a collection, to be used by all the texts in the collection; an individual text may also have an overriding template of its own. The templates for HTML display are written in HTML augmented with calls to processing functions that can fetch text or meta-data. It is also possible to override portions of the SGML conversion rules for particular elements, and this too may be done at the level of the collection or of the individual document.


The point of using a digital library system is to interconnect the objects in the library. In a print library, books are implicitly connected: when a book about Sophocles refers to Aeschylus, the author indicates the play and line number, and the reader is free to pull out a text of the older playwright and read the passage. In a well-implemented digital library, those connections will be explicit. When the on-line edition of the Sophocles book refers to Aeschylus, there will be a live link at the point of the citation, which the reader can follow to find the passage in Aeschylus. In the Perseus Digital Library, moreover, this link goes both ways. Someone reading Aeschylus will see a link back to the book about Sophocles which refers to the passage.

At present, these interconnections operate only within a single digital library. The Perseus group expects to make interconnections available among co-operating libraries, forming a federation. This will involve common meta-data rules and naming standards, so that for example Homer's Iliad is always identified in the same way. Each library in the federated group must make its catalog available to the others. Suppose a reader in one of the libraries requests a section of a book about Sophocles. The digital library system will display the desired section, and will link the references to Sophocles's plays to copies of those texts available at another co-operating library, for example at Perseus. The Perseus group expects to work on implementing such a system after the text processing system is in use at the Stoa.

Citations are not the only kind of interconnections among texts, or between texts and other objects. The Perseus text processing system can recognize place names, dates, and keywords in texts and link them automatically to appropriate tools, based on the contents of the digital library. These implicit searches will work in any digital library that uses the Perseus text processing system. As with citations, however, it is also quite desirable to make these searches work throughout a group of co-operating libraries; this is a future direction.


The Perseus digital library system is a powerful tool for on-line publishing of structured text. The Stoa will be installing this system for its own texts, and will ultimately join in a federated digital library group with the Perseus Digital Library. To make the Perseus system work for Stoa collaborators, those collaborators will need to embrace structured markup.

1 Or XML; the differences between SGML and XML are not significant for the present paper.

2 Most SGML validators cannot verify the semantics of the markup, however. For example, they cannot check whether text marked as prose is really verse, or whether a quote marked as Latin is really French. Moreover, they usually cannot check whether syntactically valid SGML conforms to the project's guidelines. For example, it is valid according to the TEI DTD to mark a phrase as a quotation, without indicating a language. When the quotation is in a different language from the surrounding text, the Stoa's standards require that the language be indicated. SGML validators cannot verify this because all they know is that the DTD allows quote elements with or without the lang attribute.

3 This is generally true of text processors. Programs that can parse natural language, or that have access to dictionaries for the main language of the text or of foreign languages that might appear in it, can often disambiguate italics as humans do.

4 e.g. Attic equivalents for the comic Spartan dialect in Lysistrata