Text Markup-Data Structure vs. Data Model

Text Markup—Data Structure vs. Data Model

Allen
Renear

GSLIS/University of
Illinois
renear@uiuc.edu

2003

University of Georgia

Athens, Georgia

ACH/ALLC 2003

editor

Eric
Rochester

William
A.
Kretzschmar, Jr.

encoder

Sara
A.
Schmidt

SUMMARY
For over ten years there have been several carefully argued critiques of the
descriptive markup approach to text representation that have for the most
part not been adequately answered. Recently an integrated development of
some of these criticisms has been presented at length, and with considerable
ingenuity and extension, in a widely read article by Dino Buzzetti,
published in New Literary History (2002, 33:
61–88). In our paper we analyze some of Buzzetti’s criticisms, accepting
some but resisting others.

BACKGROUND
When the SGML or “descriptive markup” approach to text began to be actively
promoted in the mid-1980s, quite a number of criticisms were developed in
response to this approach, or, more exactly, to the accompanying theoretical
claims, express or implied. In retrospect, it appears today that some of
these criticisms have been answered by the proponents of SGML descriptive
markup, others have simply faded in apparent significance, and still others,
while live issues to some scholars, seem neverless to generate only
unproductive repetitive discussions.
However there is a particular family of related criticisms that upon careful
re-consideration still seem fresh, deep, and, at least given their positive
reception, plausible—and yet these criticisms have also been, strangely,
largely ignored by the markup community. Some of these arguments were first
presented in the late 1980s in a paper titled “Markup Considered Harmful”
privately circulated by Darryll Raymond in response to Coombs et al 1987.
They were later presented again, in more detail in 1992 in “Markup
Reconsidered” (Raymond, Tompa, and Wood 1995); and then were further refined
in “From Data Representation to Data Model: Meta-Semantic Issues in the
Evolution of SGML” (Raymond, Tompa, Wood 1996).
In brief these papers argue that “Markup is not a data model” but rather a
“data representation” (Raymond et al 1992), and that while “SGML provides
standard representations for documents” a “standard semantics” is required
as well (Raymond et al 1996). And then, generalizing from this conclusion,
the authors go on to suggest that many of the expectations for SGML-based
markup systems, such as TEI, are misplaced.
Despite the fact that these were pointed and clearly argued positions, the
response from the TEI community, and perhaps more importantly from the
theorizing researchers whose ambitious claims were particularly targeted
(Coombs et al 1987, DeRose et al 1990, Sperberg-McQueen 1990), appears to
have been largely to simply ignore these criticisms.
Recently Dino Buzzetti has been re-presenting these arguments, supplementing
them with further theoretical extensions, and integrating them with other
criticisms of descriptive markup approach, in particular those of Jerome
McGann, most recently presented and extended in a new book Radiant Textuality: Literature Since the World Wide Web (New
York 2001)—which Buzzetti takes as supporting Raymond’s and his own
criticisms. Buzzetti has in fact presented his concerns in several venues
over the last 5 years (e.g. Buzzetti 1999a, Buzzetti 1999b), but it is in
“Digital Representation and the Text Model” (New Literary
History, 2002 33) that he presents these criticisms in a form,
and a forum, and a language, in which they can no longer be ignored.
Interestingly, this article comes precisely at a moment when some of the
theorizing descriptive markup proponents are themselves very actively
working in a similar vein. Sperberg-McQueen, who has called for the
development of a SGML/XML semantics since at least 1992 (Sperberg-McQueen
1992), is leading a project to develop such a semantics now, and Renear, who
is member of this project, has also recently claimed, in language similar to
that of Buzzetti and Raymond et al, that an SGML document instance per se
provides “only a data structure, not a theory” (Renear 2001).

THE CRITICISM
Raymond et al argue, as described above, that markup is not a “data model”,
but only a data representation that is “fully dependent on external
information for meaning”. They note several characteristics that they say
confirm that markup is not a data model in the sense that that expression is
used in database theory. These include inadequate notions of equivalence,
lack of redundancy control, and lack of defined algebraic operators. This
claim should not of course be taken simply as a narrow assertion that SGML
markup does not meet criteria for being a certain sort of thing as defined
in some field or other—the larger point, and in the context of the
discussion a plausible one, is that SGML markup will be inadequate for the
kinds of theoretical tasks it is being assigned in virtue of failing to have
these characteristics. Raymond et al (1995) state that this failure can be
remedied either by adding a “formal structure on top of markup systems” or
by defining a new abstraction from scratch. Raymond et al prefer the latter
because they believe that markup systems are not only a representation
rather than a data model, but they are in fact a relatively poor technique
for representation. Raymond et al 1996 characterizes this addition of a
formal structure as adding “semantics” to SGML.
Buzzetti accepts Raymond et al’s criticism of SGML markup. But he goes much
further, both in the extent of his theorizing and in the specific nature of
his criticisms. With respect to the latter Buzzetti argues that descriptive
markup theorists are actually confused, and profoundly so: they conflate the
“structure of the representation” with the “structure of the object
represented”, or, alternatively, they confuse the “expression” with the
“content” of that expression. Supplementing Raymond’s criticism from the
perspective of database theory with an opposition from Saussure Buzzetti
argues that descriptive markup is an expression whose form is a data
structure. But the form of the content of that expression is a “data model”.
Neither the expression itself nor its form however, is a data model—as
mistakenly believed by the promoters of SGML descriptive markup and the
defenders of the OHCO model of text. (Buzzetti goes on to make this point
with juxtapositions of other theoretical terms: “format” vs. “formalism”,
“syntax” vs. “interpretation”). As examples of this confusion Buzzetti
singles out in particular the TEI Guidelines, Coombs et al 1987, and DeRose
et al 1990, and Renear et al 1996.
“In summary,” Buzzetti says, “we may assert that strongly embedded markup
systems [are] inadequate in reference to both the exhaustivity and the
functionality of the text representation and model.”

OUR RESPONSE
We believe that that there is some truth to both Raymond, Tompa, and Wood’s
criticisms, and to Buzzetti’s application of the criticisms. However we
believe that both Raymond et al and Buzzetti substantially overstate the
case, Raymond somewhat and Buzzetti much more. Such a response may seem too
mild to deserve public airing, but it is for two reasons much more important
than it sounds: first because of the radical form the criticism takes in
Buzzetti, and second because our response might clarify some long confused
issues in text encoding.
There can be little doubt that many SGML markup enthusiasts have been now and
then confused, and, even more frequently, confusing, about the
representational nature of SGML document instances. This is mostly because
representation itself is conceptually difficult, and subject to a kind of
semiotic oscillation between expressing, referring, and exhibiting. And it
may also be partly because of carelessness and genuine lapses of clarity and
understanding.
But Buzzetti makes too much of the fact that we are sometimes awkward about
what document instances are, representationally, and exactly how markup does
what it does. Notice that for the most part everyone manages the ambiguities
quite well, and gets on, successfully more often than not, with various
tasks and projects. How could this be the case if there were such a
consistent systematic widespread conflation of expression and content? And,
in fact, a close examination shows that Buzzetti does not produce evidence
for such a systematic conflation, other than the various awkward or
unfortunate characterizations mentioned above—and those, while admittedly
signs of some failure to fully conceptualize key notions, are far from
convincing evidence for systematic conflation.
Moreover, there is no evidence that this confusion, such as it is, is
connected, as either cause or effect, with a radical unsuitability of SGML
markup to express theories about text. (According to Buzzetti “SGML is a
data structure representation language” and does not provide a data model;
and he echoes the views of Raymond et al that it is, moreover, a bad data
structure representation language.) Part of the problem is that Raymond and
Buzzetti rely too much in their argument on the specific notion of a “data
model” from database theory, and therefore over-react to the failure of SGML
to have precisely that sort of data model, or to have a good (i.e.
completely or exactly defined) one. Part of the problem is that they demand
formalization before admitting that there has been sucessful representation.
It cannot be denied that we use SGML document instances to express express
facts and theories about texts rather effectively: that is simply an
empirical fact. Part of the problem is that in claiming that SGML and XML
lack semantics and a defined set of operators, Raymond et al and Buzzetti
ignore the fact that both SGML and XML define the semantics of the markup
vocabulary in use as an intrinsic part of the application's document type
definition. Unlike Raymond et al and Buzzetti, the SGML standard does not
assume that the full semantics of an SGML application are captured in full
by the SGML formalisms used to declare elements and attributes. Nor does a
demand for an exhaustive set of algebraic or other operations seem a
promising start for a critique of a markup discipline intended to encourage
the reuse of data and the separation of data representation from processing
concerns.
In this connection Buzzetti’s criticism of the ordered-hierarchy of content
objects (OHCO) hypothesis is particularly illuminating, because strikingly
unconvincing. Unlike SGML markup, OHCO is obviously not intended as a
representation syntax, but rather an abstraction claimed to be,
structurally, the general form taken by all texts. As such it actually
appears to be roughly the kind of thing that Buzzetti wants for a data
model, though without the specific features characteristic of data models in
database theory. Buzzetti’s interpretation of OHCO however has it as “not a
model of the text, but a possible model of its expression”, just as SGML on
Buzzetti’s account, describes not the text, but “the structure of the text’s
expression”, consistent with its distinctive role as “a data structure
representation language”. But whatever the failures of OCHO as a data model,
this is no more than a tortured effort to make things seem worse then they
are. The data structures in question have untyped purely logical
parent/child relationships, whereas “hierarchy” in OHCO clearly refers to
containment—and that is not a purely data
structural relationship, despite the obvious formal similarities (asymmetry,
transitivity, etc.) that makes tree data structures convenient for
representing hierachical containment.
It is true that additional formalization of the semantics of SGML markup is a
good thing. Formalization will assist in precision, clarity, and
computation. It is for that reason that Sperberg-McQueen and others are
working to explicate the semantics of XML markup. But while we may regret
that such work has not made more progress, or that not everyone understands
the need for improved formalization of semantics and “data models”, or that
some of our colleagues are sometimes confused about subtle matters of
representation, or that we ourselves are now and then confused, or even
often confused, we should not conclude that things are worse than they are.
Scholars working in SGML descriptive markup are working with languages that
do have semantics, and do have “data models”, even if inadequately
formalized, or not quite the sort of thing database researchers prefer. And
these scholars are at least as often as not quite clear on the difference
between expression and content.

CONCLUSION
The criticisms of inadequate formalization that began with Raymond et al and
that have recently been so intriguingly developed and extended by Buzzetti
are important ones that have been mysteriously neglected by the markup
community. And the more ambitious criticisms made by Buzzetti of our failure
to fully theorize, or at least attempt to clarify, the complexities of
representation are perhaps even deeper and more important, and also
unfortunately neglected. But the claim that in these failures we see the
signs of either a systematic category mistake endemic in the text encoding
community, or evidence of a disastrous inadequacy in current techniques,
simply has not yet been proven. On the contrary, as we have suggested here,
that claim itself seems to be based upon the mistaken notion, ironically
reminiscent of the third man argument, that one more formal system will
finally bridge the gap between thought and object. There may or may not be a
gap, but if there is it will not be closed simply by another formal system,
however useful that system may be.

REFERENCES

Dino
Buzzetti

Text Representation and Textual Models

ACH-ALLC'99 Conference Proceedings

Charlottesville

1999

Dino
Buzzetti

Digital Representation and the Text Model

New Literary History

33

2002

James
H.
Coombs

Allen
H.
Renear

Steven
J.
DeRose

Markup Systems and the Future of Scholarly Text
Processing

Communications of the ACM

30
11

November 1987

Steven
J.
DeRose

David
G.
Durand

Elli
Mylonas

Allen
H.
Renear

What Is Text, Really?

Journal of Computing in Higher Education

1
2

1990

Jerome
McGann

Radiant Textuality: Literature Since the World Wide
Web

New York

2001

Darrell
R.
Raymond

Frank
W.
Tompa

Derrick
Wood

Markup Reconsidered

First International Workshop on Principles of Document
Processing, Washington, D.C., October, 1992

1992

Darrell
Raymond

Frank
Tompa

Derrick
Wood

From Data Representation to Data Model: Meta-Semantic
Issues in the Evolution of SGML

Computer Standards and Interfaces

18
1

January 1996

Allen
Renear

David
Dubin

C.
M.
Sperberg-McQueen

Claus
Huitfeldt

Towards a Semantics For XML Markup

Proceedings of the ACM Symposium on Document
Engineering ACM, November 2002

2002

Allen
Renear

Elli
Mylonas

David
G.
Durand

Refining Our Notion of What Text Really Is: The Problem
of Overlapping Hierarchies

Research in Humanites Computing

Oxford

1996

Allen
Renear

Raising the Bar, Text Encoding from a Logical Point of
View

CLiP 2001

2001

C.
M.
Sperberg-McQueen

Back to the Frontiers and Edges

Closing remarks at “SGML '92: the quiet revolution”
conference sponsored by the Graphic Communications Association
(GCA), Boston, 29 October 1992

1992

Available on the Web at

C.
M.
Sperberg-McQueen

Claus
Huitfeldt

Allen
Renear

Meaning and Interpretation in Markup

Markup Languages: Theory and Practice

MIT Press
2000

C.
M.
Sperberg-McQueen

Allen
Renear

David
Dubin

Claus
Huitfeldt

Drawing Inferences on the Basis of Markup

Proceedings of Extreme Markup 2002

2002