Text und Textgestaltung
C. M. Sperberg-McQueen
Draft of a Festvortrag to be given in Potsdam
in October 2018, at the ITUG meeting.
This draft includes some rough text in English and some in
German. Neither is finished. Any paragraph containing just a
“-” or “{}” in the German is intended to be omitted
from the oral presentation; it might be supplied in a possible
written version.
In the current state of the draft, some topics come up more
than once, and some points may be made more than once. This
may be a sign of carefully constructed recurrent theme, which
gets set up in one place and alluded to again and again later.
It is more likely, however, that the author has merely lost
the thread of the discussion and that the paragraphs need to
be resequenced (and some paragraphs merged or cut entirely).
If you are reading these words before the talk is actually
given, it means I've asked you to read the draft, and that you
are cordially invited to suggest improvements to the flow.
I have not gathered all the images I want for the slides;
in some places where I know I'll want a slide to illustrate a
point, I've inserted a dummy slide, typically of a piece of
marble.
Right now, I believe that section 2 (the shape of text on
the page) is far too long; if sections 3 and 4 were worked out
at equal length, the talk would run for two or three hours.
So I expect to try to cut section 2 quite a bit, though
there are also some important missing topics I'd like to
add. A rough time budget would be
- 1 Introduction (3 min)
- 2 The shape of text on the page (15 min)
- 3 The shape of text in the machine (15 min)
- 4 The shape of text on our heads (10 min)
- 5 Conclusion (2 min)
- 1 Introduction (3 min)
- 2 The shape of text on the page (20 min)
- 3 The shape of text in the machine (12 min)
- 4 The shape of text on our heads (8 min)
- 5 Conclusion (2 min)
- 1
- 2
- 3
- 4
- 5
1
Introduction ·
Einleitung
It is an honor to be invited to address you
this evening on the occasion of the 25th anniversary of the
International Tustep User Group, in the year that
Prof. Wilhelm Ott turns 80. My congratulations and best
wishes all round.
Es ist eine Ehre, eingeladen zu werden, Ihnen an diesem
Abend anläßlich des 25. Jahrestages der Internationalen
Tustep User-Gruppe vorzutragen, in dem Jahr, in dem
Prof. Wilhelm Ott seinen 80. Geburtstag feiert. Ich
gratuliere.
I should confess at the outset that I am not a proficient
user of TUSTEP. Improving my TUSTEP skills has perpetually
been on my to-do list, and I regret that so far in my career
other obligations have prevented me from spending as much
time on that task as I would wish. But I have known about
TUSTEP almost since the time I began working seriously in
what we now call the digital humanities, and I have been
deeply impressed with what skilled users can accomplish with
this tool. And the knowledge I have of Tustep, incomplete
thought it is, has influenced me over the years whenever I
have had the occasion to work on infrastructure that could
benefit digital humanists and other scholars who work with
electronic text.
Ich sollte gleich am Anfang bekennen, dass ich kein
TUSTEP-Experte bin. Meine TUSTEP-Kenntnisse zu verbessern
steht seit Jahren auf der privaten Aufgabenliste, und ich
bedaure, dass bis jetzt andere Obliegenheiten es unmöglich
gemachte haben, dazu zu kommen. Von Tustep habe ich schon
früh in meiner Arbeit in der Digital Humanities gelernt, und
das, was Tustep-Könner mit diesem Werkzeug erreichen können,
hat mich immer tief imponiert. Auch die partielle Kenntnis
von Tustep hat meine Arbeit an Infrastruktur für die Digital
Humanities und andere wissenschaftliche Gebiete
beeinflusst.
Part of what Tustep seems to me to illustrate is what can
be achieved if in writing software we make it our first
task to meet the requirements of serious textual
scholarship. Software for scholarly text processing has
an interesting and intimate relation to our understanding
of texts as linguistic, historical, and cultural objects:
to support textual work effectively, software must embody
or reflect a serious understanding of text, and conversely
thinking about how our software works and how it models
textual structures can help us think about texts and
textuality more effectively and can help lead us to a
deeper understanding of text.
Unter anderem zeigt Tustep (wie mir scheint), was man
erreichen kann, wenn man bei der Herstellung von Software
zuallererst versuch, den Anforderungen des
wissenschaftlichen Umgangs mit Text gerecht zu
werden. Software für wissenschaftliche Textverarbeitung
hat eine interessante und intime Beziehung zu unserem
Verständnis von Texten als linguistische, historische und
kulturelle Objekte: Um Textarbeit effektiv zu
unterstützen, muss Software ein wissenschaftliches
Verständnis von Text zugrunde liegen. Umgekehrt kann das
Nachdenken darüber, wie unsere Software funktioniert und
wie es textuelle Strukturen modelliert, uns zu einem
tieferen Verständnis von Text verhelfen.
That is my topic for this lecture: Text, the shape of
text, and the shaping of Text.
Daraus ergibt sich mein heutiges Thema: der
Text, die Textgestalt, und die Textgestaltung.
In this day of multimedia, of high-resolution full color
scanning, of automatic image recognition, why worry
about text?
Warum heute, in dieser Zeit des Multimedialen, des
Vollfarb-Scannens in höchster Auflösung, der automatischen
Bilderkennung, warum heute uns noch mit Text beschäftigen?
The computer scientist Tim Bray, with whom I had the
pleasure of collaborating a few years ago, once had a
start-up company with the motto
“Knowlege is a text-based application.”
Der Informatiker Tim Bray, mit dem ich vor einigen
Jahren eng zusammenarbeitete, hatte einmal eine
Start-up-Firma mit dem Motto
“Das Wissen ist eine textbasierte Anwendung.”
When Alan Turing set himself the question “at what point
will it be fair to say that computers can think?” his
answer (known now as the Turing Test) involved the
exchange of written messages — that is, texts
— between the computer and a human being. Turing
has been criticized for equating thinking with mere verbal
facility — what about visual thinking and other
non-verbal forms of intelligence? — but it may be
noted that he did not in fact suppose that all thought is
verbal. He suggested only that if a machine exhibits
sufficient verbal facility, then it will be fair to call
what the machine is doing a kind of thinking. (After all,
our libraries are full of works by people we call the
thinkers of the past, and the main evidence we have of
their thinking are the texts they left behind.)
Als sich Alan Turing die Frage stellte "Wann wird es fair
sein zu sagen, dass Computer denken können?", war seine
Antwort das, was uns jetzt als der Turing - Test bekannt
ist. Dieser Test beruht auf den Austausch von
geschriebenen Nachrichten - also Texten - zwischen dem
Computer und einem Menschen. Turing wurde dafür
kritisiert, das Denken mit bloßer verbaler Fertigkeit
gleichzusetzen - es gibt ja da visuelle Denken und andere
nichtnonverbalen Formen der Intelligenz! - aber Turing hat
doch nicht angenommen, dass alle Gedanken verbal sind. Er
schlug nur vor, dass es fair sei, wenn eine Maschine eine
ausreichende verbale Fähigkeit aufweist, das von der
Maschine Geleistete "Denken" zu denkt. (Schließlich sind
unsere Bibliotheken voll Werke von Menschen, die wir die
Denker der Vergangenheit nennen, und der Hauptbeweis, dass
sie gedacht haben, sind die Texte, die sie hinterlassen
haben.)
I do not say that thinking is the same
thing as the production or consumption of texts — thinking
can also take other forms and use other tools. But it does
seem clear that producing texts and consuming them are two
activities very prominent in the way most of us go about
thinking. It may be possible to think without creating a
text, but it is difficult to imagine producing a text
without something that we would have to call thought.
Ich sage nicht, dass das Denken dasselbe ist wie die
Produktion oder der Konsum von Texten - das Denken kann
auch andere Formen annehmen und andere Werkzeuge
verwenden. Aber es scheint klar zu sein, dass das
Hervorbringen von Texten und das Konsumieren von Texten
zwei Tätigkeiten sind, die für die meisten von uns sehr
wichtig sind. Es mag möglich sein, zu denken, ohne einen
Text zu erstellen, aber es ist schwer vorstellbar, einen
Text zu produzieren, ohne das wir etwas vollbringen, das
wir als Denken bezeichnen müssten.
(Even texts which we hyperbolically call mindless or
thoughtless — whether they are advertising material,
bureaucratic prose, or the writings of our political
adversaries — involve mental processes which as far as we
know exceed the capacity of any non-human animal and of
any existing computational device.)
(Auch Texte, die wir hyperbolisch "hirnlos" oder
"gedankenlos" nennen - ob es sich um Werbematerial,
bürokratische Prosa oder die Schriften unserer politischen
Gegner handelt - erfordern mentale Prozesse, die, soweit wir
wissen, die Fähigkeiten jedes nichtmenschlichen Tieres und
jedes existierenden Computergerät übersteigen.)
Texts play a centrol role in the practice
of all scientific and scholarly disciplines (not only
literature and linguistics) and in the study of their
history.
Texte spielen eine zentrale Rolle in der Praxis aller
wissenschaftlichen Disziplinen (nicht nur der Literatur
und Linguistik) und in der Erforschung ihrer Geschichte.
So today I would like to talk a little bit about text,
about the nature of text and how we shape text. It is, I
hope, a good way to honor both TUSTEP, one of the most
sophisticated and complex tools we have for working with
text, and the man to whom we owe its existence.
Deshalb möchte ich heute ein wenig über Text sprechen:
über die Natur von Text und darüber, wie wir Text
gestalten. Ich hoffe, es ist eine geeignete Art, TUSTEP,
eines der raffiniertesten und komplexesten Werkzeuge, die
wir für die Arbeit mit Texten haben, und den Mann, dem wir
seine Existenz verdanken, zu würdigen.
[Need more / better motivation: what problems does a
model of text solve? what is the harm if a model is bad?
]
2
The shape of text on the page ·
Die Textgestalt auf der Seite
If we want to know what text is, what shape it takes,
there are worse ways to start than looking at the shapes
text takes when transmitted from sender to receiver.
Wenn wir wissen wollen, was Text ist, welche Gestalt er
hat, können wir damit einen Anfang machen, indem wir
schauen, in welcher Gestalt wir Texte vom Sender zum
Empfänger senden.
2.1
The linearity of language ·
Die Linearität der Sprache
If we consider oral texts, we are likely to infer
quickly that whatever else may be involved, text is a
one-dimensional sequence of some fundamental units or
other. We may take the fundamental units to be
sentences, or words, or sounds; this affects the details
of the analysis, but one feature remains invariant:
For every item in the text, whether
sentence, word, or phoneme, there is exactly one
next item. (There is one exception: if the
text is finite, as thus far all human texts have been,
then the last item in the text has no next item.) And
conversely, for every item (except the first) there is
exactly one immediately preceding item.
Wenn wir mündliche Texte in Betracht ziehen, werden wir
wahrscheinlich schnell darauf schließen, dass Text, was
auch immer er sonst sein mag, eine eindimensionale
Abfolge fundamentaler Einheiten sei. Wir können als
fundamentale Einheiten Sätze oder Wörter oder Töne
ansetzen. Die Details der Analyse werden danach
varieeren. Eine Eigenschaft aber bleibt
invariant: Für jedes Element im Text, ob Satz, Wort oder
Phonem, gibt es genau ein nächstes
Element. (Eine Ausnahme: wenn der Text endlich ist, wie
bisher alle menschlichen Texte, dann hat das letzte
Element im Text kein nächstes Element.) Und umgekehrt gibt
es für jedes Element (außer dem ersten) genau ein sofort
vorhergehendes Element.
This idea feels natural, almost unavoidable: all natural
human languages appear to involve a one-dimensional
sequence of sounds, if only because human beings have
only one mouth, and the sound signal of human speech
unfolds in real time.
Sometimes the one-dimensionality of an object of regret,
or the object of a critical attack: George Landow's
account of hypertext contrasts it with conventional
text, which is one-dimensional and which must be read in
a single order, from start to finish, without deviation
from the path. A similar hostility to one-dimensional
text can be found in the work of Ted Nelson. Nelson
often appears to blame paper for this
one-dimensionality,1 and Landow
appears to blame the codex, though in fact the kind of
one-dimensionality they talk about is a function of
human language rather than a result of its being
recorded on paper.2
Diese
Idee scheint natürlich, fast unvermeidbar zu sein: Alle
Natursprachen scheinen aus eindimensionalen Tonfolgen zu
bestehen, schon allein deshalb, weil der Mensch nur
einen Mund hat und das Tonsignal der
menschlichen Sprache sich in Echtzeit entfaltet. Manche
bedauern die Eindimensionalität der Sprache, oder machen
sie zum Objekt eines kritischen Angriffs: in seiner
Arbeit zum Hypertextbegriff kontrastiert George Landows
den Hypertext mit konventionellem Text, der
eindimensional ist und in einer einzigen Reihenfolge
gelesen werden muss, von Anfang bis Ende, ohne
Abweichung vom Weg. Eine ähnliche Feindseligkeit
gegenüber die Eindimensionalität findet sich in dem Werk
Ted Nelsons. Nelson scheint dem Papier, Landow der
Kodexform des Buchs den Schuld für die
Eindimensionalität zu geben, obwohl die
Eindimensionalität, von der sie sprechen, eher eine
Funktion der menschlichen Sprache ist als eine Folge
ihrer Aufzeichnung auf Papier.
On the contrary, paper offers more opportunities than a
single human voice for saying more than one thing
‘at the same time’, and the
distinguishing feature of the codex as a book medium is
that it provides effectively random access to the
content, in contrast to the purely sequential access
afforded by the scroll. The shift from magnetic tape to
magnetic disk for mass storage of information echoes the
earlier shift from scroll to codex, and with many of the
same consequences.
Ganz im Gegenteil: das Papier bietet bessere
Möglichkeiten, als die menschliche Stimme, mehrere
Aussagen ‘gleichzeitig’ zu
machen, und der Kodex unterscheidet sich als Textträger
grundsätzlich dadurch von der Rolle, dass er einen direkten
Zugriff zu beliebigen Stellen des Inhalts bietet, während
die Rolle prinzipiell nur sequentiellen Zugriff bietet.
Der Übergang von Magnetbänder zu Platten bei der
Speicherung von Grossdaten ähnelt in Vielem den Übergang
von der Rolle zum Kodex, und viele der Folgen sind auch
überraschend ähnlich.
2.2
The non-linearity of the page ·
Die Nichtlinearität der Seite
In this context, it may be informative to
look at a written text which reflects this idea fairly
directly, as in this 1494 printing of a German translation
of Apollonius of Tyre.
In diesem Zusammenhang mag es informativ sein, einen
schriftlichen Text zu betrachten, der diese Idee
ziemlich direkt widerspiegelt, wie in diesem 1494 Druck
einer deutschen Übersetzung des Romans von Apollonius
von Tyrus.
Slide: Apollonius of Tyre 1494
The type on the page is perfectly
legible, there is nothing the least bit unconventional
about the layout, and yet the page does not look at all
like a modern page, even setting aside the use of Fraktur.
If text is fundamentally a one-dimensional sequence of
characters, why does this page, which treats text as
precisely that not feel more familiar? It may not be
immediately obvious just why the page looks foreign to us.
The source of our feeling may become more evident if we
examine a conventional printed page.
Die Schrift auf der Seite ist gut lesbar, am Layout ist
nichts unkonventionell, und dennoch sieht die Seite gar
nicht wie eine moderne Seite aus (auch wenn man den
Einsatz von Fraktur ausser Acht lässt). Wenn Text im
Grunde eine eindimensionale Folge von Zeichen ist, warum
fühlt sich diese Seite, die den Text genau in diesem
Sinne widergibt, nicht vertrauter an? Es ist vielleicht
nicht sofort klar, aus welchen Gründen die Seite uns fremd
erscheint. Die Quelle unseres Gefühls kann deutlicher
werden, wenn wir eine konventionelle gedruckte Seite
betrachten.
Slide: standard modern page
Slides needed: closeups of
page numbers, running heads, paragraph
breaks, and section heads.
Among the things a modern eye misses in the incunabulum
(which we can notice because they are present on this
page) are
page numbers
running heads
the use of whitespace to mark paragraph breaks
Das moderne Auge vermisst in diesem
Wiegendruck unter anderem
die Seitenzahlen
die lebenden Kolumnen
die Gestaltung von Absatzgrenzen durch Zeilenbruch und Einrückung (o.ä.)
2.2.1
Paragraph and section breaks ·
Absatz und Textunterteilung
Slide: paragraphus symbol, modern para break
Modern paragraph breaks make use of
whitespace in ways that incunabula often do not: the last
line of the one paragraph is not filled out, and the new
paragraph begins on the next line, after an indentation.
Because of the shift to the next line, the modern
paragraph break shown here is a two-dimensional phenomenon
-- the same goes even more strongly for the section head
shown. We can of course shoehorn these into a linear
stream of characters, by treating functions like
“carriage return”, “line feed”, “vertical
tab”, and “horizontal tab” as special kinds of
characters which are embedded in the text stream like any
other. This is the way Teletype machines handled issues
of text formatting. And indeed, all standard coded
character sets for computer use reserve space for
‘control characters’ like these.
In modernen Büchern werden Absätze mit Leerzeichen
gestaltet, wie es bei Inkunabeln oft nicht der Fall
ist: Die letzte Zeile des einen Absatzes wird nicht
ausgefüllt, und der neue Absatz beginnt in der
nächsten Zeile nach einer Einrückung. Wegen der
Verschiebung in die nächste Zeile ist der hier
gezeigte moderne Absatzbruch ein zweidimensionales
Phänomen - das gleiche gilt für den abgebildeten
Abschnittskopf noch stärker. Wir können diese
natürlich in einen linearen Zeichenstrom einfügen,
indem wir Funktionen wie "Rücklauf", "Zeilenvorschub",
"vertikaler Sprung" und "horizontaler Sprung" als
spezielle Arten von Zeichen behandeln, die in den
Textstream eingebettet sind Irgendwelche anderen. Auf
diese Weise haben Teletype-Maschinen die
Textformatierung verwaltet. Und in der Tat reservieren
alle standardisierten codierten Zeichensätze für
Computergebrauch Platz für "Steuerzeichen" wie diese.
The absence in inunabula of anything
resembling a modern paragraph break or section break
serves as a useful reminder that our modern typographic
expectations are just that — a recent development,
not more than a few centuries old. To a large extent, the layout conventions we
find natural today require thinking in two dimensions, not
only in one.
Dass Wiegendrucke keine modernen Absatzbrüche oder
Abschnittbrüche haben, dient als nützliche Erinnerung
daran, dass unsere modernen typografischen Erwartungen
genau das sind - eine jung Entwicklung, die nicht mehr
als einige Jahrhunderte alt ist. Die
Layout-Konventionen, die wir heute als
selbstverständlich empfinden, erfordern zu einem
großen Teil das Denken in zwei Dimensionen, nicht nur
in einem.
We not only render paragraph breaks differently in
modern books, we also have, I believe, more of them,
just as we have more subdivisions of the text at larger
sizes. Literary texts now almost always are divided
into chapters and paragraphs; long poems have been
divided into cantos and books since Dante and Tasso and
Vergil. Homer is also divided into books, but the
divisions are a post-Homeric (Alexandrian) innovation,
as are the divisions of Beowulf into fittes, and the
division of the Nibelungenlied into
aventiuren. How does it happen that
we now subdivide our texts at fine granularity into
chapters and sections and sub-sections? Is it that the
regular, unbroken block of text has become so unwelcome
that we introduce structural breaks in our texts in
order to have an excuse for introducing whitespace on
the page?
Wir gestalten nicht nur in modernen Büchern
Absatzbrüche anders, wir haben auch, glaube ich, mehr
Absatzbrüche, genauso wie wir mehr Unterteilungen des
Textes in größeren Einheiten haben. Literarische Texte
sind jetzt fast immer in Kapitel und Absätze
unterteilt; lange Gedichte sind seit Dante und Tasso
und Vergil in Cantos und Bücher aufgeteilt. Homer ist
auch in Bücher unterteilt, aber die Spaltungen sind
eine posthomerische (alexandrinische) Neuerung, ebenso
wie die Aufteilung von Beowulf in Fittes und die
Aufteilung des Nibelungenliedes in Aventiuren. Wie
kommt es, dass wir unsere Texte nun feingranular in
Kapitel und Abschnitte und Unterabschnitte
unterteilen? Ist der regelmäßige, ungebrochene
Textblock so unwillkommen geworden, dass wir
strukturelle Brüche in unseren Texten einführen, um
eine Entschuldigung für die Einführung von Leerzeichen
auf der Seite zu haben?
2.2.2
Page numbers ·
Die Seitenzahlen
Page numbers illustrate a pervasively important
property of printed books: if you and I have copies of
the same edition, I can refer you to a particular
passage of the work by mentioning its page
number.3 This dramatically simplifies
the discussion (and thus the study) of texts. Since
manuscript copies of a work do not, as a rule,
preserve the pagination of the exemplar, such page
references are not a prominent feature of scribal
cultures, and in order to support references to
specific textual locations in a scribal culture it is
necessary4 to construct
some system for labeling specific passages in a text.
The Eusebian canons used in Western
Europe to allow references to Biblical passages and
the creation of Gospel harmonies and rudimentary
concordances are an example of such a system. Only
the most important texts will be worth the effort of
creating such a system. In a print culture where all
pages are numbered, any text will have a
convenient reference system.5
Slide: running heads
In books printed nowadays, running heads often identify
both the book and the chapter of the book. This is rare
(if probably not unheard of) in manuscript books, and it
is tempting to speculate that it reflects the relative
scarcity or abundance of books. In a monastic community
in which one is allowed a single book in one's cell, for
reading during silent periods of the day, it may
relatively seldom be the case that one wishes to be able
to tell what work the book contains, by glancing at a
page. If I have just one book in my cell, how hard is it
going to be to remember which book it is? It is the
reader whose desk is littered with open books who is more
obviously in need of such metadata. When one has easy
access to many books, there is an obvious convenience in
being able to see at a glance which is Aristotle and which
is Plutarch and which is Augustine.
The gradual adoption of the title page reflects, I
think, the same development: title pages are important
for quick identification of books, when there are many
lying around. The printers themselves may have been
the first beneficiaries of title pages: once printing
becomes a business, a printer with several books on
hand for sale, will find it easier to manage the stock
if it's easy to see which book is which. Page numbers
also serve a purpose in the printing house, making it
a little easier to check that the imposition of the
forme is correct and that the quires have been
assembled in the right order. The shift from leaf
identifiers like Ai, Aii, Aiii, Aiiii to page numbers
like 1, 2, 3, 4 may be linked to the dawning
recognition that page numbers also have other uses for
other readers.
The use of two-dimensional, visually prominent
paragraph breaks, the use of running heads to identify
chapter or section, and the use of page numbers all
tend to make things easier for a reader who is
scanning through the book looking for something. They
bring less benefit to readers working their way
systematically through the book (though they do help
give us a sense of our progress through the book).
2.2.3
Are the peripheral phenomena? ·
Wesentlich oder unwesentlich?
Neither page numbers nor running heads fit comfortably
into the model of text as a one-dimensional stream of
words without further internal structure. We can
accommodate them into a mechanical representation of
text only if we are somehow able to distinguish them
from the other material in the text stream. The
method most widely used today is some sort of inline
markup (XML, HTML, wiki markup, or other).
We could, perhaps, dismiss page numbers and running
heads and the modern style of paragraph breaks and
section headings as peripheral issues. Yes, of course,
when we print a text on multiple pages it can make sense
to label the pages. But (we might argue) it is the
pages being labeled in this way, not the text: the page
numbers and running heads are artefacts of the material
representation of the text, and not part of the text
itself. (We do not insist that the page numbers remain
the same in a new edition of a text.)
There is no logical flaw in this view; it's merely a
question about the point at which we draw a boundary
between the text and the cultural
practices by means of which we create, preserve,
disseminate, and consume texts.
It can be useful to distinguish clearly between texts
and books. But if we are interested in texts, and if we
are interested in software to help us produce, manage,
or study texts, then we will also be interested in
books, and other forms in which texts can be
represented,
and our software will need to know about books as well
as about texts, or else its depiction of texts will
always feel slightly awkward and foreign to us, in much
the same way as that 1494 printing of Apollonius
of Tyre.
That is, we can narrow the meaning of the word
text, if we like, to exclude
things like page numbers, but then we will need other
broader terms to describe what we are interested in, if
we want to understand texts as cultural objects.
2.3
More complex pages (and texts) ·
Komplexere Text, komplexere Seiten
2.3.1
Annotation, parallel streams, apparatus ·
Anmerkungen, Paralleltexte, apparatus criticus
The idea that text is a simple
sequence of words or characters becomes even less
tenable when we examine texts that are important or
central cultural objects. In almost any culture,
important texts acquire commentaries. Often they
acquire multiple commentaries and levels of annotation,
as illustrated by this page of a [fifteenth-century
Venetian ? check, confirm, or omit] edition of the Bible
(in Latin), together with an interlinear gloss, the
glossa ordinaria attributed to
Walafrid Strabo, and the Postilla of
Nicolaus de Lyra, one of the most popular of medieval
Bible commentaries.
Glossa ordinaria
If one sticks a pin into this page at a random
position, what is the likelihood that the
‘next’ word in the text is the one
immediately to the right of the pin-prick, on the same
line? The model of text as a one-dimensional sequence has
some difficulties with page numbers and running heads, but
it implodes completely when confronted with a work like
this one.
In this page of the glossa ordinaria, we can distinguish
seven text streams which must be synchronized page by page:
the base text of the Bible
an interlinear gloss, explaining the words or constructs of the text
in the inner margins, letters serving as references to corresponding passages in the glossa ordinaria or the commentary of Nicolas de Lyra
the glossa ordinaria
the commentary of Nicolas de Lyra
in the outer margins, references to parallel passages in other books of the Bible (marked by intratextual symbols in the Biblical text)
also in the outer margins, bibliographic references to authors and works mentioned in the glossa ordinaria or by Nicolas de Lyra
Slide: Complutensian Bible 1522
In the case of the glossa ordinaria,
there is a clear primacy of one text over the others: the
Bible text is central (both intellectually and
topographically), the others peripheral. But in other
cases (here the Complutensian Polyglot Bible of 1522),
what we have are multiple versions of what is, in some
sense, the same text (here, the book of Deuteronomy).
These examples of complex works suggest that for
purposes of study, we may find it useful to think of
text not as a one-dimensional stream of characters but
as a bundle of one or more such streams, which are
synchronized at various points. The trick for laying
such material out on the page is to ensure that any
given synchronization point falls on the same page for
all of the parallel streams. This is a complex task, to
put it mildly, which may be one reason there are so few
books which work with as many parallel streams as the
Complutensian Polyglot.
Slide: critical edition
If even the simplest conventional
modern printed books are hardly one-dimensional, we
should perhaps not be surprised to find that more
complex texts are even less one-dimensional, nor that
scholarly editions, intended to support serious work
with the text, provide good examples of typographic
complexity. Conventional printed scholarly editions
provide a number of different kinds of information in
addition to the base text, and rely on typographic form
to distinguish the kinds of information and the
interrelations:
line numbers
prose annotation
apparatus criticus (sometimes there are several apparatus, each devoted to a different class of information, sometimes by classes of text witnesses and sometimes by type of textual variation)
I describe this page as complex by comparison with
other modern books; it is of course not a particularly
complex page by Tustep standards.
2.3.2
Textual variation ·
Textvariation
An extremely rich source of complexity for texts is
the inescapable (if often unwelcome) fact that texts
vary. They shift, they change, they twist and turn.
No text extant in more than one text witness will be
reducible to a single sequence of words, because when
there are multiple witnesses, the witnesses will not
all agree on the exact sequence of words.
[Discussion of ways to think about textual variation.]
[Need simple example, preferably in NHG.
Failing that, using MHG / Minnesangs Frühling.
Wolfram's Tagelied? Unter den linden?]
[Complete parallel texts.]
[Zeilensynoptik - line by line overview of variation.]
[Rhine Delta data structure / variant graph.]
[Parallel segmentation (Thaller).]
2.3.3
Relative cost of stability and change ·
Stabilität und Änderung
[Note: I rather like this comparison, but unless it
has more to do with the topic than appears at the
moment, it probably needs to be cut for time. If
kept, it should be compressed.]
One of the key contributions of printing to textual
scholarship was that it contributed decisively to
textual stability.6 Printing pushes texts toward fixity
by making it possible to disseminate multiple copies
of a given work carrying the identical text (or, in
the case of stop-press variants, nearly identical
texts). If we measure the rate of textual change in
textual variants per copy of a work, printing reduces
the rate drastically -- not to zero, but to a much
lower rate than typical manuscript traditions exhibit.
Textual criticism was conceptually possible before
printing, but the results were difficult to
disseminate because manual copying of a purified text
will introduce new textual changes, and eventually the
details of the text-critical work will be lost.
Textual criticism is somewhat easier in print culture,
because its results are more durable.
In an oral culture, a story will often be expressed in
different words each time it is told; as some of you
will recall, the classicists Milman Parry and A.B. Lord
observed this phenomena among the epic singers of the
Balkans and developed a theory of formulaic poetic
diction to explain how they were able to improvise
metrical narrative at the speed of performance. The
gist of the story is typically (relatively) stable, but
the words will change from singer to singer and from
performance to performance. It is easier to cast the
story in our own words than the remember the exact words
used when we heard the story ourselves.
(Short texts, like lyric poetry, form an obvious
exception to this generalization.)
In a scribal culture, it is much less difficult to
reproduce the wording of the examplar than in oral
transmission: after all, the exemplar is right there
in front of the scribe, and no great feat of memory is
required. Actually, fidelity to the words of the
exemplar is perhaps slightly easier than changing
them, because changing the words requires thinking a
bit about how to phrase the sentence. Reproducing a
word or sentence in one's own dialect, on the other
hand, is perhaps a little easier than preserving the
details of a foreign dialect (as well as, perhaps,
less useful). So from a mental point of view,
preserving the words of the exemplar and changing them
are much more evenly balanced than in the oral
situation. From the physical point of view, it takes
the same amount of physical effort to write out the
words of the exemplar without change as it does to
change them -- except that omitting a word is less
work than copying it. The net effect appears to be
consistent with what we find in our libraries: scribal
copies are typically close enough to their exemplars
that we recognize them effortlessly as carrying the
same text, but absolute fidelity to the exemplar is
rare — essentially non-existent.
Printing changes this equation dramatically: it is much
easier to pull one more sheet through the press without
changing the text than it is to stop the press, make a
correction to the text, re-tighten the forme, and print
the next sheet — and that, in turn, is much simpler and
faster than setting the text again from scratch. That
is, the relative costs of making multiple copies which
are textually identical is much lower than the costs of
making multiple copies each different from each
other.7 The result is that texts,
while still subject to change from edition to edition,
tend to have much more fixity in print than in
manuscript.
The advent of digital media changes the equation
again, to something more like the scribal or even oral
state: the physical cost of making a copy is the same
for faithful and for modified copies: the economic
forces that pushed printed texts to fixity do not
apply. It requires less thought to make a copy
without introducing new changes, so there is still
some pressure toward fixity. But a great deal depends
on the context: the first dictionaries to be available
in electronic form were valuable resources for many
people working in what we now call computational
linguistics, and in order to make the dictionary work
properly with the system in use, computational
linguists routinely reformatted the dictionaries in
various ways, and added new information needed for the
work they were doing. The result is that the textual
history of the electronic dictionaries that were
shared from project to project in the 1970s and 1980s
looks a lot more like the textual history of a
manuscript transmission than that of a purely printed
work: for some early electronic dictionaries,
essentially no two electronic copies of the work are
the same.
3
The shape of text in the machine ·
Die Textgestalt innerhalb der Maschine
If we want to work with text using software, that software
needs to have some notion of what text is. Looking at
possible and actual representations for text in electronic
form can be a useful way of thinking about what text is.
3.1
Characters and character sets ·
Zeichen und Zeichensätze
Picture of punched card
When scholars first began using
information-processing equipment (punched cards, and later
computers) for work with texts, the most pressing issue
for text representation was finding ways to represent the
characters that occur in texts but did not occur on the
keyboards of punched-card machines.
‘Special’ characters, when I began
working with computers, were any that required special
treatment. Since punched cards were limited to the 26
letters of the uppercase Latin alphabet used in standard
English, A through Z, even the distinction between
upper-case letters and lower-case letters required special
measures for some early projects. Such special measures,
of course, required more effort, which cost money or time
or both. For these reasons, the Trèsor de la langue
française, for example, decided against marking the
difference between uppercase and lowercase letters: their
corpus was intended to support lexicographic work on the
dictionary of French being created by the French Academy,
and since they expected to re-check every citation against
a good printed edition before including it in an entry,
they saw no need for the electronic text, intended solely
for project-internal use, to carry case distinctions.
As some of you will remember and others can imagine,
the situation was even more more challenging for
characters with diacritics, like ä, ö, ü, or for
characters used in some Latin scripts but not in English,
like the ess-zett of German or the eth and thorn of
Icelandic.
The development of standardized Greek, Hebrew, Arabic, and
other scripts took even longer.
The development of coded character set standards was slow,
and commercial interests resisted the standardization of
characters they felt were not important for contemporary
commercial use. They objected, for example, to proposals
to provide standard representations for the Cyrillic
characters whose use was abolished shortly after the
Bolshevik Revolution in Russia.
Under those circumstances, scholars interested in
electronic transcription of older texts and texts in
minority languages found it necessary to find their own
ways to represent the characters they needed, via
transcription (as in the Beta Code used in the Thesaurus
Linguae Graecae) or by replacing a given
‘special’ character with a longer
sequence of characters which would not (one hoped) appear
in naturally occurring texts. Tustep's representations
for Greek, Hebrew, and rarer Latin characters fall into
this class; Tustep early developed a larger inventory
of predefined characters than any other scholarly text
processing software I'm familiar with. (Other tools
left the users to their own devices.)
Character set standardization is important, and the
development of the Universal Character Set (UCS) of ISO
10646 / Unicode is a huge advance for everyone interested in
text processing by machine. But if we are interested in
ensuring that our tools can be applied to texts from all
languages, from all periods, and for all scholarly purposes
— and as digital humanists, those goals are natural enough
— then it is essential to recognize that no finite
enumeration of characters will ever suffice to allow us to
achieve those aims. There are writing systems we do not yet
fully understand.
For certain kinds of paleographic work we
must sometimes preserve information about letter shapes
which would normally be omitted from our electronic
representations (which normally work at or near the
graphemic level). The character inventories used in the
Medieval Nordic Text Archive (MeNoTA) distinguish not just
between insular or curved d and erect
(or Continental) d, but four
different forms
of lowercase e, six
different forms of lowercase
a, and
nine forms of lowercase f.
[Delete?]
If our text processing tools are well grounded, we
should very seldom need to define new characters for our
work, because the characters we need will already be
part of the system. (At least, if we are using Tustep.)
But tools intended for scholarly use must also allow us
to add more characters, when and as we need them. It is
for these reasons that XML processors are required to
accept characters in the Unicode ‘Private Use
Area’, even though this thought scandalizes
some people. And it is for these reasons that TUSTEP
developed, so early in its history, such a broad
coverage of the historical Latin, Greek, and Hebrew
alphabets, but also allows user-defined characters when
that is essential.
[Delete?]
[Verify this claim about TUSTEP.]
3.2
OHCO ·
OHCO
As regards the organization of the text above the
character level, over the decades a variety of data
formats were developed, each reflecting in its way a
particular conception of text and of the appropriate
operations on text which need to be supported. I will
not attempt a historical survey of these, although I
believe the hermeneutics of data formats is a rewarding
exercise, and the examination of such formats can
sometimes help shed light on the nature of text (if only
in a negative way), apart from their intrinsic interest.
Instead, I would like to skip forward several decades
and consider what is still, I think, the most serious
attempt so far to describe concisely a general model of
text. I mean the model offered by Steven DeRose, David
Durand, Elli Mylonas, and Allen Renear, then all of
Brown University, under the title “What
is text, really?” (‘Was ist Text,
eigentlich?’).8
As many of you may know, the authors define a model of
text as an “ordered hierarchy of content objects”
(‘geordnete Hierarchie von
Inhaltsobjekten’), widely discussed since as
“the OHCO model”. The authors briefly describe a
number of alternative models of text, and argue for the
superiority of their model, before describing some
remaining challenges for the future. The crucial claim
associated with the paper (and widely called “the OHCO
thesis” or “the OHCO hypothesis”) is: “we
think that our point can be scarcely overstated: text
is an ordered hierarchy of content
objects.” The OHCO model of text is clearly an
attempt by the authors to explain the virtues they see
in the Standard Generalized Markup Language (SGML) --
and now in the successor to SGML, called the Extensible
Markup Language (XML) — and the OHCO model is
frequently described as the philosophical underpinning
of SGML and XML.
The idea that a text of any complexity can be reduced to
a simple hierarchical arrangement of parts strikes many
students of text as ludicrous, so the OHCO model, and
the SGML and XML specs which are thought to be based on
it, have been widely criticized as simplistic,
uninformed, even authoritarian. (Some humanists have a
reflexive distrust of anything called a hierarchy.) Any
book chosen at random will show examples of structural
units which do not nest properly and thus form no
hierarchy: texts are divided into pages (whose cultural
significance we have already noted) and also into
paragraphs, which quite often extend across page
boundaries. But if the elements don't nest, they don't
form a hierarchy.
I would like to defend both the OHCO model and the SGML
and XML specifications against this criticism.
It is probably true that the OHCO model provides a
pretty good match for the approach to text encoding and
processing used by most early adopters of SGML and XML.
The developers of SGML and XML vocabularies identify
different kinds of textual objects, and (in the simple
case) represent them as SGML or XML
elements, which nest within each other and
thus form a containment hierarchy. In a document
production environment, pages are the result of
processing, not the input to processing, and as the
document changes from version to version the pagination
can and should change. So early adopters of SGML
routinely marked paragraphs, but not pages, as
structural units of text.
But the match is not perfect. To believe that SGML and
XML are useful in practice, it is not necessary to claim
that text by nature is an ordered hierarchy
of content objects. Nothing in either the XML or SGML
specifications makes any such claim, explicitly or
implicitly. To find these technologies useful, it
suffices to believe that whatever the underlying
Platonic reality of text (assuming it has one,
which of course Aristotelians and post-modern
constructivists can doubt), text can usefully be
modeled as an ordered hierarchy of content
objects.
And that texts can usefully be modeled using SGML and
XML elements is, perhaps, adequately demonstrated by
thirty years of successful SGML applications.
Se non è vero, è ben trovato.
The SGML spec, moreover, defines some concepts which
directly contradict the claim that a text is a single
hierarchy of objects: the optional feature
CONCUR can be used to allow multiple
element hierarchies to be defined in a document. If we
wish to record not just the play / act / scene / line
hierarchy of Shakespeare's plays, and also the page and
gathering boundaries of the First Folio (since the
correct interpretation of the spelling and text layout
in the First Folio varies from page to page), we can do
so with two parallel hierarchies.
It's also true that even without CONCUR, it is an
oversimplification to believe that SGML (or XML) define
structure only using nesting elements. Both specs
define a mechanism for assigning identifiers (IDs) to
elements and referring to those identifiers from
elsewhere. (HTML hypertext anchors are perhaps the most
widely known use of this idea.) The fundamental data
structure most suitable for reasoning about SGML and XML
documents is not the tree, but the directed graph. The
element structure forms an often useful substructure of
this graph.
If SGML had been based on the assumption that a text
has, by nature, one hierarchical set of content objects,
then the existence of the CONCUR feature and of ID/IDREF
links would be impossible to explain.
It should also be pointed out that while the authors of
the OHCO paper do speak of text as an
ordered hierarchy of objects, their concluding remarks
include the observation that “many documents have more
than one useful structure”. Since they obviously do
not regard this observation as undermining their earlier
claim that “text is an ordered hierarchy
of content objects”, it seems safe to infer that by
“an” ordered hierarchy they meant “at least
one” ordered hierarchy, not “at most one” or
“exactly one” ordered hierarchy.
While SGML and XML appear to be the focus of interest
for both the authors of the OHCO paper and for the
critics of that paper, the model of text as an
ordered hierarchy of objects has a number of other
instantiations which should be mentioned here at least
in passing.
The software system developed in the 1960s by Douglas
Engelbart at Xerox PARC (NLS for "oNLine System", later
re-christened "Augment") structured documents using a
generic three-level hierarchy: unlike SGML, Engelbart's
system did not label nodes in the tree with a type name
(like "chapter", "chapter-title", or "paragraph"), and
it did not allow nodes to nested arbitrarily deep. But
it implemented a model of text as an ordered hierarchy
of anonymous objects. One of the first widely used
PC-based concordance systems, Word Cruncher, also used a
three-level organization for texts. This made it easy
for Word Cruncher to handle texts like Shakespeare
plays, where passages are conventionally identified by
act, scene, and line number, or like the Bible, where
passages are identified by book, chapter, and verse.
But perhaps the instantiation of this model most
familiar to this audience is TUSTEP itself: the
organization of a Tustep file into numbered pages and
lines is a good example of the model. And the ways in
which Tustep can work with pages and lines and their
numbers illustrate how a more powerful model of text
makes it possible to write more powerful software.
3.3
Variations on OHCO ·
OHCO-Variante
One informative way to learn from the OHCO hypothesis is
to consider each part of its description in turn, and
ask what happens to the concept of text if we modify
that part of the phrase.
3.3.1
Unordered hierarchy? ·
Nichtgeordnete Hierarchie?
ordered. What would result if we
assumed an unordered hierarchy of content
objects? The answer is surprisingly clear: the result
is a database management system supporting a
hierarchical model. The object-oriented databases of
the 1980s and 1990s come to mind. The model is
interesting, but it is pretty clearly not a model of
text. For at least the last forty years, with the
rise of the relational database model, databases have
explicitly disregarded the sequence of records in a
table and the sequence of columns in a row.
But texts do not disregard sequence in that way: we
cannot rearrange the chapters of a novel without
changing the novel. We cannot reorder the words of a
poem without destroying the poem.
“Heide
auf
Röslein
Röslein
rot
Röslein
der
Röslein”
is not the opening line of a poem by Goethe.
There are some novels with multiple
reading sequences, like the Argentinian modernist
Julio Cortázar's Rayuela
(‘Hopscotch’
[dt. ‘Himmel-und-Hölle’ oder
‘Hickelkasten’]), illustrate in part the
principle that assigning a different order to the
chapters (and their events) makes a different story.
3.3.2
Ordered hierarchy of ... ? ·
Geordnete Hierarchie von ... ?
content objects. What would result if we
assumed that text were not composed of content
objects? Would it be different? What would it
be composed of?
I find it hard to answer this question, because the
authors offer no definition of the term
content object, and I don't know
with certainty what they meant by this phrase.
Perhaps they mean ‘elements of the logical
structure’ of the text, as opposed to the
elements of the layout structure (so: paragraphs, not
pages, for the typical prose document). This would be
consistent with the rhetoric of most proponents of
SGML at the time they were writing.
If text were not a hierarchy of content objects it
might then be in contrast a hierarchy of page-layout
or rendering objects. This, I believe, is the model
of text implicit in PDF, or of TeX device-independent
files, and might be visible in the internal data
structures of software for those formats.
It is explicitly the model of the World Wide Web
Consortium's XSL Formatting Objects specification.
The XSL FO specification defines a set of XML elements
which correspond to regions on a page, which are to be
filled in a particular way with formatted text.
In recent decades, there have been scholars who have
argued that in order to interpret literary texts we
must study them in their original physical
presentation; they may find meaning in the page layout
and the choice of paper in the first editions of
William Butler Yeats, and they argue strenuously that
the poems of William Blake cannot be divorced from the
script and the drawings in the engravings in which the
poems were first presented. Some people identify this
school of criticism as being interested in the
materiality of text. The model of text
as a hierarchy of layout objects — as, in
effect, a sequence of pages — will appeal to
those interested in materiality in this sense.
As a medievalist, I am also interested in the
materiality of text: a good understanding of the
material conditions of textual transmission is helpful
both in textual criticism and in understanding how
literature was produced and consumed. But I am
reluctant to identify text as a structure (whether
hierarchy, sequence, or something else) of typographic
or more generally text-rendering objects. When we
have no authorial manuscripts, when instead we have
thirty-four manuscripts with different page layouts,
which of the thirty-four manuscripts will be the
text? Is the Nibelungenlied one
literary work, or thirty-four? Materiality is of
great interest, for medievalists as for others. But
precisely because there is almost never a single,
authoritative witness to any ancient or medieval work
(they may be single, but seldom authoritative),
medievalists and classicists find it useful to
distinguish systematically between text (an abstract
object) and text carriers (physical objects that
instantiate the text).
It is helpful to be able to encode pages and other
material objects, in addition to but not instead
of the more abstract and more purely textual objects
like cantos, stanzas, or paragraphs.
Perhaps by “content objects” DeRose et al. mean
merely ‘the objects which occur in the
content’ (of the text, or of other objects), or
‘the objects which carry or realize the content
of the text’. That would, it appears, leave
open the identification of just what objects, and what
kinds of objects, those are. This would be consistent
with the fact that in SGML and XML, it is the user (or
the designer of the SGML/XML vocabulary), not the
specification or the software, who determines what
kinds of objects will be recorded as occurring in the
text. If we are interested in the materiality of our
exemplar, all we need to do is define its content
objects as material ones. This is why XML can be and
is used for the description of page layout or other
specialized kinds of structures.
If we think of text not as a structure of content
objects, and not as a structure of objects at all, could
we still have a model of text? In many computing
contexts, we face the choice between representing
information as data and representing it as a function or
procedure. There have certainly been proposals to model
text as procedures, but they have not been well
received. Procedures are notoriously harder to think
about than data, and it is difficult in many cases to
determine, from examining two procedures, whether they
calculate the same result or not. It can be difficult
to define general conditions for regarding two texts as
the same text (with some textual variation) or as
different (but similar or overlapping) texts. But most
textual scholars think the question is meaningful; it
would I think be a mistake to adopt a model of
text-as-procedure which makes it unanswerable in
principle.
3.3.3
Not a hierarchy? ·
Eine Nichthierarchie?
hierarchy. What if text were not
arranged in a hiearchical structure, not a tree? What
if it has some other form of organization? What other
forms of organization are plausible candidates?
The linguist Igor Mel'cuk observes, in a discussion of
modeling human language, that hierarchical structures
— that is, trees — occupy an interesting
middle ground between completely unconstrained
directed graphs and the most tightly constrained of
directed graphs. “A [semantic representation] is an
(almost) arbitrary network, a graph with practically
no formal restrictions imposed on it. On the contrary
a [phonetic representation] is obviously a string of
phonetic symbols, i.e. a chain, the simplest of all
possible graphs, with the maximum of restrictions
imposed on it ... To establish a many-to-many mapping
between arbitrary networks and chains, a convenient
bridge is needed, that is, a graph formally situated
halfway betwen arbitrary networks, on the one hand,
and chains, on the other. This happens to be a tree,
a formal entity traditionally used to depict sentence
structures ...”
If we wish to think about alternatives to tree
structures for our model of text, we can usefully turn
our attention first to the family of models which
impose tighter restrictions, modeling text as a
sequence (chain, in Mel'cuk's terminology). Note that
a sequence is what we get if we constrain a tree with
the rule that every node except a leaf has exactly one
child.
There will only ever be one leaf. Every node,
except the leaf, will have just one child node, just
as in a sequence every element but the last has just
one next node.
3.3.3.1
Ordered sequence of content objects? ·
Geordnete Folge von Inhaltsobjekten?
We have encountered already the idea of text —
indeed, or all language — as a sequence.
[Delete?]
Some linguists argue that treating spoken utterances
as a one-dimensional sequence of phonemes is an
over-simplification: among other things it omits
pauses, changes of pitch, and changes of speed. (In
conventional writing, punctuation supplies partial
information about these things, but not nearly
enough to represent the prosody of an utterance
completely. Naturally occurring writing are only as
unambiguous as they need to be, to allow written
messages to be deciphered reliably. They need not,
and typically do not, offer anything like a complete
record of an utterance.) Sign languages, meanwhile,
are not spoken with a single mouth but with two
hands, a body, and a face; it is possible to reduce
them to a single sequence of written symbols, but
the process resembles the multiplexing of several
channels of information into a single channel: the
existence of a sequential written form for a
language does not mean the language is best modeled
as a one-dimensional stream of words, morphemes, or
phonemes.
[Delete?]
A wide variety of sequence models have been
defined or implemented; they all model text as
“a sequence of ... X”, but they provide
different values of X.
Perhaps the simplest model of this class identifies
text as a sequence of characters. This model is at
the heart of text editors like emacs or vi on Unix
systems, and ‘programmer's
editors’ of various kinds on other
operating systems.
It is also visible in the string datatypes of our
programming languages, in standard email (I mean
email without HTML markup), and in many Unix
utilities (where it competes with the alternative
model of text as a series of lines).
de
Perhaps the earliest model, historically speaking,
represents text as a sequence of punched cards, with
as many words on each card as feasible, or for verse
texts, with one verse line per card. In the Pfeffer
Corpus of modern spoken German, each card showed the
use of one word in the context of a sentence; when
the sentence was more than 80 characters long,
elisions were used to make it fit.
[No, I do not have an example; cited from memory.]
After punched cards ceased to be used, the model of
text as a sequence of lines remained
prominent, if only in the data structures of text
editors, which represented text files as arrays of
line pointers. It is also visible in a great many
Unix command-line tools like grep, which searches
for strings of characters in a file and returns all
the lines which match the given
pattern.
(And elsewhere?)
de
Beginning in the 1960s, programs began to be written
to allow documents to be formatted for printing by
the computer's printer. Most such programs treated
their input as a sequence of lines, which the
program divided into two groups: some lines
contained the text to be printed, and others
contained processing instructions to guide how the
text should be printed. In the programs Runoff,
roff, nroff, troff, and Script, versions of which
were available for essentially every mainframe
computer, formatting instructions were identified by
a full stop occurring in the first column of the
line (on the theory that in natural language, no
words begin with a full stop; this is not quite true
for English, it turns out, but it
is easy enough for the user to avoid beginning an
input line with such a word.
The most recent of the batch formatting programs,
Donald Knuth's TeX, retains the division of the text
into data and commands, but makes commands
recognizable in other ways, thus eliminating the
structural significance of line breaks in the input.
Its model is (oversimplifying slightly) that text is
a sequence of characters, interspersed with
formatting commands.
Slide of two-dimensional text (to illustrate Busa)
An interesting variant on the card or line model was
developed early on by Roberto Busa. In it, each
card is devoted to a single word, which is recorded
(in its surface form) in the initial columns of the
card. Later columns of the card, in a kind of
tabular format, provide other information about the
word: perhaps its lemma or lookup form, perhaps its
grammatical features, perhaps its syntactic function
or the word number of the word it depends on
(adjectives point to their nouns, subjects and
objects to their verb, etc.). Because the
conventional text can be read vertically down the
beginnings of the lines, this format is sometimes
called a ‘vertical text’.
[citation of Anderson?], but because the additional
information on each card provides a second dimension
orthogonal to the first (what Busa called the
“internal hypertext” of the word), I think it
might better be called two-dimensional text.
Nowadays, the most widespread sequential model of
text in computing, the one most human beings use for
drafting and revising documents, is the common word
processor model. In this model, a text is a
sequence of paragraphs (or paragraph-like objects),
and paragraphs are sequences of characters. In this
model, different paragraphs may be styled
differently: they may have different margins,
different styles of indentation (flush left, flush
right, centered), different leading [Durchschuss],
and different default character stylings. Within a
paragraph, different characters (or sequences of
characters) can be assigned a particular font, font
size, or treatment.
Although the style menu in a typical word processor
may contain entries like Chapter title, Section
title, Subsection title, and so on, word processors
typically have no inbuilt notion that chapters
contain sections, or that sections contain
subsections; they thus cannot enforce a rule that
(for example) a chapter should not directly contain
a sub-sub-section. Indeed, the typical word
processor knows nothing at all about chapters,
sections, and sub-sections, only about chapter
titles, section titles, and sub-section titles.
From a scholarly point of view, or for modeling
purposes, this is perhaps the biggest drawback to
all these flat sequence models: in these models,
there is no natural sense in which the first
paragraph of the first chapter of a novel is
part of or in the first
chapter: the chapter is not an element in the
sequence which constitutes the text, and thus does
not appear in the model. We can of course define
the chapter as a sequence of paragraphs, but that
amounts to changing the model, and it will be
difficult to make the details of the extended model
come out right.
3.3.3.2
Several hierarchies of content objects? ·
Mehrere Hierarchien von Inhaltsobjekten?
If we consider less constrained structures, two
possibilities come easily to mind.
Slide of two-tree text (to illustrate CONCUR).
[replace with German example]
First: instead of a single hierarchy of content
objects, we can have several: two, three, or
arbitrarily many. The image shows a haiku as a
simple example: a haiku with three lines and two
sentences. Blue nodes mark the linguistic
structure, pink nodes the verse structure; white
nodes are shared and occur in both structures.
As has been mentioned, SGML already defines multiple
hierarchies as a possibility, and while the CONCUR
feature was seldom implemented and was dropped from
XML as an unnecessary complication, it is not
conceptually difficult to extend XML to support the
same thing. Oliver Schonefeld and Andreas Witt
proposed an XCONCUR some years ago. More recently,
at least one project has reported using a pure XML
notation to record multiple tree structures in their
documents: the ‘dominant’
structure uses conventional XML elements, and the
‘recessive’ structures use empty
milestone elements. For different kinds of work,
different dominant structures are appropriate, so
the project automatically transforms documents from
one XML structure to the other when needed. This
allows the conceptual model of CONCUR to be applied
in a pure XML context, so that standard XML tools
can be used for processing.
3.3.3.3
Directed graph of content objects? ·
Gerichteter Graph von Inhaltsobjekten?
Second, we can allow the structure of text to be not a
tree, and not several trees, but a graph or directed
graph which is not required to be a tree. Some years
ago, Claus Huitfeldt and I proposed a directed-graph
structure for text, under the name Goddag (generalized
ordered-descendant directed acyclic graph, also
Norwegian for "hello"); the model has been useful for
thought experiments, but we have had neither the
resources nor the desire to build for Goddags the same
kind of infrastructure that already exists for XML:
parsers, validators, transformation languages and
transformation engines, databases and database query
languages, and so on.
A team at the Huygens Institute in the Netherlands
is now undertaking such an ambitious project to
support their own new model of text, under the name
"Text as Graph." In TAG, text is modeled not as a
single tree but as a hypergraph, from which various
trees with different structures can be projected.
[More on Text as Graph would be helpful here.]
3.3.3.4
A practical compromise? ·
Eine pragmatische Lösung?
One substantial drawback of unrestricted graphs is
that they are rather difficult to process
efficiently. Multiple concurrent trees are easier
to process, but are still more complex than handling
a single tree. So even if we regard the underlying
reality of our texts as unconstrained networks, or
the equivalent, we may find trees attractive as
tools for working with text: they allow us to expose
part of the reality in a simpler form, and they
provide a convenient way to mediate between an
unrestricted graph structure for text and the linear
order of bytes which we need in order to store a
document in a file using a file system which models
files as flat sequences of bytes.
In this context, I should mention TUSTEP's use of a
hierarchical model for the management and manipulation
of text, coupled with its ability to transform a
document from one hierarchy to a different hierarchy
(while retaining traces of the original structure, to
allow re-transforming the document in the other
direction). This allows a remarkable flexibility in
the management of documents: there is always a
hierarchical structure of ‘pages’ and
‘lines’, but markup can also be
supplied in the document, and used to transform the
document to a different structure, under user control.
For a skilled user, therefore, Tustep offers the
ability to work with either a model of text as a tree,
or as a forest of multiple concurrent trees, or as a
graph. The power thus given to users is I think one
reason for TUSTEP's success.
4
The shape of text in our heads ·
Die Textgestalt innerhalb des Kopfs
Texts are many things — cultural objects, literary
objects, utilitarian objects — but perhaps above all
they are linguistic objects.
Conventional texts consist of words,
which combine lexical information with morphological
and syntactic information; these words are organized
into sentences which have syntactic, semantic, and
pragmatic information. Sentences convey information
and ideas, directly as part of their meaning,
less directly as part of their entailment,
and very indirectly by their contribution to our
view of the world. Texts quote other texts,
or misquote them, appeal to their authority or refute
them, interpret or misinterpret them. A full
account of the shape of text must include some
mention of these linguistic, aesthetic,
denotational, expressive, conative, phatic,
metalinguistic, and intertextual functions; a fully complete
representation of text as data — if such a thing is
possible — must provide ways to represent such
aspects of text and process them.
... Eine vollständige Beschreibung der Textgestalt
muss diese sprachliche, ästhetische, ... Funktionen
irgendwie erwähnen; eine wirklich vollständige
Repräsentation von Text als Datentyp — wenn
so ein Ding überhaupt denkbar ist — muss es
ermöglichen, diese Texteigenschaften darzustellen
und zu verarbeiten.
These aspects of text are only partially represented in
the written form of a text. Print is, after all, a
technology for writing. And writing systems are almost
always incomplete representations of linguistic
utterances: the function of any writing system is to make
it possible to decipher a message, not to provide a
complete representation of all aspects of the utterance.
Accordingly, writing systems are almost always
underspecified vis a vis the language(s) they are used to
represent: the writing system disambiguates just as much
as it needs to, to be practical, and usually as little as
it can get away with. That means that written utterances
are frequently (strictly speaking) ambiguous, and we rely
on context to help us understand what was meant. A given
Babylonian numeral, for example, can mean 1, or 60, or 360,
and the writing system relies on context to make clear
which meaning is to be assigned to a given occurrence
(25 plus 35 is ... not 1 and not 360, but 60). The token
"M." may mean many things, but when followed by a
nomen gentilicium (e.g. "Tullius")
it can only mean "Marcus".
Writing systems for humans to use will almost always omit
information when they can, to make it easier to produce
written messages. But a writing system for machines can
be more explicit while remaining practically usable, if we
use software to assist in writing and reading it. What
would be involved if we tried to make our electronic
representations of text represent them more completely
than is customary (or perhaps possible) in print? What
might we then be able to do with our electronic texts?
Phonologists may be interested in a phonetic rendering of
the text, either a strictly phonetic one,
which transcribes in more or less detail the sounds
actually made by a speaker, or (as in the text shown in
the image) a phonemic one, which
distinguishes the phonemes of the language and indicated
(for those who can read the International Phonetic
Alphabet) at least approximately how they are pronounced.
Phonemic representations of this kind are, if I
understand things correctly, sometimes used to transcribe
material from languages which do not have an established
writing system.
Students of manuscripts might benefit from a distinction
similar to that made by phonologists between phonetic and
phonemic transcription: for some purposes, a graphetic
transcription is useful, which distinguishes the several
forms of lower-case a shown
earlier. For other purposes, a graphemic
transcription is more convenient. It is sometimes
nice if both levels can be accommodated in the same
electronic representation.
We may be interested in the morphological structure of
the text, which can be modeled in several ways. The image
shows the surface form of the word on the left, in the
center the ‘shallow morphological
structure’ as specified in the Meaning/Text Model
of language created by Igor Mel'cuk and Aleksandr
Zholkovsky, and on the right the ‘deep
morphological structure’ in the same model. (The
full morphological representation of a sentence in the
meaning/text model requires also some representation of
prosody, but I omit it here for brevity.)
The most common linguistic annotation in existing
language corpora consists of assigning a part of speech to
each word. Here we see the Stuttgart/Tübingen part of
speech tag set applied to a sentence from Heidegger, which
has apparently slightly confused the part-of-speech tagger.
(Specifically, on the word
woher.)
The first such corpora to be produced were the tagged
Brown and LOB corpora of modern American and British
English, in the 1970s and 1980s, but many more have
followed. You may notice that the part of speech tags used
have (as here) only a passing resemblance to the eight parts
of speech identified by the ancient Latin grammarians and
which you may have learned in school.
Surface syntactic representation of Melcuk sentence 7
Another popular form of annotation for language corpora
consists in producing a syntax tree for each sentence in the
corpus, either a phrase-structure tree (as in the Penn
Treebank and some others) or as a dependency tree.
(Dependency trees have a popularity in computational
contexts which appears to be out of all proportion to the
popularity of dependency grammars among practicing
linguists.) Shown here is the surface syntactic
representation of a Russian sentence (whose
morphology was glimpsed earlier). Each oval represents a
word in the sentence (or strictly speaking an item in the
deep morphological representation of the sentence). The
surface syntactic structure is modeled here as an unordered
tree: any information expressed by the ordering of words in
the sentence must be represented either in the dependencies
or in the classification (labels) of the dependencies.
(Most dependency treebanks retain the order of the words,
and perhaps as a result can use a much smaller set of
dependency classes.)
The nodes of the theme (topic) are colored light blue;
those of the rheme (comment) are colored pink. Within the
rheme, a subordinate theme and rheme are identified; these
are identified by double ovals in blue or red,
respectively.
I won't attempt to explicate this diagram in detail; as
you can perhaps see, it is a dependency tree rather than a
phrase-structure tree; it differs from dependency trees
created in other approaches primarily in being unordered, in
having (in consequence) a richer set of dependency labels,
and in systematically using not the surface form of the
words but the forms of the deep morphological structure as
nodes. For the purposes of this discussion, it suffices to
notice that the dependencies will require that we record
directional links from each dependent to its governor.
Some linguists define two levels of syntax trees: a
surface level which is relatively clearly related to the
surface form of the sentence, and a deep level which is more
distant from the surface form, but which makes the semantic
structure of the sentence clearer. This example, again from
Igor Mel'cuk and illustrating the meaning/text model,
regularizes some surface syntactic and lexical variations.
The deep structure uses a very restricted set of dependency
types, and replaces some nodes (like "very energetic"
modifying "help") with lexical functions
(Magn).
Several arguments of verbs which are implicit in the
surface form of the sentence and omitted from the surface
syntactic structure are made explicit here, so the words
NAROD, referring to the American people, and STRANA,
referring to African countries, each appear three times, and
are in turn connected by links in the deep anaphoric
structure (represented again here by dotted lines).
Some treebanks attempt to make the semantic structure of
the sentence accessible for processing by annotating the
surface syntax tree; others, as I've just shown, produce a
second ‘deep’ syntax tree. Some
linguists will wish for a separate representation of the
semantics of the sentence; the image shows the
semantic structure postulated by Mel'cuk for
the sentence whose syntactic trees have already been shown.
(In the meaning/text model, this semantic representation
will be the same for all sentences
‘synonymous’ with the one shown earlier.
In this diagram, the semantic structure is modeled as a
directed graph; it would not be impossible to render it as a
set of sentences in logical form, if one wished to reason
formally about the meaning expressed.
I have used the meaning-text model to illustrate the
kinds of information we would need in order to have a
notionally complete representation of the linguistic
structure of text, partly because it is possible to find
fairly clear descriptions of each of its descriptive
layers. Some linguistic approaches may offer simpler,
possibly less complete, descriptions: not all students of
syntax will distinguish two syntactic structures and also a
semantic structure. The ‘semantic’
annotations to the Penn Tree Bank, for example, offer a
single structural representation intended to describe both
the syntactic structure and the associated semantic roles.
And of course we may be interested, in a given practical
project, only in the morphology, or only in the lexis,
or only in the surface syntax of our text.
If a more or less complete representation of the
linguistic structure of a text is so complex and raises
so many questions, we may decide that a complete
representations of everything is a chimera
not worth chasing.
How much chance do we have of being able to process a
query that demands “show me all the passages in Musil,
if there are any, which are best understood as allusions
to Faust,” even if we restrict ourselves
to the views of a single literary historian? How likely
is it that we could make explicit everything
we think about any given text?
Mathematicians sometimes say that for all the mathematical
discoveries of the last centuries,
we are like swimmers who explore one part of a small cove
in detail and know almost nothing at all about the ocean
just beyond. We have only scratched the surface.
Similarly, with text, I think, it is possible that we have
only scratched the surface. Our current methods of
textual representation constitute a set of hard-won
achievements. But there is a great deal more work to be
done. For that work we need the best tools we can have,
and among those tools, I think it is safe to say, is
TUSTEP.
5
Conclusion ·
Schluss
Some years ago, the prominent digital humanist Willard
McCarty (editor of the Humanist discussion list) used to say
humorously that software often reflected not only the
thinking but also the personality of its makers. To
drive the point home, he would say "Think of Word Perfect,
for example. Have you ever seen such a Mormon
piece of software? It is clean, well-behaved, polite almost
to a fault — and also unprepared to deal with anything that
does not fit into its world view, and absolutely,
irredeemably convinced that it knows the Right Way to do
things." To the extent that software is a tool, it
reflects the convictions of its makers about what tasks are
worth doing. Not only that: software will also necessarily
reflect the makers' beliefs about how one can (or should, or
must) set about performing those tasks, and (a little less
directly) about the human being or beings who will be using
the software: what they find important, what they find less
important, how they think about their task and how much
responsibility they wish to take for it.
There are not many pieces of software that make me wish
to live up to the software's implied image of
its user. TUSTEP is one of these.
On this analysis, every piece of text processing software
reflects an idea of text, of the things we might want to do
with text, of the ways we might want to shape text by
processing and (for those programs that produce output) the
ways we might want to shape it on the page. Sometimes, the
ideas reflected are simple, facile, even simplistic. Not
many programs offer a concept of text and the shaping of
text as detailed, as deeply felt, as coherently thought
through as that of TUSTEP. As always, though, programs
reflect the thought of those who make them, and it seems
fair to say that with TUSTEP, Wilhelm Ott has shown himself
to be one of our day's great thinkers about text, and the
shape of text, and the shaping of text.
Please join me in honoring Wilhelm Ott and thanking him
for his work.