A Fresh Computational Approach to Textual Variation

If there is one thing that can be said about the entire
literary output of the world since the invention of
writing it is that literary works exist in multiple versions.
Such variation may be expressed either through the
existence of several copies of a work, through alterations and errors usually in a single text, or by a combination of the two. A textual feature of this degree of importance
ought to be at the forefront of efforts to digitise our
written cultural heritage, especially at a time when
printed media are becoming less important. Until now literature has been represented digitally through systems of markup such as XML, which are ultimately derived from formal languages developed by linguists in the 1950s (Chomsky 1957; Hopcroft and Ullman 1969); but over recent years it has gradually become clear that the hierarchical structure of such languages is unable to accurately represent variation in literary text. Alan
Renear (1997), for example, admits that variation is one exception that does not fit into his hierarchical model of text; likewise Vetter and McDonald (2003) conclude that markup provides ‘no entirely satisfactory method’ for
representing variation in the poetry of Emily Dickinson.
More general discussions of the shortcomings of
hierarchical markup, including the problem of variation, have recently been made by Dino Buzzetti (2002) and Edward Vanhoutte (2004).
An alternative approach, not yet tried, is to use graphs to represent variation. Graphs were first studied in the 18th century by the Swiss mathematician Leonhard Euler, who is best remembered for his solution to the famous
‘Bridges of Königsberg’ problem (Trudeau 1993). The type of graph which most closely resembles textual
variation does not appear to have yet been described by anyone; however, it can be derived from the following example. Consider four versions of the simple sentence:
A The quick brown fox jumps over the lazy dog.
B The quick white rabbit jumps over the lazy dog.
C The quick brown ferret leaps over the lazy dog.
D The white quick rabbit playfully jumps over the dog.
Collapsing the five versions into collapsing the four versions. Such repetitions are clearly undesirable. If they were present in an electronic edition each time one copy was changed, an editor would have to check that the other copies were changed in exactly
the same way. If all this redundancy is removed by
collapsing the four versions wherever the text is the same,
the following graph results:
Figure 1
This is a type of ‘directed graph’, which we call a
‘textgraph’. Its key characteristics are:
a. It has one start and one end point.
b. The ‘edges’ or ‘arcs’ are labelled with a set of
versions and with a fragment of text, which may be empty.
c. There are no ‘directed cycles’ or loops.
d. It is possible to follow a path from start to end for each version stored in the graph, which represents the text of that version.
In figure 1 version D contains an insertion: ‘playfully’ and a deletion ‘lazy’. These are represented in the graph
as empty edges. In fact insertions and deletions are
the same thing viewed from different perspectives: every deletion is an insertion in reverse and vice versa.
Transpositions, as in version ‘D’ - the transposition of ‘white’ and ‘quick’ in relation to version ‘B’, can be viewed as a deletion of some text in one place and its insertion in another. All that is then needed is some way to refer back to the original text to avoid copying, e.g. by ‘pointing’ to it. This feature has been shown in figure 1 by drawing the transposed text in grey, which does not change the structure of the graph.
This model is equally applicable to variation arising from a single manuscript or from the amalgamation of
multiple manuscripts of the same work. Its biggest
advantage is that it can handle any amount of overlap without duplicating text. One example of a rigorous test of this model can be found in the archives of the ‘Digital Variants’ website. The poem ‘Campagna Romana’ by the modern Italian poet Valerio Magrelli exists in four drafts, the first of which is shown in figure 2.
Figure 2
In original manuscripts like this it is often unclear how variants are to be combined. For example, in the line
‘Il suo arco sereno/certo/scandito/ ha la misura d’un
sospiro/misura la sera/’ it is impossible to say if there ever was a version: ‘Il suo arco scandito misura la sera’. The sensible way to proceed here is simply to provide a mechanism for recording any possible set of readings,
and to leave the interpretation up to the editor.
Figure 3
Documents which are based on this graph structure
we call ‘multi-version documents’ or mvd’s. One
application of this format is the applet viewer shown in
figure 3 (Schmidt, 2005). This currently allows the user
to view one readable version or layer of text at a time.
In reality only the differences between each layer are
recorded, and the user can highlight these using red to
indicate imminent deletions and blue for recent insertions.
The text is also searchable through one version or all
versions simultaneously. This visualisation tool is in
an early stage of development and as yet it can only
handle plain text. However, because it cleanly separates
the content of the document (represented by the edges of
the graph) from its variation (represented by the graph’s
structure), the same method could also be used to record
versioning information in almost any kind of document
- including XML, graphical, mathematical and other
formats:
Figure 4
This allows a multi-version document to utilise existing
technology. By removing variability from a text, and
effectively representing it as a separate layer, the mvd
format allows technologies like XML to be used for what
they were designed to do: to represent non-overlapping
content. One way this could be achieved would be
to edit the text in an existing editor but to modify the
editor slightly so that instead of reading and writing
the document directly it would read and write only one
version at a time to an mvd file, as shown in figure 4.
There are a couple of possible objections to the overall
technique described here. Firstly, because it is not based
on markup, it is no longer practical for the user to see the
contents of the document in its merged form. Secondly,
existing XML technology currently utilises markup to
record information about the status of individual variants.
This data would have to be re-encoded as characteristics
of the bits of varying text, since the document content
would no longer carry any information about variation.
However, the very idea of ‘variants’ embedded in the
text is a structure inherited from the critical edition,
which is now widely regarded as obsolescent (Ross
1996; Schreibman 2002). Through the printed medium
traditional philology advanced the notion of textual
‘truth’ in its effort to restore a lost original, whereas our
model is directed toward the fruition of the text as it
really is. As we move forward into an age when
digital text has the primary focus, some of the old ideas
associated with paper-based methodologies may have to
be revised or given up entirely (Fiormonte 2003).
In conclusion, the use of ‘textgraphs’ to represent variation
appears to overcome the problems of redundancy and
overlap inherent in current technologies, and to reduce
document complexity. Thus far, a file format has
been devised and has been demonstrated in a working
multi-version document viewer for plain text, which
is capable of representing original documents of high
variability. By separating variation from content it also
has the potential to leverage existing document handling
technologies. This technique represents a new method
of handling textual variation; it is mathematical and
wholly digital in character, and unlike what it purports to
replace, it is not based on the inherited structures of the
printed edition. References
Buzzetti, D. (2002) Digital Representation and the Text Model, New Literary History, 33(1): 61-88.
Chomsky, N. (1957) Syntactic Structures, Mouton & Co: The Hague.
Fiormonte, D. (2003) Scrittura e filologia nell’era
digitale, Bollati Boringhieri: Turin.
Hopcroft, J.E. and Ullman, J.D. (1969) Formal
Languages and their Relation to Automata,
Addison-Wesley: Reading, Massachusetts.
Renear, A. (1997) Out of Praxis: Three (Meta) Theories of Textuality in Electronic Text, Sutherland K. (ed.), Clarendon Press: Oxford, 107-126.
Ross, C. (1996) The Electronic Text and the Death of the Critical Edition in R. Finneran (ed.), The Literary Text in the Digital Age, 225-231.
Schmidt, D. (2005) MVDViewer Demo, available at: http://www.itee.uq.edu.au/~schmidt/cgi-bin/MVDF_sample/mvdviewer.wcgi
Schreibman, S. (2002) The Text Ported, Literary and Linguistic Computing, 17: 77-87.
Trudeau, R.J. (1993) Introduction to Graph Theory, Dover: New York.
Vanhoutte, E. (2004) Prose Fiction and Modern
Manuscripts Limitations and Possibilities of Text-Encoding for Electronic Editions in Unsworth, J., O’Brien, K. O’Keeffe and Burnard, L. (eds.), Electronic Textual Editing. (forthcoming - available at http://www.kantl.be/ctb/vanhoutte/pub.htm#arttg)
Vetter, L. and McDonald, J. (2003) Witnessing
Dickinson’s Witnesses, Literary and Linguistic Computing, 18: 151-165.