Image Description at the William Blake Archive

The larger objective of this paper is to follow the lead of Matthew Kirschenbaum and the participants in his panel at last year's ALLC/ACH conference in Charlottesville, Virginia, in thinking about visual electronic resources as structured data. More specifically, I propose to turn the spotlight on a significant portion of the William Blake Archive's metadata: its SGML-encoded image descriptions. Introductory remarks will emphasize the image descriptions as a design feature intended to maximize use of and complement the Archive's other image resources, but the balance of the paper will consider more foundational issues: what exactly are image descriptions in formal terms? What is their relationship to their first-order objects? What kinds of problems do they present for the project team? What functions do they serve? My aim is to provide a broad overview of our practices and the challenges we face (both theoretical and practical) in describing images. The full version of this presentation will attend closely to the discursive features of the image descriptions and their principles of inclusion of pictorial data on the grounds that they serve as a barometer of the general reliability of the search and retrieval functions.

The image descriptions and their characteristic terms - which are best considered as a unit - are an integral part of the Archive's paradigm of image search and retrieval. General descriptions are available via a link to a separate window from all Object View pages in the Archive, and more specific descriptions are returned in the course of an image search. The descriptive commentary is also available to the user who invokes an Inote session from any of several windows and pages. (Inote, which has received much press in humanities computing circles since its public debut, is a java-based image annotation tool developed at IATH that superimposes a four-quadrant grid on an image. Clicking an area of interest opens a separate annotation viewer containing the editorial commentary targeted to the selected region.) The image descriptions are conceptually inextricable from the Archive's characteristic terms, a menu of which the user selects from when launching an image search. Though it is tempting to devote a portion of the paper to an evaluation of how the Archive's controlled vocabulary measures up when compared to other established classification and vocabulary browsers (such as IconClass, the Art and Architecture Thesaurus, and the Library of Congress Thesaurus for Graphic Materials), that subject is sufficiently complex to warrant separate treatment at some future date.

Strictly speaking, the Archive's search engine consults only the SGML-encoded characteristic terms when returning hits to a user. The exact function of the prose that appears between <illusobjdesc> tags is more difficult to define. My provisional account of the purpose of the descriptions is threefold: first, they illustrate that the visual-to-verbal transposition isn't a simple one-step operation: more complexly, images are first captured in prose and from the prose we extract smaller searchable units. Second, the descriptions provide a system of checks and balances that allow the user to cross-reference them with the characteristic terms, making the underlying logic of the hits returned in the course of a search session explicit. Third, they offer us a space where Willard McCarty's metadata rule of disambiguation can be violated: as I will explain in more detail in my talk, a fair amount of uncertainty and doubt is built into the descriptions as a response to iconographic ambiguities in the source material; this interpretive uncertainty is, if not effaced entirely, at least diluted at the characteristic level. These functions aside, it is my sense that the descriptions aren't being mined for information as effectively as they could be, though the beta version of WBA 2.0, scheduled for public release later this summer, should make it possible to harvest their data in new ways.

Because the Archive's image search and retrieval software, Dynaweb, must consult an SGML information base that is, of course, textual rather than pictorial in content, the descriptions and their corresponding characteristic terms serve as a verbal proxy for their visual objects. To put it otherwise, the annotations function as metatext (rather than, as in conventional print relationships, paratext) to the primary data, creating an unusual onus of responsibility for the editors and assistants to provide descriptions and terms that show as much fidelity as possible to the originals in order to guarantee optimum results for the user searching across the visual collection.

This process of faithfully, accurately, and thoroughly translating from one medium (the visual) to another (the verbal) is, as one would expect, beset by difficulties. But we are not without precedent or aid. In particular, the problems we encounter in the course of composing our image descriptions can be profitably understood by looking to the growing body of literature in Humanities Computing on encoding transcripts of source materials. My suggestion assumes that at heart the two endeavors (describing images and transcribing texts) are first and foremost acts of translation - the former from pictorial data to linguistic data, the latter from - to take the example of the Beowulf Project - chirographical marks in a codex manuscript to the descriptive codes of SGML - and as such are prey to all the difficulties inherent in the translation process.

At last year's ALLC/ACH conference, Paul Caton and Julia Flanders raised the issue of the measurability of data as an important consideration for project teams interested in encoding renditional information: they stressed the necessity of

"having a clear sense of what the units of information are, how to identify and distinguish them consistently, and how to record them accurately. It [the criterion of measurability] represents/requires an attempt to decide . . . what threshold of perceptibility will be maintained: . . . what will be considered either too small or too costly or difficult to record."

At the Blake Archive, our efforts to extract iconographic content from the source material for the purposes of image description are routinely complicated by just the kinds of measurability and perceptibility difficulties outlined by Caton and Flanders. Should Blake's interlinear motifs be captured in as much detail as the large-scale designs? Should the hatched lines that appear meaningless in the 100 dpi image but representational in the 300 dpi be described (and thus made searchable)? Should we even avail ourselves at all of the 300 dpi images as resources to transform our unaided human eyes into Argus eyes, or is this approach contrary to the central tenets of diplomatic editing? What about the left arm of that central figure on plate 16 in copy D of The Marriage of Heaven and Hell, which was printed but colored on the impression so that it blends almost imperceptibly with the adjacent figure's gown? Is that wispy line extending from the ascender of the letter "b" a bird or a vine, or is it both? Indeed variations on the last question may reflect the most common kind of perceptual ambiguity the annotator of Blake encounters, attributable in part to Blake's almost profligate use of metaphors; his peculiar syncretic insight makes it dangerous to ever rule out the possibility that a design motif has multiple referents: flora and fauna are often indistinguishable, a human figure may be a composite of several characters, and a lock of hair easily morphs into a coiled serpent.

The final version of the paper will address our working solutions to the problems enumerated in the foregoing paragraph. I will conclude by briefly shifting the emphasis from the creation of image descriptions to their potential use, offering some parting thoughts and caveats on the kind of knowledge that the image descriptions might produce when used in conjunction with the Archive's other image tools and resources.

References

Caton, Paul and Flanders, Julia (1999). "Encoding Rendition in Primary Source Texts Using TEI." Unpublished essay, 1999. Presented at ACH-ALLC 99. University of Virginia, Charlottesville. 12 June 1999.
Eaves, Morris, Essick, Robert N., and Viscomi, Joseph (eds) The William Blake Archive <http://www.iath.virginia.edu/blake/>.
Inote. Software from the Institute for Advanced Technology in the Humanities. <http://www.iath.virginia.edu/inote/>.
Kirschenbaum, Matthew, et al. (1999). "Refining Our Notions of What (Digital) Images Really Are." ACH-ALLC 99. University of Virginia, Charlottesville. 10 June 1999.
McCarty, Willard. "Humanities Computing as Interdiscipline." <http://www.iath.virginia.edu/hcs/mccarty.html>.