Stylometry in manuscript studies!?
When I signed up for the workshop on “Digital Manuscript Analysis” (organized by the University of Hamburg’s Centre for the Study of Manuscript Cultures at DH2012), I expected this to be something relatively disconnected from text analysis in literary studies. Author attribution studies as they are being widely practised today are very much focused on transcribed text, which allows to establish frequency lists of words or letters for each text, and to calculate statistical distance measures for different texts based on those frequency lists.
How do manuscripts come into play here? Many times, there is no transcribed text, but rather a digital facsimile and some metadata about it, especially about material aspects of the manuscript, and about contextual information. However, as far away as parchment and ink or infrared imaging of palimpsests may seem from stylometry, they are in fact very close in some respects, or at least in some uses of them.
For my part, I was unaware of such uses. For instance, there is a lot of work going on in manuscript studies, where sub-letter features are automatically retrieved from a manuscript: for ideogrammatic writing systems, aspects like stroke angles, stroke lengths, stroke width, average stroke width, etc. can be retrieved. And more generally, things like line heights, average line width, range of line width, etc. can be retrieved. Also, the chemical composition of inks or of parchment and paper can be analysed for each manuscript or manuscript fragment and recorded as metadata.
Some projects (at this workshop, this was especially the Tibetan Buddhist Resource Center, New York) are using such information to calculate statistical distance measures between various manuscripts or manuscript fragments. Of course, the research questions here are slightly different from author attribution studies as we know them: rather than wanting to attribute an author to a given as yet anonymous text, manuscript scholars are more interested in finding fragments belonging together for some reason, whether by scribe, place, time, or scroll, to create more complete manuscripts or to create context through a number of closely related manuscripts.
What I don’t know, and would love to find out, is what kinds of algorithms these projects are using for statistical distance measure. Are these the same algorithms as in authorship attribution, like Burrows Delta, or are they different? If they are different, is that for historical and institutional reasons, or really because different algorithms are better for such data? Could the algorithms and tools from textual authorship attribution be used in manuscripts?