Parallel witnesses are in many ways a luxury: several instances of a textual source mean fewer problems due to lacunae or lost material, a chance to compare for meaningful differences and less vulnerability to idiosyncrasies or errors in one manuscript. At the same time, they present both technical and methodological challenges in the evaluation of their contents. Much research and software development related to parallel witnesses has been interested in textual criticism, stemmatology and the construction of a critical apparatus, see for example Bakker 1996, Clement 2011; and for software see Juxta, Wheeles 2014, as well as the Versioning Machine (, and TEICHI (see Pape et al. 2012 for discussion). This paper will be concerned with a different aspect of work with parallel witnesses and textual criticism – their consequences for encoding redundancy and subsequent behavior in quantitative analysis.

The case of Coptic SCRIPTORIUM

As a case study on the issue of corpora containing diplomatic transcriptions of parallel witnesses, we will consider the guidelines currently used in Coptic SCRIPTORIUM, a collaborative project publishing open access diplomatic transcriptions of Coptic manuscripts online ( The diplomatic transcriptions are available from the project website in several formats, including Epidoc TEI following Cayless et al. (2009), and are made searchable and visualized using ANNIS, a browser based corpus search tool (Krause & Zeldes, to appear). Figure 1 illustrates an excerpt from a document body in TEI, and a corresponding visualization from ANNIS (the word diabolos ‘devil’ has been highlighted in Greek script in both views to show the same position).

Figure 1: The word diabolos in code and visualization

Visualization and TEI fragment for part of a manuscript from Archmandrite Shenoute’s I See Your Eagerness (Witness GL 29-30).

One of the advantages of publishing diplomatic transcriptions is that, unlike editions which remain under copyright, materials can be offered over an open access license. At the same time, we do not have the luxury of an edited, authoritative text: for documents that have multiple witnesses, witness encoding becomes an issue, rather than encoding facilities for a critical apparatus. Luckily, metadata standards are generally capable of handling witness information, which we encode in the header to the fragment above as follows, using the TEI <listWit> element as part of the <sourceDesc> specification, shown in Figure 2 for a manuscript abbreviated GL, which parallels another manuscript known as XJ.

Figure 2: Parallel witness encoding

Encoding parallel witness information in the TEI header with <listWit>

The problem begins when we consider that parallel witnesses do not cover the exact same span of text (there are parts in witness GL which do and do not correspond to witness XJ), and ask how we can search through our corpus for quantitative research.

The ANNIS platform allows us to search for the categories annotated in the corpus and get frequency breakdowns for items, such as a ranked frequency list for words of Greek origin (notice the xml:lang attribute in Figure 1). A search for Greek words in all data from the work I See Your Eagerness gives the frequency list in Figure 3, which shows ‘diabolos’ appearing twice. However, this is a direct result of the parallel witness to the section mentioning the Devil, thus skewing the quantitative results.

Figure 3: Frequency breakdown

ANNIS frequency list for Greek words in I See Your Eagerness.

In order to deal with this issue, we require a mechanism to ‘count things only once’. The problem is that we cannot simply exclude entire manuscripts as redundant, since in the current case, each manuscript contains some parts that are not found in any other manuscript. We therefore decided to divide each manuscript into minimal portions that correspond to the parts that do and do not have parallel witnesses, annotating each portion with parallel witness information. We consider only the most complete witness of every section to be the ‘primary’ source in answering queries such as the frequency list above, whereas parallel portions are considered ‘redundant’. This is illustrated in Figure 4.

Figure 4: Partitioning parallel witnesses

Parallel witnesses. The primary witness for the middle, overlapping section is on the left in Portion 2, while Portion 3 is marked as ‘redundant’. Portion 3 in the second MS is unique, and therefore non-redundant.

In the ANNIS interface, the user can choose to limit queries to non-redundant information, or to search through all the data, including duplicates.


The approach outlined above has been implemented in the ANNIS database and interface for Coptic SCRIPTORIUM corpora, but it is a point for open discussion what the best way is to encode parallel witness information in TEI XML markup, and how to designate information about ‘redundant’ copies. What is the best way to mark up a part of a document as having other witnesses? How should we mark up the privileged ‘primary’ witness portions and the ‘non-primary’ ones? Can we ‘mix and match’ in multiple documents? Is this a type of metadata or inline annotation? Methodologically too, the issue of choosing what to call redundant is non-trivial, as are alternative solutions we might have chosen. A possible alternative for quantitative evaluation is to consider all witnesses in every search and normalizing frequencies by number of witnesses, such that frequencies from a passage attested twice are weighted to be ‘worth half as much quantity’. This will produce different results if passages are marked as parallel which contain small differences. We hope that further discussion of these and other strategies will draw attention to, and improve the handling of, parallel witnesses in quantitative research.


I would like to thank Caroline T. Schroeder for contributing to the data model and markup decisions in this dataset and David Brakke and Rebecca S. Krawiec for providing the data and annotations described in this paper. I am also grateful to the National Endowment for the Humanities (NEH) for their continuing support of this project.


[Bakker 1996] Bakker, H. P. S. (1996). Towards a Critical Edition of the Old Slavic New Testament: A Transparent and Heuristic Approach. PhD Thesis, University of Amsterdam.

[Cayless et al. (2009)] Cayless, H., Roueché, C., Elliott, T., & Bodard, G. (2009). Epigraphy in 2017. Digital Humanities Quarterly 3(1). [cited 13 July 2015].

[Clement 2011] Clement, T. (2011). Knowledge Representation and Digital Scholarly Editions in Theory and PracticeJournal of the Text Encoding Initiative 1. [cited 13 July 2015]. doi:

[Krause & Zeldes, to appear] Krause, T., & Zeldes, A. (to appear). ANNIS3: A New Architecture for Generic Corpus Query and Visualization. Digital Scholarship in the Humanities . ["early access" online publication 25 October 2014. doi:]

[Pape et al. 2012] Pape, S., Schöch, C., & Wegner, L. (2012). TEICHI and the Tools Paradox. Journal of the Text Encoding Initiative 2. [cited 13 July 2015]. doi:

[Wheeles 2014] Wheeles, D., & Jensen, K. (2014). Juxta Commons. Journal of Digital Humanities 3(1). [cited 13 July 2015].

Author's keywords for this paper:
TEI; parallel witnesses; quantitative; metadata; computational linguistics; information retrieval; Coptic

Amir Zeldes

Assistant Professor of Computational Linguistics

Georgetown University

Amir Zeldes is a computational linguist specializing in corpus linguistics. His main area of interest is the syntax-semantics interface, where meaning and knowledge about the world are mapped onto lexical choice and syntactic structure in language-specific ways. He is also involved in the development of tools for corpus search, annotation and visualization, and has worked on the development of standards for the representation of textual data in Linguistics and the Digital Humanities.