The problems of fine-grain text to image alignment and its recording in standard interchangeable XML format were brought about by the Oriflamms Research Project aiming at establishing an ontology of forms and at analyzing the graphical structures of Medieval scripts (Latin and vernacular). Such an ontology would be useful for further studies in linguistics and various disciplines concerned with the history of scripts (palaeography, epigraphy, diplomatics, etc.). It would also provide valuable data for the development of software for optical recognition of medieval handwriting (HTR). One of the main project deliverables is a corpus of transcriptions aligned to facsimile image zones at word and character level, encoded in a standard interchangeable format (the choice was made to use TEI XML) and distributed under a free license.
To attain its goal, the project brought together researchers and engineers from several
research labs in the fields of Humanities (palaeography, epigraphy and linguistics)
computer science (image analysis), as well as from a commercial enterprise specialized
optical character recognition. The partners of the project contributed a number of
or inscription transcriptions and corresponding digital images of the source documents.
the transcriptions were TEI XML encoded but the precise markup schemas varied considerably
depending on their purposes (historical, linguistic or literary studies), on the philological
traditions followed (
critical text editing) and
on their age (from late 1990-s to 2013).
Input Transcription Types and Formats
Before presenting technical issues that had to be solved in the project, we will briefly describe the basic transcription types used in paper editions, and the ways they can be marked up and combined using computer technologies. The details and examples provided are mostly taken from the editions of Medieval Latin and French manuscripts but similar traditions exist for other European languages.
Palaeographic (Allographetic) Transcription
Transcriptions of this kind are also called
imitative but we prefer the
allographetic that corresponds to the linguistic status of letter
variants in the framework of the theory of graphemics [Coulmas 1999].  This
kind of transcription represents very faithfully the aspect of the source document
It distinguishes major letter variants, like
s (ſ vs. s) or
r ( vs. r),
reproduces abbreviation markers, original punctuation and word segmentation. As a
allographetic transcriptions reproduce (or represent somehow) the layout of the original
document (including line breaks, scribal correction markers, marginalia, etc.).
Allographetic transcriptions are rare in paper editions and can mostly be found in
handbooks and albums for palaeographers (e.g. [Koschwitz 1879] and [Careri et al. 2001]). In electronic textual editing
transcriptions are also very rare: the original manuscript transcriptions of the Charrette Project are one the
few examples that we could find. More often, some allographetic features can be found
multi-layer electronic transcriptions that will be discussed later.
This kind of transcription is the easiest to align with image zones. However, problems may arise in the process of word tokenization, as the white spaces that appear in manuscripts (and reproduced as they are in allographetic transcriptions) do not always correspond to word limits in terms of modern linguistic analysis. Another problem may concern abbreviation markers, some of which (but not all) are superscribed over the baseline characters. This has however more to do with the technique of image segmentation than with the transcription itself.
Diplomatics is a scholarly discipline that studies the tradition, the form, and the production of written acts (especially historical documents) [VID 1997: 21]. In diplomatic editing, it is important to preserve as much information as possible from the source document without overcharging it with information not relevant for its interpretation (such as, typically, letter variants, or allographs). The practice of diplomatic editions has been far from being stable along its history and different national traditions. Some normalization efforts were undertaken by the Commission internationale de diplomatique (http://cidipl.org) but their recommendations [Bautier ed. 1984] are very general and allow considerable variation in local practices. It is nevertheless possible to point out some typical features of diplomatic transcriptions. For instance, the abbreviations are expanded but the letters supplied are typographically marked (using italics or some kind of braces). The original document layout is usually preserved. The punctuation and word segmentation may be normalized according to modern typographical rules. Some diplomatic editions also use character disambiguation (introducing u / v and i / j phonetic distinction and adding some diacritics that help reading the text) and capitalize the proper names. In some cases the editor may supply a missing word or part of word, or point out to a repeated word or phrase but these emendations are always clearly marked typographically. As an example of modern diplomatic edition we can quote Les plus anciens documents linguistiques de la France published online by Zurich University (http://www.rose.uzh.ch/docling).
Diplomatic transcriptions can in principle be aligned to image zones, provided that the abbreviation expansions, scribal corrections and editorial emendations are properly tagged. The punctuation marks may need to be ignored if the punctuation is modernized in the transcription. The neutralization of allographs, disambiguation of some characters and capitalization of proper names do not affect the alignment but need to be addressed in the markup model.
Normalized (Critical) Transcriptions
Transcriptions of this kind are the most widespread in the field of literary text
editing. As far as the Old French texts are concerned, the first
normalization recommendations (although the term was not used) were
published by a commission of Société des Anciens Textes
Français in 1926 [Roques 1926]. These recommendations were
precise enough at some points (like using diacritics) but a number of questions (like
separation) were not mentioned at all. In early 2000s, École Nationale
des Chartes published a series of recommendations for editing various kinds of
medieval texts (Latin and vernacular, literary and documentary) [Vielliard ed. 2001, Guyotjeannin ed. 2001 and Bourgain ed. 2002] but they are not always followed in practice.
Without speaking of
Lachmannian) critical editions based on the comparison of multiple witnesses, which are by
definition not alignable to an image of a single manuscript, even relatively
faithful normalized transcriptions are hardly usable for the purposes of
alignment. Whereas major editorial emendations to the source document are usually
marked (e.g. supplied words or letters are placed into square brackets), some information
important for alignment is lost: the expansions of abbreviations are unmarked, as
scribal corrections, and the manuscript page layout is rarely preserved (except, to
certain extent, the verse lines). However, thanks to digital markup, critical transcriptions
may become more
alignment-friendly. In this case, they should not be referred
plain critical but rather as
Hybrid Digital Transcriptions
The possibility offered by computer technologies to have both abbreviation markers
expansions, corrected or regularized and original erroneous or irregular forms is
one of the
main arguments in favor of producing digital editions. The TEI provided special elements
these purposes since its very early versions. Initially, the alternative versions
encoded by means of attributes (e.g. the
<expan> element had an
abbr attribute), but the P5 version introduced a
<choice> element in order to group alternative encodings of the same segment
of a text. This mechanism is much more flexible and powerful than the old attribute-based
but it may require more complex processing. TEI allows using more than two child elements
<choice>s may nest. Some restrictions
had to be added to the content model of
<choice> for the alignment project,
such as allowing only two child elements. Some more restrictions and requirements
alignable transcription format will be described in section “Target Format”.
The actual transcriptions used in the project can all be qualified as
hybrid but precise encoding practices differed considerably. Characteristic
features of three of them are listed below.
The Princeton Charrette Project manuscript transcriptions were initially encoded in
TEI P3 SGML with extensive use of SGML entities for
These transcriptions were purely allographetic. They were automatically converted
(extended TEI P4X) in 2002, and are currently available in this format under a Creative Commons BY-NC-SA
2.5 License from the project legacy website. These transcriptions provide information on allographs
grouping (e.g. that
s are variants of the same letter), and supply expansions
for some abbreviations, but the latter have never been checked and contain many errors.
The following example illustrates a verse line transcription from the Charrette project:
<l n='1' id='MS-A-196-r-a-1' key='FU-31'> <chr_large size="8" color="" detail="">A</chr_large> un jor dunea<chr_var letter="s" var="long">∫</chr_var>cen<chr_var letter="s" var="long">∫</chr_var>i<chr_abbr type="" class="" expan="" cert="no">[apost]</chr_abbr>o </l>One can see that project specific elements are used for abbreviations, allographs, and large capitals, and that no expansion is provided for the abbreviation in the end of the line.
Graal Multi-Layer Transcriptions
The digital edition of the Queste del Saint Graal
(manuscript Lyon, Bibliothèque Municipale, P.A. 77) [Marchello & Lavrentiev ed. 2013]
allows downloading its TEI XML source files under a Creative Commons BY-NC-SA
3.0 License. The Old French part of the edition is encoded in TEI P5 XML using
the BFM-MSS extension. This extension includes, in turn, three extension elements
by the Medieval Nordic Text Archive
Initiative Project (MENOTA):
that are wrapped into
<choice> and placed inside
tag every word or punctuation mark of the text, as shown in the example below:
<w type="NOMpro" xml:id="w106_000286"> <choice> <me:norm>Lancelot</me:norm> <me:dipl>lanc<ex>elot</ex></me:dipl> <me:facs>lan<bfm:mdvAbbr>c̅</bfm:mdvAbbr></me:facs> </choice> </w>
The three Menota elements correspond roughly to the three transcription types
described above. One can observe that the initial l is
capitalized in the normalized layer (because Lancelot is
a personal name), that the abbreviation marker in the form of horizontal bar placed
the c letter is represented by the corresponding Unicode
character in the allographetic layer, and that the expansion letters are tagged with
<ex> in the diplomatic layer.
The format of the Graal edition has the advantage of
being explicit and easy to process, but it is very verbose and is not
as defined in in the TEI Guidelines [TEIP5: section 23.4].  It should be noted that the
facs (i.e. allographetic) transcription layer is only provided for the first nine
columns of the text, so different alignment rules had to be applied to the rest of
Fontenay Charters Transcriptions
The Fontenay charters were encoded in TEI P5 XML on the basis of a normalized edition, then enhanced with allographetic features, focussing on abbreviations and the distinctive allographs which were already identified as relevant for the research [Stutzmann 2011: 253-5]. Its origins make this transcription uneven: tagged named entities and normalized punctuation and capitalization, but with line breaks, abbreviations, and allographs. A fragment of the input transcription is provided below:
<lb n="14"/> Actum est <app> <lem>hoc</lem> <rdg wit="B" rend="omis"/> </app> <placeName>Eduę</placeName> in plena ſynodo <date> <app> <lem>.XV.</lem> <rdg wit="B">XV°</rdg> </app> kl. julii</date>, laudanti<choice> <expan>bus</expan> <abbr>b;</abbr> </choice> archidiaconi<choice> <expan>bus</expan> <abbr>b;</abbr> </choice> <persName> Gaufrido</persName>,One can note simultaneous use of apparatus elements, name and date tags, abbreviation marks and expansions, and the presence of modernized punctuation (which is not tagged). The use of whitespaces in elements with mixed content is relevant for tokenization.
The scope of this format is to enable text to image alignment at word and character level avoiding as much as possible any manual editing of the input transcriptions. It defines a set of layout elements necessary for positioning a text fragment on the page surface, a number of rules for distinguishing alignable and non-alignable elements of the transcription, and the mechanism for recording the alignment at word and character level.
Layout Markup and Linearity Conflicts
In order to associate a transcription of a page with an image file, a <pb/> must be
provided at the beginning of each transcribed page. It must have an
attribute for stand-off alignment and may have a
facs attribute for direct
linking to the corresponding image file or to the corresponding
zone element in
<cb/> must be used in the beginning of each text column if there are
more than one of them on a page.
<lb/> must be used in the beginning of each text line. It must have an
n attribute indicating the line number. In most cases this attribute may be
generated automatically but sometimes the physical organization of text segments in
does not correspond to the logical order of the text structure. This actually happens
border of some text divisions where the title of the new division (often written with
ink, as a rubric) overlaps with the end of the previous
division (See Figure 1).
Figure 1: Rubric Overlapping with Text Divisions
In this case, the linearity of the transcription
follows the logical order of the text, and two
<lb/> elements with the same
value of the
n attribute are used in the rubric and in the text body, as in the example below:
<div> <p>[...] <lb n="4"/> est desus ia ne sera si fo<lb break="no" n="5"/>rt tour .</p> </div> <div> <head>comm<ex>en</ex>t la t<ex>er</ex>re <lb type="rubrique" rend="align(right)" n="6"/> crolle et fent.</head> <p> <lb n="6"/> <hi rend="lettrine">O</hi>re <lb n="7"/> entendez donq<ex>ue</ex>s [...] </p> </div>Thanks to this simple mechanism, the alignment software can find the physical lines corresponding to the transcription of rubrics and similar cases. The
typeattribute on the first occurrence of hte
<lb/>allows sorting out the line breaks that different from those in the
maintext flow. The optional
rendattribute may be used to indicate the position of the following text in the physical line. In our example, the text
crolle et fent.is placed to the right of
Oreand thus overlaps with the logical order of the transcription.
Another case where the linearity of the transcription does not correspond to the
physical order of lines in the manuscript is that of interlinear or marginal additions.
These must be marked up using the
<add> tag and are considered
non-alignable at the current stage of the project, as the image analysis software
only detects the
regular lines of the text.
A yet more complex situation is created in the case of scribal corrections of word order usually indicated by special transposition markers (See Figure 2).
Figure 2: Transposition
If the transcription follows the
final text order (which is usually the case),
special markup is necessary in order to ensure the possibility of alignment. A combination
<seg> elements is proposed for this purpose in
the Oriflamms project:
<ptr type="transposition-orig" target="#tps106_163d2_1"/> <w type="ADVgen" xml:id="w106_006642">plus</w> <seg type="transposition-target" xml:id="tps106_163d2_1"> <w type="ADVgen" xml:id="w106_006643">apertement</w> </seg>
The following XSL templates can be used to rearrange the text in the
<xsl:template match="tei:ptr[@type='transposition-orig']"> <xsl:apply-templates select="id(substring-after(@target,'#'))"/> </xsl:template> <xsl:template match="tei:seg[@type='transposition-target']"/>
Manuscript pages can contain inscriptions that do not belong to the text being
transcribed. These include catchwords and posterior annotations (such as library shelf
marks). In some cases these elements may be transcribed (using
<add> tags) but in most cases they are not, and the alignment software has
to guess which lines on the manuscript page are
suspect and dropped from
Alignable and Non-Alignable Elements
In addition to the transription stricto sensu, XML
documents may contain a whole number of text containing elements that do not correspond
any graphical object in the source manuscript. These include the metadata recorded
<teiHeader>, editorial notes (
<note>), text supplied by
the editor for missing or illegible passages (
<supplied>). These are of
course non-alignable. On the contrary, the elements
<surplus> contain parts of text which should not appear in the edition
(e.g. repeated or expunctuated words), but should be taken into account by the alignement.
The transcription of these text segments had to be added to the source documents where
Some transcriptions include critical apparatus providing variants from different
manuscripts of the same text tradition (cf. the Fontenay
transcription fragment provided above). These are encoded
elements. In this case, the alignable text content is situated in the
<rdg> where the
wit attribute contains a reference to the
base manuscript of the transcription.
<choice> elements are used, only one of their child elements is
<sic> (and not its
orig (and not its
<expan> sibling). However, if
<expan> are used alone, they are alignable at word level, even though some
characters in the transcription will not match those on the image. The
element used alone may be alignable at word level but as is contains a segment of
somehow corrected by the editor, there is no guarantee that the characters contained
element have any correspondence on the image.
The abbreviations may be particularly hard to align, as some of the markers occupy
position on the horizontal text line (as baseline or spacing superscript characters)
others are placed above other characters (as combining diacritics or superscript letters).
The former are alignable, the latter are not. When the transcription represents the
abbreviation markers directly, the Unicode areas labelled as
combining can be
used to identify the non alignable characters. When the transcription only contains
expansion with the supplied letters marked up using the
<ex> tag, the only
way to identify the alignable expansions (which are a minority) is to provide a list
them. Here is the list that has been used in the project so far:
overturned con the baseline for
et(⁊ or  (MUFI)),
angular tilde for
er(͛ ou  (MUFI)),
double curb tilde for
ur(᷑ ou  (MUFI)),
Word and Character Level Tokenization
<pc> tags are used on every alignable word
and punctuation mark. All of them are equipped with an
xml:id attribute which
is used for alignment. The same technique is applied at character level using the
<c> tag. The tokenization at both levels is performed automatically using
pre-processing XSLT stylesheet librrary.
Processing and Implementation: Oriflamms Alignment Software
The Oriflamms software allows the automated alignment of TEI words with images, the manual and automated validation and correction of the alignment [Stutzmann 2013 and Leydier 2014]. It also offers the possibility to display all image occurrences of a query string and to export comprehensive statistics on the validation and correction actions performed by the user.
To create an alignment project, one needs a TEI file and the set of images that are
referred to in the
<pb> elements. The software then automatically detects the
bounds of text columns in the images using the count of
<cb> elements. Median
lines are then computed in the columns. If there are more lines on the image than
<lb> elements, then the most uncommon lines are ignored in priority (e.g.
catchwords or library shelf marks inscribed on the page). Words are not segmented
spaces between them are at best irregular and often nonexistent. Along each median
signature is computed. It describes the coarse shape of the strokes: dots, curves
or right, vertical lines above, bellow or crossing the median line. The same kind
is extracted from Unicode text lines using a lookup table. The signatures of Unicode
text lines are compared with the Levenshtein distance from which the alignment is
automatically inferred. Because of the kerning, vertical segments are unsuited to
words and characters, so we use curves that follow the lightest path (considering
text is darker than the background).
The graphical user interface displays the alignment on the images with boxes on top of which sits the transcriptions. A mouse click on a words makes it cycle through three states of validation: correct, erroneous and unvalidated.
Figure 3: Oriflamms Software Word Alignment Validation and Correction Interface
The same interface allows validating and correcting text-to-image alignment at character level:
Figure 4: Oriflamms Software Character Alignment Validation and Correction Interface
A second way of validating is also offered via a tabular view. It displays a list of all the words that can be sorted alphabetically, by the number of occurrences or by length.
Figure 5: Oriflamms Software Tabular Word Alignment Validation Interface
When a word is selected, all its image occurrences are displayed in several clusters
sorted by graphical similarity. The user may click (or click and drag) to tag the
wrongly aligned occurrences. The
manual validation of the alignment can be expanded automatically with simple rules
a word lying between two validated words is valid.
On the image view, the user can enable the edit mode that allows to correct the median lines and the word alignment. In edit mode, the words' bounding boxes are rectangular to allow a simple correction of the frontiers. The curvy frontiers are recomputed automatically when a word's alignment is modified. When a frontier is edited, all preceding or following unvalidated or wrongly aligned words are realigned consequently.
Statistics can be exported to a spreadsheet. They describe the alignment correctness of validated word including rate for each letter beginning or ending words. The average correction (in pixels) of the front and back frontiers is also displayed. These statistics allow us to analyse which words, letters or sequences of letters are baldy aligned in order to improve our algorithms.
The research project that inspired the reflection on the problems of text to image
still going on, and the proposed markup solutions presented above are to a certain
work in progress. We hope that sharing this experience at the Balisage conference
may result in the improvement and consolidation of the proposed format and of the
software that will eventually benefit the whole community of scholars working with
and transcriptions of text-bearing objects.
[Bautier ed. 1984] Bautier, Robert-Henri, ed.
Folia Caesaraugustana, vol. 1: Diplomatica et Sigillographica.
Travaux préliminaires de la Commission internationale de diplomatique et de la Commission
internationale de sigillographie pour une normalisation des éditions internationales
éditions de documents et un Vocabulaire internationale de la diplomatique et de la
sigillographie. Zaragoza. Institución
Fernando el Católico.
[Bourgain ed. 2002] Bourgain, Pascale and Vielliard, Françoise (ed.) Conseils pour l'édition des textes médiévaux. Fascicule III, Textes littéraires. Paris. CTHS, École nationale des chartes. 2002.
[Careri et al. 2001] Careri, Maria, Fery-Hue, Françoise, Gasparri, Françoise, Hasenohr, Geneviève, Labory, Gillette, Lefèvre, Sylvie, Leurquin, Anne-Françoise, and Ruby, Christine. Album de manuscrits français du XIIIe siècle. Mise en page et mise en texte. Rome. Viella. 2001.
[Guyotjeannin ed. 2001] Guyotjeannin, Olivier, ed. Conseils pour l’édition des textes médiévaux, Fascicule II, Actes et documents d'archives. Paris. CTHS, École nationale des chartes. 2001.
[Koschwitz 1879] Koschwitz, Eduard. Les plus anciens monuments de la langue française publiés pour les cours universitaires par Eduard Koschwitz. Henninger. Heilbronn. 1879. http://gallica.bnf.fr/ark:/12148/bpt6k3733294
[Leydier 2014] Leydier, Yann, Eglin, Véronique, Bres, Stéphane, and Stutzmann, Dominique. "Learning-Free Text-Image Alignment for Medieval Manuscripts", 14th International Conference on Frontiers in Handwriting Recognition (ICFHR), 2014, p. 363‑368. http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=6981046, doi:https://doi.org/10.1109/ICFHR.2014.67
[Marchello & Lavrentiev ed. 2013] Marchello-Nizia, Christiane and Lavrentiev, Alexei, ed. Queste del saint Graal. Édition numérique interactive. [online]. 2013. http://catalog.bfm-corpus.org/qgraal_cm
[Roques 1926] Roques, Mario.
règles pratiques pour l'édition des anciens textes français et provençaux. Rapport
de la 2e
Romania, vol. 52, 1926.
[Stutzmann 2011] Stutzmann, Dominique. "Paléographie statistique pour décrire, identifier, dater… Normaliser pour coopérer et aller plus loin?", in Codicology and Palaeography in the Digital Age 2, Norderstedt, 2011, p. 247-277. http://kups.ub.uni-koeln.de/4353/
[Stutzmann 2013] Stutzmann, Dominique. "Ontologie des formes et encodage des textes manuscrits médiévaux. Le projet ORIFLAMMS", Document numérique, 16/3, 2013, p. 81-95. doi:https://doi.org/10.3166/DN.16.3.69-79"
[Vielliard ed. 2001] Vielliard, Françoise and Guyotjeannin, Olivier, ed. Conseils pour l'édition des textes médiévaux, Fascicule I, Conseils généraux. Paris. CTHS. École nationale des chartes. 2001.
 The results presented in this paper were obtained in the framework of the Oriflamms Research Project (ANR-12-CORP-0010-02) financed by the French National Research Agency (ANR). The authors are also grateful to the ASLAN project (ANR-10-LABX-0081) of Université de Lyon, for its financial support within the program "Investissements d'Avenir" (ANR-11-IDEX-0007) of the French government operated by the ANR.
 In disciplines like Epigraphy this kind of transcription is usually referred to
interpretative), however, we prefer to use this term
for the intermediate transcription type described in the next section.
 Named after Karl Lachmann (1793-1851), a German classical philologist considered to be the founder of the method.
 The BFM-MSS proper extension elements are not relevant for the scope of this article and will not be described here. Complete ODD documentation of the BFM-MSS encoding schema is available at http://portal.textometrie.org/bfm/files/BFM-MSS-ODD.xml
<pc> element was not available at the time when the encoding
schema of the project was defined.
 The Graal transcriptions qualify as a TEI extension and are close to being TEI conformable.