Among the disciplines that study the past
through its textual heritage, palaeography focuses on the
original script of manuscript books. For a thorough introduction on palaeography see Bischoff
1990.neither script,
nor any other single aspect of a book, is an adequate basis
for [final] judgement
(quoted in Brown 1993, 19), the goal of the
palaeographical method remains the dating and localisation of
manuscripts through the analysis of the physical features of
their writing.
The broader interests of palaeographical research range from the origins and alterations of manuscripts to their owners and uses. The cultural materiality of the manuscript itself (see McGann 2003), however, directs the research. All physical aspects of the book as an individual object with its own history are inseparable from the script and its creation (See Parkes 1991). The techniques used in a manuscript's manufacture, the notes made by its scribes or illuminators, the indications borne by its kalendars and litanies, its provenance and textual tradition, its decoration and illumination—all are valuable guides to palaeographic interpretation.
Palaeography's most fundamental method is
the graphic comparison of one or more manuscripts against a
dated (or datable) and localised corpus. This involves a
formal and stylistic comparison
—as Supino
Martini stresses in her statements about the palaeographical
method—between what has reached us with a close
paternity (or date) and what is supposed to be possibly
brought back to the same paternity (or date)
.confronto formale e stilistico fra
quanto ci è giunto con paternità (o data) vicina e quanto
si presume possa essere ricondotto alla stessa paternità
(o data)
(Supino Martini 1995, 18).
Thus the discriminating palaeographical
eye, trained by experience of observation, the synoptic
examination of manuscripts, and the practice of analogy, is
able to see order in what might otherwise appear
undifferentiated. It is an eye that weaves a logical plot
around what seem to be intuitive likenesses or unlikenesses,
thanks to what has been called a See Ginzburg 1979, 59-106; Petrucci 1991, 7; Mastruzzo
1995, 460.
A palaeographical study begins with the
choice of a sample corpus. Although the corpus is chosen to
provide an adequate sample of handwriting, the manuscripts in
question can be very different in provenance, date, and script.
In identifying and classifying graphically significant
features, therefore, nomenclature becomes crucial. Indeed, in
palaeography, the definition of terminology used to create
significant categories is a high priority. See the landmark first
polyphonyof the real manuscripts (see Sperberg-McQueen 1991) needs to be abstracted and reduced as consistently and systematically as possible, without constructing
arbitrary boxes to squeeze facts into(Gumbert 1976). A consistent system of reference or ontology of information for palaeographic analysis does not exist, however.
For the ontology of information designed for the analysis of old Roman cursive see Terras and Robertson 2004.
Publications on palaeographical subjects
need to be accompanied by illustrations. Indeed, even the
pioneer volumes on palaeography of the seventeenth and
eighteenth centuries bore accurate imitative drawings (see Petrucci
1984), while the treatises of the early nineteenth
century featured hand-engraved facsimiles. At the end of the
nineteenth century, the development of the technique of
photography made available full-page reproductions, For a bibliography on the limits of reproduction as
secondary evidence compared to verbal description see Tanselle
1989 and the brief comments in Petrucci
1991, 14.
Generalisation, however, requires
abstraction. For the sake of generalisation and dissemination,
the original handwriting of the actual manuscripts needs to be
reduced to a prototype for classification and study. The
leading strategy has been the one of expressing the variations
by means of hand-drawn letterforms. The tension between the
individual handwritten sample and its abstract model is thus
resolved by relying on partial visual representations created
The undertakings of digitisation, of
digital editions and the provision of online facsimiles, are
already dealing in part with this fundamental issue of visual
representation and with the general problem of access to
historical material. For a survey and discussion of the future developments of
electronic scholarly editions towards the use of enhanced
image-pages see Robinson 2004. Examples of projects which have envisaged comprehensive
online access to palaeographic materials from different
perspectives are CDFP (Wooley et al. 2002, and Cuneiform Digital
Forensic Project 2004), CEEC (Codices Electronici Ecclesiae
Coloniensis), and MANCASS C11 2004. Unfortunately, the granularity of digital page-images does
not always satisfy the palaeographer's need for close
examination, nor for comparison among related
manuscripts. For the definition of the discipline see two issues of
Although the present research is focused on a specific case study, the methodology explored here claims a more general value. What in fact the results have shown is that the digital models produced by the SPI provide a considerable support to the analytical description of letterforms and to the systematic comparison among scripts which differ in small details. As a consequence, the nature of the traditional palaeographical method is enriched rather than diminished or undervalued.
The methodology explored in the current work could be applied not only to other case studies and corpora, but may in fact enhance already existing projects on image-based digital editions of medieval manuscripts and documents by offering some effective tools for palaeographic interpretation.
To date, quantitative studies in
palaeography have been limited, apart from some pioneering but
isolated works (e.g. Gilissen 1973; CNRS 1974), to
the production and format of books—the so-called
In fact, as is known, a discipline named
See the debate in
Within the current research, the value
of the trained palaeographic eye is not in doubt. The present
project aims to explore the possibility of supporting the human
expert by quantifying the graphical signs and providing tools
that, rather than diminish its interpretative insights, support
or guide them. The foundational idea is that the examination of
specific explicit criteria can help to bring sensible and
meaningful order to our understanding of script. Explicit criteria are intended here not as rigid
conceptual containers, but generative types as opposed to
The The development of the application has been described in
some dissertations and in a scientific paper by Aiolli et
al. 1999.
The present research was started with the aim of experimenting with this application, in order to establish a methodology and to produce results that eventually will lead toward a future project involving the author as humanities computing expert and one of the researchers from the University of Pisa as specialist in machine learning. While the concluding paragraphs of the current paper summarise the strengths of the application and methodology in use, they also point to the need to redesign the entire program to make it possible for other scholars to benefit from its use and to envisage further developments.
At the moment, the SPI has two main functions:
- to show and quantify graphical relationships among manuscripts bearing different styles of handwriting;
- to provide objective measures of similarity between manuscripts of unknown date or provenance and a corpus of stored models of localised and dated manuscripts generated by the system.
The system consists of a set of modules
communicating through a shared database. It can be conceived of
as a set of
User interaction is limited to two main
tasks: inserting bitmap files into the database following a
previous phase of capture of the relevant manuscript leaves,
and extracting relevant letters (called
While exploring the database, the palaeographer may browse among various images of the manuscripts that have been inserted on the system and among the corresponding folios; the extracted characters may be visualised, and the settings for the generation of models and classifications can be changed.
To achieve the present analysis, a
corpus of manuscripts has been chosen and digitised according
to specific criteria. Around forty codices —without taking
into account the fragments—dating from the 10th to the 12th
century, held at the The only existing catalogue, not exhaustive and rather
concise, but quite useful for a first study, is Avitabile et al. 1970. Some codices are listed
also by Cao et al. 1996. The suppression of religious houses in Tuscany in the
Napoleonic period is discussed in Biagianti
1985.
Five codices from this larger corpus
have been selected for experimentation by the current author
for her master's thesis (Ciula 2003, 2004a, 2004b, and
2004c): FIII3, FIII13, FV2, FV8 and FV21. All were
held at the monastery of S. Eugenio just outside Siena in the
fifteenth century, as the notes of possession in them
attest. A historical introduction to the monastery can be found
in Kurze
2002.
This sample includes, above all, liturgical texts. The main text is usually written by more than one scribe, in either one or two columns, and it is often framed by contemporary or later marginal and interlinear glosses.
The handwriting in this sample can be
placed on a continuum, from the not-yet-formalised Caroline
script of the ninth century through intermediate trends to
the Gothic hands of the late twelfth century. The palaeographical literature confines
According to the classic nomenclature (see Bischoff et al. 1954, Bischoff
1990, Cencetti 1997 and Derolez
2003),
On the basis of earlier studies concerning these codices, it is possible to identify three groups of scripts:
- Non-formalised Caroline handwriting (9th-10th centuries)
- Intermediate groups
- 1st group: 11th century;
- 2nd group: first half of the 12th century;
- 3rd group: second half of the 12th century;
- Pre-Gothic (early or incipient Gothic script; late 12th-early 13th century).
Out of this approximate classification, there is a group of codices that have been generally dated to the twelfth century, again according to earlier studies.
The five codices from the monastery of S. Eugenio studied here belong to three of the five groups: FV8 (end of ninth/beginning of tenth century) and the first part of the codex FV2 (tenth century) belong to group 1; FIII3 (anno 1017) and FIII13 and the second part of FV2 (both eleventh century) belong to group 2a; FV21 (end of the twelfth century) belongs to group 3 (all dates as in Avitabile et al. 1970).
The following paragraphs will show in more detail how SPI may work during a palaeographical enquiry and can support:
- the morphological analysis of the model related to a single letter or group of letters;
- the comparative analysis of a predefined set of several models;
- the classification of new manuscripts by means of a set of classifiers that work on single letters or groups of connected letters.
For SPI to work and produce models of letterforms, it is necessary first to separate the letters out, that is to say to segment the continuous script of the manuscript into individual occurrences of letterforms. Some premises need to be stated to justify such an approach, which treats the single letter form as the basis for morphological analysis.
In medieval palaeography it is widely
recognised that many variations in handwriting are not
random, but consciously produced. Indeed, the medieval
scribe had in mind a script to pursue, a
constructed
letterforms, made in several strokes
.
The segmentation process allows us
then to focus on distinctive letterforms, which will vary
minimally for the same style of handwriting. Nevertheless,
despite the experienced scribe's taught skills,
handwriting, like any human activity, varies and is never
the same. No letter can be re-written by hand in exactly
the same way. This intrinsic variability complicates all
automatic image processing of any human handwriting. It is
a concrete example of what is called the problem of
inconsistency in automatic pattern recognition (for
an introduction to document image analysis see Bunke and
Wang 1997). From a palaeographical point of view,
the issue is to distinguish between the differences due to
natural human inconsistency and those due to changes of
style.
For this study I have chosen not to segment the manuscripts in their entirety, but only the single letters or groups of connected letters that are relevant. The SPI segmentation module, which works quite successfully with non-cursive medieval hands, goes through these steps:
- The palaeographer chooses a letter to segment within the digital page from the alphabetic list on the toolbar, and this starts the training of the system.
- The area of the image that contains the letter, that
is to say the frame that delimits the effective extension
of the character, its
blob of ink , is then selected, either manually or automatically.The identification of the box may be made by automatic selection, in which the system searches for the optimal local frame after the mouse is positioned in the approximate middle of the letter, or by manually drawing a box around the letter of interest and adjusting the size of the frame with the mouse. The automatic selection is achieved by a technique independent of the morphology of the character, based exclusively on the local approximation of the letter's dimensions using criteria related to the type of handwriting, such as the thickness of the stroke, the module, and the modular rapport between the height and the width of the letter. Even if the cropped frame contains part of other letters, the following phases are not compromised. It is instead of crucial importance that the frame encloses the whole character in question, including linkages with any connected letters.
- The segmentation algorithm suggests a segmentation
for the chosen letter or groups of letters.
In the last version of SPI the palaeographer can access some variations of the minimal segmentation process. These variations have resulted by adding vertical segments to the main automatic segmentation. Therefore, at the beginning, just a minimal segmentation is shown. Starting from there, there is the possibility of deriving alternative segmentations that will be proposed to the palaeographer together with the minimal solution, and the solution suggested by the system. Finally, it is up to the expert to opt for the best segmentation.
The main parameters for the
segmentation relate to the typology of pattern to segment,
that is to say on what is called projection The projection of a character on a straight line of
gradient For those letters or groups of connected letters whose
morphology is highly dependent on the singular hand,
and which are expected to have a consequent variable
histogram—as it is often the case of the letter 〈g〉, or as it occurs when there
is an alternation of vertical 〈d〉 and 〈d〉 with
sloping ascender—it is necessary to change modality.
However the actual interface does not provide an option
for this change.
Before the second phase of model generation, it is necessary to repeat the segmentation process by selecting a certain number of characters representing the same letter.
In the training phase, a
Hence, on the basis of digital images of medieval book hands, the system generates some graphical models related to single letters or to groups of connected letters.
Once the models are created, they may be used for either interpreting models that already have been generated and saved so far (that is to say as objects for the palaeographer to observe) in order to analyse their morphological characteristics; or for creating clusters of reference for the system itself to refer to when the classification of manuscripts takes place.
The models and, in particular, the
figures by which they are represented—the centroyds—may be
subjected to a series of transformations. This procedure of
graphical transformation is called
More specifically, what the system is
able to visualise are projections on the subspace of the
model. The common characteristics, that is to say the
tangents of the class, remain invariant, while the model is
allowed to show its possible changes within its tangents'
subspace. In mathematical terms, the variety obtained as the
transformations are applied to the pattern is
approximated within a linear space or tangent space,
which has the same number of dimensions as the number
of defined transformations.
A window of the application allows this graphical visualisation of the tangents and shows how the character's prototype changes while its position varies within a determined subset of tangents.
It is always possible to display and to
browse the models that have been generated and stored.
However, the
The
A preliminary phase is necessary during which the software computes the matrix of distances among the models taken in pairs. A measure of distance, which represents the degree of similarity, is then associated with every model compared to the other ones and it appears at the top of each model frame. The same calculation of distance is used by the next tool, the dendogram.
The dendogram is based on a clustering algorithm that splits the set of models in subsets, and orders them in a hierarchical structure of clusters. Every cluster represents a subset of models sharing a certain morphological similarity. The result of such a computation is visualised as a binary tree that is actually the dendogram itself.
Besides being useful for palaeographical
analysis and for graphical comparison, the models are the
references for the last phase, that is the effective
classification phase, as briefly anticipated above.
(
During the operative phase of classification, the palaeographer extracts some samples of letters from the subject manuscript, this time not for training the system but for testing its classificatory skills and interpreting the results. At this point, the system
- applies transformations to the samples, establishing
the possible variants of the
pure character or centroyd; - compares characters extracted from the subject sample
with stored prototypes from the comparison sample on which
the training phase had been performed
The stored models have to be activated and they have to be compatible with the examined characters from the subject sample: same type of letter, same image format, and same number of tangents.
- retrieves manuscript models which are morphologically similar to the subject characters.
This kind of classification is known as
In applying SPI to the corpus of manuscripts under study, the first stage to define and perform is obviously the digitisation process: for every codex belonging to the subject corpus, four pages were scanned for each style of script. For manuscripts showing a homogeneous style, this meant four pages per manuscript; in more heterogeneous manuscripts, four pages were scanned per style. Pages were scanned at 300 dpi and an archival collection was built on CD-ROM in TIFF format for preservation purposes. Image files used in the SPI were converted into bitmaps with two levels of bit depth, as required by the program. The digital leaves were cropped and introduced into the database in sections corresponding wherever possible to manuscript columns so as to facilitate image management and further analysis.
After the setting up of the digitisation process, it has been possible to perform the segmentation for every page, starting with the letter 〈a〉 (the evolution of which is rather emblematic in the development of the Caroline handwriting) and continuing with 〈d〉, 〈f〉, 〈g〉, 〈m〉, 〈n〉, 〈p〉 and 〈u〉. It would have been ideal to perform the segmentation and subsequent phases for all characters in the alphabet, ligatures, and, eventually, punctuation. The available version of the software does not allow this at the moment, however.
Once the relevant letters had been segmented and automatic models generated, qualitative interpretations of the quantitative representations were elaborated, borrowing terms and methods from the palaeographical literature and intertwining them with the new methodology based on the digital models generated by SPI. The following sections shows these results for the letter 〈a〉 in FIII3.
The first basic stroke of the letter 〈a〉 is the shaft or back. In this hand, it slopes to the left. It may have also an upcurving endstroke at the bottom that often links with the following letter. The top, when not linking with the previous letter —e.g. in 〈ea〉 or 〈ta〉 —is shortened, making what is nearly a single-compartment form. The bow is quite triangular and flattened. Its shape is thin at the top and bottom and, by light contrast, thick in the middle of the stroke.
An analysis of the first tangent
shows that the endstroke of the shaft may be either
reduced or extended upwards, so as to link with the
following letter on the left side. This analysis also
demonstrates the existence of an inconsistency either in
the slant of the back (which may be more or less
straight), or in the module related to the height of the
〈a〉 (which may be taller or
shorter):
On the other hand, the second
tangent, which features a fixed slight leftwards-slant,
shows a variation related to the width of the letter,
either more or less flattened and, consequently, more or
less tending towards a single-compartment shape:
Both tangent variations highlight the corresponding differences in the shape of the counter bow, narrowed or wider depending on the changes in the whole letterform.
After the analysis of letterforms,
followed by the generation of the corresponding models and
the observation of their variants, a comparison among the
various manuscripts was carried out with the support of the
diagram and dendogram tools. For interpreting the measures,
the comparison of different structures of the
diagram—structures which can be obtained by changing the
model at the start of the linear measurement—has been
useful. In the application, diagrams and dendograms feature
tag titles, which show the association between the
model and the corresponding manuscript or graphic
unit.
Variations in the morphology of the letter 〈d〉—limited to the straight 〈d〉 and excluding the features of the Uncial 〈d〉, which is present in all the graphic units except for FV1 (1st part)—are affected especially by the shape of the lobe, which may be rather small and flat, or large and round. In addition, the letterform can be accentuated by a contrastive execution as happens for the letterform 〈b〉. The serif of the ascender may feature a club rather than a wedge shape, or a line serif rather than a beak. These variations often occur within the same manuscript and may confirm that the script was not yet canonised.
This diagram is quite exemplary, since the first order shown seems to offer a gradual variation in the slant of the shaft, from the rightwards-trend in the first model of FV2 (1st part), to the slight leftward-position in FV21 (88.05). Following the same direction, the lobe seems to grow in module, at the expense of the shorter shaft, and to acquire contrast. It is interesting to note that the same phenomenon—with the slight variation of the swamped position of FV2 (2nd part) and FIII3—related to the slope, to the module of the lobe and to the shading, has been noted also for the analysis of the letterform 〈b〉 (see figure 6 above).
As is again true of the letter 〈b〉, FV2 (1st part and 2nd part), FV8 and FIII3 are clustered together on the left of the dendogram, while FIII13 and FV21 are confined to the right side of the graph. According to the measures shown by the diagram, however, FIII3 is this time closer to the second group containing FIII13 and FV21, due to the thickness of its execution and to the wideness of the lobe; while FV2 (1st part) is in its own branch on the left, because of the accentuated inclination of its shaft:
The following figure shows a comparable analysis for 〈b〉:
The classic subjects of palaeographical
analysis include the graphic elements which might contribute to
the definition of a script's
While concentration on the minute
features of letterforms and their models is enhanced by the
use of this computational tool, the overall appearance of the
script and its immediate context is not taken into account by
the SPI. From the segmentation process, where the cropping of
characters is still accomplished within the digital
image-page or within a section of it, all processing by the
SPI focuses on the individual letters, ignoring the
fundamental setting of the script, from the alignment of
letters on up. This new methodology is best used therefore,
when it can be supported by the traditional discipline of
Besides those methodological considerations, the SPI also is affected by several limitations. In order to assure a greater longevity for the application (see O'Donnell 2004), these improvements could be carried out:
- the software runs only on Windows 98 and it should be possible to make it work on other platforms;
- the graphical interface has been conceived by computer
scientists and needs to be redesigned according to the main
principles of accessibility and usability
This is one of the inconveniences of testing an application after its implementation rather than during it. Indeed, the accessibility and usability design process ought to be intimately interconnected with the building of the application itself.
- the segmentation works well, but would be improved by
the provision of additional filters for the processing of
damaged or otherwise difficult-to-process manuscript
images:
Damaged manuscript pages and distortions occurred during capture. Difficulties in the script (biting, ruling, interlinear glosses, conflicting lines). - the alphabetical grill should be widened and, eventually, made flexible so as to accept the addition of specific letters and ligatures;
- the fields that can be filled by descriptive notes should be marked up, so as to make possible better terminological accuracy and a structured search function;
- it should be possible to automatically compare diagrams and dendograms and to display more than one at the same time.
This experiment with the SPI has been successful in that it has shown how the traditional qualitative palaeographic paradigm can be strengthened and assisted by the creation of graphic models that are quantitative in nature but able to be arranged visibly on the screen. This allows the program to show the infinite variations of handwriting and interpretations connected to them, opening the analysis to criticism and reasoning. It has demonstrated, in short, how quantitative models can be created that call for qualitative exegesis.
The palaeographical comparative method is enriched and changed. Yet, from the beginning, terminological coherence is needed to carefully describe the models and to compare them. Moreover, the comparison of many models allows us to see the data from new perspectives, by structuring the corpus using multiple criteria. Finally, the interpretations are made concrete and precise by both digital representations and quantitative measures.
The testing of SPI on the specific corpus of manuscripts held in Siena has produced detailed descriptions of the letterforms and of the handwritings under examination, descriptions that may lead to further interpretation of the unknown origin of the codices and of the development of Caroline minuscule. However, the methodology, its advantages and limits, can be generalised and the combination of classical palaeographical analysis and digital models applied to new case studies for the analysis and identification of other book hands or non-cursive documentary hands. An ideal database could actually include samples and models of various Western European handwritings, with punctuation and other signs, so as to facilitate the work of scholars looking for similar origins and provenances, for clues to copying activity and to the circulation of culture.