An Interpretable Deep Learning Approach for Morphological Script Type Analysis

Abstract

Defining script types and establishing classification criteria for medieval handwriting is a central aspect of palaeographical analysis. However, existing typologies often encounter methodological challenges, such as descriptive limitations and subjective criteria. We propose an interpretable deep learning-based approach to morphological script type analysis, which enables systematic and objective analysis and contributes to bridging the gap between qualitative observations and quantitative measurements. More precisely, we adapt a deep instance segmentation method to learn comparable character prototypes, representative of letter morphology, and provide qualitative and quantitative tools for their comparison and analysis. We demonstrate our approach by applying it to the Textualis Formata script type and its two subtypes formalized by A. Derolez: Northern and Southern Textualis

Full text

An Interpretable Deep Learning Approach
for Morphological Script Type Analysis
Malamatenia Vlachou-Efstathiou1,2[0000−0002−9397−356], Ioannis
Siglidis2[0009−0002−2278−5825] , Dominique Stutzmann1[0000−0003−3705−5825] , and
Mathieu Aubry2[0000−0002−3804−0193]
1Institut de Recherche et d’Histoire des Textes, Paris, Île-de-France, France
{malamatenia.vlachou, dominique.stutzmann}@irht.cnrs.fr
2LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, Marne-la-Vallée, France
{mathieu.aubry, ioannis.siglidis}@enpc.fr
https://learnable-handwriter.github.io/
Abstract.
Defining script types and establishing classification criteria
for medieval handwriting is a central aspect of palaeographical analysis.
However, existing typologies often encounter methodological challenges,
such as descriptive limitations and subjective criteria. We propose an
interpretable deep learning-based approach to morphological script type
analysis, which enables systematic and objective analysis and contributes
to bridging the gap between qualitative observations and quantitative
measurements. More precisely, we adapt a deep instance segmentation
method to learn comparable character prototypes, representative of letter
morphology, and provide qualitative and quantitative tools for their
comparison and analysis. We demonstrate our approach by applying it
to the Textualis Formata script type and its two subtypes formalized by
A. Derolez: Northern and Southern Textualis.
Keywords: Latin Palaeography ·Computer Vision ·Palaeographical
Analysis ·Character Prototypes ·Textualis Formata
1 Introduction
The concept of script type is of central importance to palaeography, which
studies handwritten documents in relation to their context of production such
as date, origin, and scribal hands, to support historical discourse. Adapting
M. Parkes, we define a script type as “the model which the scribe has in their
mind’s eye when they write” [
43
,
56
], that is, if we restrict ourselves to characters,
the set of prototypical forms of each character towards which they are working
when they write. In order to discretize the continuum of handwritten forms
and establish script types and their classification criteria, the palaeographical
method compares handwriting samples, describes, and analyzes the variations
of letter forms. Palaeographers often resort to the idea of script types as ideal
prototypes [
6
], with the delineation of artificial alphabets, i.e., sets of abstracted
letter forms. Several typologies have been proposed and refined over the years.
arXiv:2408.11150v1 [cs.CV] 20 Aug 2024

2 M. Vlachou-Efstathiou et al.
Most recently A. Derolez proposed a taxonomy for Gothic book scripts based on
letter morphology [
16
], where some letters with distinctive visual elements serve
as the basis for classification.
In this paper, we introduce a methodology that leverages deep learning for
the analysis of morphological script types. More precisely, we learn aligned char-
acter prototypes from documents and present different methods for qualitative
and quantitative analysis. This enables us to confront different documents to
existing typologies, potentially adding nuance or complementing them. Indeed,
existing typologies present persistent methodological issues [
16
,
53
,
58
], mainly
the ambiguity arising from relying on a “global impression” to discern scripts,
inconsistencies in nomenclature across scholarly traditions, and difficulties in
describing minute morphological differences using natural language [
19
,
56
]. These
challenges underscore the potential benefits of methods such as ours, which
could enhance palaeographical analysis through a systematic and objective ap-
proach, and facilitate the integration of quantitative measures with qualitative
observations.
This is in line with the position of pioneers like Léon Gilissen [
20
,
21
] and
others [
46
,
42
,
52
,
66
,
15
,
37
,
59
,
39
], who experimented with statistical measurements
and the modeling of measurable elements of script. Such measurements of scripts
pose significant challenges, such as defining a set of descriptors or discriminative
handwriting features and ensuring comparable objects and magnitudes [
55
].
Contrary to classification tasks such as writer and geographical attribution,
which are formulated as discriminative learning problems, script type analysis
cannot be reduced to a classification problem [
60
,
26
,
61
] and adequate modeling of
variations is crucial [
57
]. Indeed simply matching external samples to pre-defined
script types does not help better understanding and questioning the classification
criteria.
We thus propose a method for evidence-based paleography focusing on inter-
pretability rather than script classification. Our key idea is to remain close to
classical morphological approaches for defining taxonomies and introduce tools
to model and analyze letter shapes automatically. We build on the Learnable
Typewriter approach [
51
] and adapt it so that it can learn comparable character
prototypes, which requires designing appropriate finetuning strategy and filtering.
We then introduce visualizations and graphical tools, as well as an interpretable
variability measure. To demonstrate how such tools can be leveraged for palaeo-
graphic analysis, we select a corpus in Textualis Formata and present a case
study on the morphological analysis of its two subtypes, Northern and Southern
Textualis.
Contribution. In summary, our main contributions are:
–the adaptation of a deep instance segmentation method for palaeographical
script type analysis
–
a methodology for homologous comparison of characters, including visualiza-
tion, graphical, and quantitative tools

An Interpretable Deep Learning Approach for Script Type Analysis 3
Fig. 1: The Learnable Typewriter Model learns to reconstruct text lines
using a set of learned character prototypes. We demonstrate how the character
prototypes can be used for palaeographic analysis.
–
a case study demonstrating how these tools can complement the classic
taxonomy of A. Derolez [
16
] for the analysis of Northern and Southern
Textualis.
2 Related work
We first give an overview of works that develop quantitative methods for palaeo-
graphic analysis. We then present “prototype-based” approaches to document
analysis, which, although not specifically developed for paleographic analysis, are
the basis of our approach.
Quantitative methods in palaeography. In the past two decades, many automatic
methods have been developed for writer or script classification [
54
,
40
], using
texture-based features [
38
,
49
,
67
,
34
,
27
,
17
,
25
], grapheme-based features [
48
,
50
,
18
,
29
]
and deep learning classification approaches [
9
,
7
,
10
,
11
,
65
,
64
,
28
]. Some papers, such
as [
31
], make a particular effort to build interpretable features or to visualize
deep features responsible for the classification, but their interpretation remains
limited.
Another branch of studies aims at producing interpretable outputs for palaeo-
graphic analysis of letter structure. The Information System for Graphological
Identification [
36
], extracts the average shape of specific characters via curve
and contour detection, standardizing orientation and size, for automatic hand
comparison and writer identification. The Graphem project [
39
] focused specifi-
cally on script type features. [
14
] explores visually interpretable stroke analysis,
by extracting connected components to create a strokes code book, and then
grouping the strokes through graph coloring for categorization of elementary
stroke shapes. Closest to us, focusing on entire letter form variations for script
type analysis is the System of Palaeographical Inspection [
1
,
8
]. It generates an
average character prototype by computing the centroid of semi-automatically
segmented occurrences. The prototypes are used both for hierarchical clustering
of similar hands and classification of external samples.

4 M. Vlachou-Efstathiou et al.
However, the results of these approaches can hardly be compared with minute
traditional palaeographic analysis and do not provide complete automation.
Instead, our idea is to build on methods that directly learn prototypical characters
from documents and use them for actual paleographic analysis.
Prototype-based approaches in document analysis. Early methods for document
analysis [
30
,
32
,
68
,
3
,
4
] use variants of character template matching for analyzing
documents. While their main goal is often to perform optical character recognition
(OCR), such methods typically also produce finer outputs, such as character
segmentation, and learn a character template or prototype. Similar approaches
have thus been used for typographical analysis of early prints [
47
,
24
,
33
]. This
type of approach has recently been revisited with deep learning tools by the
Learnable Typewriter approach [
51
]. We build on this method and describe it in
the next section.
3 Approach
3.1 Learning comparable character prototypes
The Learnable Typewriter. We build our approach on the Learnable Typewriter
model [
51
], visualized in Figure 1. This deep learning model learns to reconstruct
text lines by compositing a set of character prototypes on a simple background.
Given as input the image of a line, it predicts the color of the background,
the characters used in the line, and for each character, its position and color.
The character prototypes are also learned by the model, and each instance of
a character is reconstructed with the exact same prototype. The model can
be trained, as in our experiments, using a set of text line images with their
transcriptions.
Each character prototype is a grayscale image and can be thought of as the
average shape of all occurrences of a character in the training data, standardized
for color, size, and position. Therefore, training a Learnable Typewriter model on
a particular corpus, such as one corresponding to a specific script or handwriting
style, will yield the average shape of each character without the need for manual
selection of specific character samples, annotation of character positions, or
binarization.
Finetuning character prototypes. We propose to compare different documents
and different scripts by comparing the character prototypes learned on various
corpora. However, directly comparing prototypes learned by different models is
not possible, because they are not aligned. Our solution is first to learn a reference
model using a reference corpus - in our case study, a set of documents in Textualis
Formata - then finetune the model to reconstruct selected documents, keeping all
network parameters frozen except those that only impact the prototypes. Since
the positioning, scaling, and coloring of the prototypes are shared, the prototypes
will remain aligned and can be directly compared, such as by computing their
difference. We define a single reference corpus to obtain reference prototypes,

An Interpretable Deep Learning Approach for Script Type Analysis 5
(a) Filtering strategy (b) Failure case identification
Fig. 2: Prototype filtering and failure case identification. We use a mask
M defined from the reference prototype
R
to remove artifacts from finetuned
prototypes
P
, yieling a filtered prototype
F
. We compute an error
e
associated
to the filtering to automatically identify potential failure cases.
and then finetune them on multiple specific corpora, that may or may not be
part of the reference corpus.
Character prototype filtering. While the reference prototypes generally are of high
quality, we observed various artifacts in the finetuned prototypes, particularly for
less common characters, when trained on a single document. While this does not
hinder the qualitative analysis of the prototypes’ shapes, it does complicate the
quantitative comparison between prototypes. To alleviate this issue, we propose
to filter the finetuned prototypes using the reference ones, as visualized in Figure 2.
Let us consider a specific character, the associated reference prototype
R
and
the associated finetuned prototype
P
for a given finetuning. Using the reference
prototype, we define a reference mask Mas
M=G∗D(R> t),(1)
where
G
is a Gaussian filter,
∗
denotes a convolution,
D
is a dilation operation
and
R> t
is the binary mask associated to pixels for which
R
is greater than a
threshold
t
. In our experiments, we use a Gaussian
G
of standard deviation 2, a
dilation
D
of 2 pixels, and a threshold
t
=0.8. Intuitively, this mask defines in a
soft way, for each character, pixels that are close to the reference prototype.
Using this mask, we define a filtered prototype
F
=
M·P
, where
·
is the
pixel-wise multiplication, which we use for all of our analyses.
Automatic identification of failure cases. While the filtering process described
above generally improves the visual quality of the prototypes without changing
the appearance of the characters themselves, there are instances where either the

6 M. Vlachou-Efstathiou et al.
(a) Graph interpretation (b) Character graph (c) Document graph
Fig. 3: Comparison graphs. The markers correspond to different document
prototypes and their coordinates to their distance to the Northern and Southern
Textualis prototypes. See text for details.
appearance is slightly altered or the finetuned prototype is of very low quality.
We want to identify such cases automatically, to avoid misinterpretations. To do
so, for a given character associated with a reference mask
M
and a finetuned
prototype Pwe compute the error edefined by:
e=∥(1 −M)·(P> t′)∥,(2)
where
∥ ∥
is the norm of an image,
·
is the pixel-wise multiplication, and
t′
is
a scalar threshold, set to 0.65 in our experiments. Intuitively, this error can be
interpreted as the number of pixels that are present in the finetuned prototype
P
(i.e., have values higher than
t′
) but are filtered out by the mask
M
. This value
e
enables us to identify: (i) finetuned prototypes whose shape is significantly
different from the reference one and are thus modified by the filtering process,
such as the ‹d› in Figure 2, and (ii) finetuned prototypes of low quality that
might not be easily interpretable, such as the ‹e› in Figure 2. In our results, we
highlight prototypes where
e >
15 in orange and prototypes where
e >
30 in red.
3.2 Character prototype comparison for palaeographic analysis
Visual comparison. We can visually highlight the morphological differences
between two prototypes by subtracting one from the other. To make this difference
easier to understand, we use a colormap that represents zeros as white, and positive
and negative values as two distinct colors, typically red and blue. This method
reveals pixel-wise differences, facilitating an initial qualitative examination of the
morphological disparities (see Table 2 and Figure 7).
Character and document comparison graphs. To quantitatively analyze character
prototypes, we introduce an adapted comparison graph, illustrated in Figure 3. In
this graph, each point represents a specific document character prototype, with its
coordinates defined as its distance in pixel space to two selected prototypes. Since
we study Textualis Formata, we use the prototypes of Northern and Southern

An Interpretable Deep Learning Approach for Script Type Analysis 7
Textualis (NT and ST) (see Section 4.1 for details), other prototypes could be
selected for different analysis. The distance to the axes can be interpreted as
visualized in Figure 3a.
We employ two complementary types of graphs for our analysis. Firstly,
character graphs (see Figure 3b) concentrate on a single character across all
documents. In these graphs, dots represent the documents selected to train the
reference, Northern and Southern Textualis models, while crosses denote the
remaining documents. Blue, resp. red, markers signify Northern, resp. Southern
Textualis documents. The identifier for each document is written near its marker.
This type of graph allows to easily identify outlier documents for a specific
character, such as NT5 for the ‹a› character, which is closer to the ST prototype.
Secondly, document graphs (see Figure 3c) focus on a specific document, where
each dot corresponds to a different character, labeled near the marker. The color of
the dots corresponds to the frequency of occurrence, with darker dots representing
less frequent characters. This type of graph facilitates the identification, for a
given document, of the characters most typical of a subtype, such as the ‹a› for
NT3.
Quantifying character variability. According to the literature, the Northern
Textualis class allows for more morphological variation across documents than
the Southern Textualis. While the visualizations and graphs can qualitatively
support this idea, we further aim to quantify the characters’ variability. Thus,
we report the standard deviations of character prototypes within one subtype,
σNT
for Northern Textualis and
σST
for Southern Textualis. These standard
deviations can be thought of as the average number of pixels that change across
two character prototypes of the same subtype.
4 Experiments
Research question and analysis framework. To analyze the results of our ap-
proach, we adopt the taxonomy formalized by A. Derolez [
16
], which provides
a framework based on morphological criteria. We select a corpus in Textualis
script type, specifically in its canonized calligraphic form Formata, due to its
more distinguishable morphological elements compared to more rapidly executed
forms. Despite the morphology-based categorization where Textualis Formata
represents a coherent group, Derolez makes “an important distinction between
two fundamentally different species”, Northern and Southern Textualis, based
on their geographical distribution and a set of minute morphological differences.
However, these differences vary according to factors such as date, geographical
origin, or language, and often intersect, blurring this fundamental distinction in
some cases. Our goal is to confront the results of our approach with Derolez’s
criteria and observations.

8 M. Vlachou-Efstathiou et al.
Table 1: Dataset Description. The ‘Doc.’ column refers to the names we use
for the different documents in this paper, NT and ST stand for Northern and
Southern Textualis, and the ‘Ref.’ column reports which documents are used to
train the reference and subtype models.
Doc. Ref. Shelfmark Language Century Date Origin Folio(s) Lines
NT1 Paris, BnF, Français 403 [62] French 13th 1226-1250 England 4r 76
NT2 Paris, BnF, Français 12400 [62] French 14th 1305-1310 Eastern France 92r 54
NT3 Arras, BM Ms. 861 (315) [12] Latin 14th - - 56r 65
NT4 Paris, BnF, Français 1728 [44] French 14th 1372 - 3r 59
NT5 Paris, BnF, Français 20120 [62] French 13th 1240-1250 Paris or Orleans 7r 81
NT6 Paris, BnF, Français 619 [35] French 14th 1375-1400 - 1v 65
NT7 Berlin, SB, Hdschr. 25 [12] Latin 15th 1451-1500 Flanders 22r-22v 25
ST1 Paris, BnF, Français 9082 [62] French 13th 1295 Rome 171r 56
ST2 Paris, BnF, Espagnol 65 [5] Navarrese 14th 1301-1310 - 5v,6r 118
ST3 Paris, BnF, Italien 590 [2] Italian 14th 1370-1410 Italy 18r 68
ST4 Madrid, FLG, mss 289 (Hand A) [22] Castilian 15th 1480 Seville 245v 86
ST5 Paris, BnF, Français 187 [62] French 14th 1350-1386 Milan or Genova 18r 16
ST6 Paris, BnF, Latin 7720 [23] Latin 15th 1390-1410 Florence 102v 74
ST7 Madrid, FLG, mss 289 (Hand B) [22] Castilian 15th 1480 Seville 274v 52
Total: 892
4.1 Dataset and Experiment Details
Data Selection. Our data was build from two open-access repositories, EC-
MEN [
63
,
62
] and CATMuS [
45
]. We selected the documents to ensure the variabil-
ity of the corpus in terms of geographic, linguistic, and chronological distribution,
resulting in seven documents for each subscript as listed in Table 1 (labelled NT
1-7 and ST 1-7). We verified and normalized the transcriptions to fit a graphemic
approach [
13
,
12
]. Despite our efforts to use diverse and representative documents,
we acknowledge that biases may exist within our dataset.
Character set choice. From the extended set of characters in medieval manuscripts
- including upper and lower case letters, ligatures, punctuation, and abbreviation
signs - we follow a standard approach for morphological analysis and focus on
the lowercase alphabetic characters where morphology is crystalized through
frequent usage. From these characters, we show results on the ones common to
all documents, thus excluding ‹j,k,v,x,y,z›.
Trained models. For our analysis, we trained multiple models to obtain character
prototypes at different levels of granularity: (i) a script type model for Textualis,
(ii) script subtype models for Northern and Southern Textualis, and (iii) document
level models for each document in our dataset. We use the Textualis script type
model as reference model, and finetune all other models from it as explained in
Section 3.1. to validate that our reference and subtype models can be effective to
analyze documents they were not trained on, we limited their training to NT 1-4
and ST 1-4, indicated in the ‘Ref’ column in Table 1.

An Interpretable Deep Learning Approach for Script Type Analysis 9
Fig. 4: Our filtered prototypes on type, sub-type and document level. The
highlighted prototypes are the ones for which filtering had a significant impact
(see Section 3.1 and 4.1 for details).
4.2 General results
Prototype quality. Figure 4 shows the character prototypes generated by our
approach across type, sub-type, and document levels. We note several limits in
these prototypes. Firstly, the prototype for ‹m› is not well modeled ( ). This issue
stems from the network being trained with a CTC loss allowing to use the same
prototype twice to model the same letter. As a result, we exclude ‹m› from our
analysis. Secondly, the two allographs ‹ / s› are represented by a single prototype
resulting in an averaged representation of the two ( ), necessitating cautious
examination. Thirdly, we meticulously scrutinized all prototypes highlighted in
orange and red, where filtering significantly impacted the outcomes. In almost all
instances, the filtering was meaningful and eliminated irrelevant artifacts from
the prototypes. Nevertheless, it is worth mentioning that: (i) the lengthy shaft
of the ST3 ‹d›( ) and the hairline extension of the limb of ‹h› in NT2 ( ) are
slightly severed; (ii) in general, documents that are not part of the reference set,
are less accurately modeled and more impacted by the filtering, especially the

10 M. Vlachou-Efstathiou et al.
Table 2: Comparison between Derolez’ criteria for Northern and Southern Textu-
alis and our subtype prototypes.
‹Ch.›
Derolez’criteria NT| ST |diff.
σNT σS T
‹a› NT: Closed form with variations like “box-‹a›”
ST: Open form or slightly closed with hairline 4.0 3.4
‹b› NT: Sloped or forked ascender tops
ST: (i) Flat ascender tops, (ii) round lobe 4.1 3.6
‹c› NT: Angular or broken lobe curves
ST: Semi-circular lobe 2.9 2.4
‹d›
NT: (i) Lengthened and (ii) concave shaft ST: (i) Shorter
shaft and (ii) almost horizontal, (iii) round bowl 3.8 3.1
‹e›
NT: (i) Diagonal direction of the hairline and (ii) angular
or broken lobe curves
ST: (i) Horizontal or no hairline, (ii) semi-circular lobe form
3.3 3.2
‹f› NT: Incurvation of the shaft foot to the right
ST: Flat foot 4.3 4.3
‹g› NT: Tendency for the closed, “8-shaped” form
ST: Tendency for the open, “Rücken -g” form
Note: Various intermediate forms and difficult to classify
5.8 5.4
‹h›
NT: (i) Incurvation of the shaft foot to the right, (ii) ex-
tended or dislocated limb and (iii) sloped or forked tops ST:
(i) Flat ascender foot, (ii) circular limb on the baseline and
(iii) flat ascender tops
4.9 4.2
‹i›
NT: (i) Accentuated (diamond-shaped or forked) headline
and (ii) extended hairline for the foot ST: (i) Approach
stroke for the headline and (ii) flat end for the foot
2.4 1.6
‹l› NT: Sloped or forked ascender tops
ST: Flat tops 2.5 1.8
‹n›
NT: Accentuated (diamond-shaped or hairlines) for the
headline and (ii) same for feet ST: (i) Approach stroke
hairline for the headline and (ii) flat feet
3.7 3.0
‹o› NT: Broken / more vertically elongated curves
ST: Circular arc forms 3.5 2.7
‹p›
NT: (i) Artificial spurs on the left and (i) decorated descen-
der feet ST: No spurs and (ii) flat descender feet 4.2 3.2
‹q› NT: (i) Lengthy and (ii) decorated descenders
ST: (i) Short and (ii) flat descenders 4.8 4.0
‹r›
NT: (i) Hairline endstroke for shaft foot and (ii) angular
horizontal stroke
ST: (i) Flat shaft foot and (ii) straight horizontal stroke
2.8 2.4
‹s›
NT: Incurvation of the shaft foot to the right for ‹ › and
(ii) closed and angular curves for ‹s› ST: (i) Flat shaft foot
for ‹ › and (ii) open semi-circular curves for ‹s›
2.7 2.4

An Interpretable Deep Learning Approach for Script Type Analysis 11
‹t›
NT: Vertical pendant hairline of the headstroke ST: No
ornaments Note: Different levels of shaft projection above
headline and length of horizontal stroke
3.0 2.4
‹u› NT: Accentuated (diamond-shaped or sloped) headline
ST: Flat or left approach stroke for headline 3.7 3.0
forked ascender tops of ‹h› and ‹b›. This information loss, while limited, should
be considered during the analysis.
Palaeographical relevance of the subtype prototypes. To showcase how our pro-
totypes can be related to classical palaeographic analysis, we systematically
compare in Table 2 Derolez’s general morphological criteria to our Northern and
Southern Textualis prototypes, highlighting their variations by visualizing their
difference. We find that Derolez’s observations closely align with the variations
that our prototypes enable us to visualize. Additionally, we report our variability
scores
σNT
and
σST
for each letter, which were consistently higher for North-
ern Textualis, which is consistent with Derolez’s claim that this script subtype
generally exhibits higher intra-class variation.
4.3 Character graph analysis.
In this section, we provide examples of how our character graphs, together with
our prototypes, can support a detailed palaeographic analysis of the variations of
a specific character (examples presented in Figure 5).
Discriminative characters. We first analyze the results for four discriminative
characters, ‹a,o,p,h›. The letter ‹a› is often considered as a distinguishing
criterion between script types, so much so that W. Oeser [
41
] distinguished
seven categories within the Northern Textualis script subtype mainly based on
allographs of ‹a›. Most striking in our ‹a› character graph is that the prototypes
for NT5 ( ) and NT7 ( ) are actually closer to the ST prototypes. This is
consistent with the observation that open ‹a› forms are standard for ST. The
dispersion of the characters on the graph also provides insight into the variability
of ‹a› in this subtype. The group associated to NT1-4 corresponds to the closed
“box-a” form in NT2 ( ) and NT4 ( ) and the double-bow variant in NT1 ( )
and NT3 ( ). NT6 presents a more vertically elongated form and stands out ( ).
While there is morphological variations across ST documents, with round shapes
(ST1 ; ST2 ; ST5 ; ST6 ), or with more angular inner bows (ST3 ; ST4
; ST7 ), the consistent use of an open form, or only closed with a hairline,
distinguishes them from the NT subtype, and all ST documents prototypes are
closer to the ST prototype.
The letter ‹o› is also particularly discriminative. The treatment of its (gener-
ally) two mirroring arcs, using broken or semi-circular strokes, often has visual
echoes in letters with lobes and arcs like ‹b, c, e, p, q›, which contributes to the
visual evaluation of a hand or script type as wide/round or narrow/angular. This

12 M. Vlachou-Efstathiou et al.
Fig. 5: Character comparison graphs.
is confirmed by the fact that all the points of documents identified as NT are
above the diagonal and all the ones corresponding to ST below, meaning that the
prototypes for each document are closer to its subtype prototype than the other
one. For NT, the forms of NT2-5 ( ) are particularly well reconstructed,
better than NT1 ( ) and NT7 ( ) which are slightly more narrow and vertically
elongated. Again, the form of NT6 stands out ( ), it consists of double broken
strokes resulting in a narrow, quadrangle shape. Similarly for ST, ST1 ( ) and
ST2 ( ) are particularly close to the ST prototype, being less wide than ST3-7
( ). This is consistent with Derolez’ assertion that angularity/roundness
separates the two subtypes, while intra-class variation is associated with different
degrees of narrowness/breadth.
Regarding the letter ‹h›, it highlights one of the limitations of our approach.
The extended limb, characteristic of NT (cf. Table 2) is clearly present in all
associated documents and prototypes. However, because its position varies in
a document, the limb appears dimmed in the NT prototype ( ). Moreover, its
position varies significantly across different documents, which can result in a
greater distance between a document prototype and the NT prototype. This
explains why the NT2 prototype is actually more similar to the ST prototype,
while NT5 is as close to both. Note that the shift of the limb in NT2 (not curved
to the left but rather extended straight down) compared to the NT prototype is so
significant that our filtering partially erases it (as can be confirmed by examining
the masked region, similar to Figure 2b), which was flagged by our automatic

An Interpretable Deep Learning Approach for Script Type Analysis 13
failure identification. This emphasizes the necessity to confront our graphs with
the visual appearance of the prototypes and the documents for interpretation.
For the letter ‹p›, we note that the consistent presence of artificial spurs at
the baseline level and in general of decorations are characteristic of NT ( ). The
two subtypes appear well separated, except for the deviation of NT5 ( ), which
we will further analyze in Section 4.4.
Non-discriminative characters. We now examine ‹i› and ‹r›. In our graphs, the
points corresponding to these letters in all documents are close to the diagonal,
i.e., they are as close to both the NT and ST prototypes. The fact that they
are almost all close to the origin indicates that they present little variation. For
‹i›, NT1 and NT3 stand out as more typical of NT, and they indeed show clear
diamond-shaped headlines ( ). The NT4 prototype ( ), on the other hand,
is actually closer to the ST prototypes, and it does not present any headline
decoration typical of NT. For both characters, the NT6 prototypes are much
further than the rest from both the NT and ST prototypes and correspond to
much narrower and elongated forms ( ).
4.4 Document graph analysis.
In this section, we discuss and interpret document graphs for the examples
presented in Figure 6. Additionally, we visualize the differences between the
document prototypes and subtype prototypes in Figure 7, leveraging them to
better understand the graphs.
Class-representative cases. We start by examining two documents that are very
typical of their subtype. For NT3, a 14
th
century manuscript in Latin, the graph
clearly shows that all the character prototypes are closer to the NT prototypes
than the ST prototypes. More in details, NT3 presents a closed, double bow ‹a›
( ), which progressively dominated over other variants from the end of the 13
th
c. [
16
], ascenders are consistently sloped on the left side ( , hairlines directed
to the right at the baselines ( ), and there are clear diamond-shaped
headlines ( ). All characters thus fully correspond to the expected NT forms.
For ST2, copied in the first decade of 14
th
c. in the Iberian peninsula, the
character prototypes are, on the contrary, all closer and conforming to the ST
prototypes: compact letters with very short flat-top ascenders ( ), flat-feet
descenders ( ) and strokes so bold, hairlines almost become invisible ( ).
Ambiguous cases. A particular interest of our document graph method is the
identification and analysis of documents that partially diverge from their assigned
subtype. NT5 stands out in the graphs, as seven character prototypes are closer
to ST than to NT prototypes, with a particular difference for ‹a,g,q› ( ).
Copied in the 13
th
c., in the Paris/Orleans area, NT5’s forms are significantly
smaller in size, an example of “pearl-script”, generally used for Parisian pocket-
Bibles. Even though its level of execution is still Formata, due to their size, certain
letters are simplified, resulting in forms that are closer to ST, like open ‹a› ( )

14 M. Vlachou-Efstathiou et al.
Fig. 6: Document comparison graphs.
- a characteristic of early NT samples -, and spurless ‹p› ( ). On the contrary,
a closer examination of ‹g› reveals it is in reality characteristic of NT, but not
well modeled by our prototypes, in part because of the high level of variation,
resulting in a blurred prototype, and in part because it was too different from
the reference documents, which was actually flagged by our automatic failure
case identification. ST6, copied in Florence at the end of 14
th
/beginning of 15
th
c., also presents two class diverging characters, ‹b› and ‹s/ ›. For ‹s/ ›, even
though not directly obvious due to the fusion of the two allographs, ‹ ›’s foot is
curved to the right ( ), a characteristic proper to NT. At the same time, ‹b›’s
slightly dislocated lobe ( ), also present in ‹h› ( ) and ‹p› ( ), characteristic of
later examples of Textualis [
16
], is disrupting the typical circular lobe shape of
ST. These dislocated lobes are clear in our Figure 7.
Later Textualis examples. Later specimens of both subtypes present notable
differences from their earlier counterparts. NT6, copied in the last quarter of
the 14
th
c. is an example of later Northern Textualis, and, while the graph shows
that all prototypes are closer to the NT prototypes than the ST prototypes, it
also clearly shows that they are very different from both, i.e., they are more on
the top right of the graph. This document particularity lies in its strict angular
forms, with diamond-shaped minim feet ( ) as well as its exaggerated narrow
shapes, with a total absence of round strokes for arcs ( ), typical of 14
th
-15
th
c. (esp. Northern) Textualis.ST7 is an example of late, 15
th
c. Iberian Textualis,
discussed separately by Derolez due to its tendency towards more angular forms.

An Interpretable Deep Learning Approach for Script Type Analysis 15
Fig. 7: Visual comparison. Subtype and document prototypes and their pixel-
wise differences with positive values in blue and negative in red.
Particularities of this type lie in the presence of hairlines and angular shapes
( ) alongside typical rounder ones ( ) with flat tops and feet ( ).
However, these characteristics are only discernible while looking at the prototypes
and their differences, and not directly in the graph, where the prototypes are
closer to ST, and not particularly poorly reconstructed. This can be understood
both by the fact that these particularities do not make the prototypes more
similar to the NT ones, and by the fact that another Iberian Textualis (ST4),
from the same manuscript but from a different hand, was used in our reference set,
and thus our ST prototype already models some characteristics of this Textualis
type. This highlights the impact of the prototype training data on our analysis,
prototypes utilized as the basis for our comparison graph axes.
5 Conclusion
In this work, we introduced a deep learning-based methodology for interpretable
script comparison and analysis. By applying it to the two subtypes of Textualis
Formata script type defined by A. Derolez —Northern and Southern Textualis—
we showed how such an approach can complement qualitative document analysis,
by quantifying specific elements and summarizing information. We believe our
approach contributes to bridging the gap between traditional and learning-based
approaches to paleography.
Aknowledgements This study was supported by the CNRS through MITI and
the 80|Prime program (CrEMe Caractérisation des écritures médiévales), and by
the European Research Council (ERC project DISCOVER, number 101076028).
We thank Ségolène Albouy, Raphaël Baena, Sonat Baltacı, Syrine Kalleli, and
Elliot Vincent for valuable feedback.

16 M. Vlachou-Efstathiou et al.
References
1.
Aiolli, F., Simi, M., Sona, D., Sperduti, A., Starita, A., Zaccagnini, G.: Spi: a
system for palaeographic inspections. AIIA Notizie 4, 34–38 (1999)
2.
Alba, R., Rubin, G., Boschetti, F., Fischer, F., Clérice, T., Chagué, A.: HTRomance,
Medieval Italian corpus of ground-truth for Handwritten Text Recognition and
Layout Segmentation [dataset] (2023). https://doi.org/10.5281/zenodo.8272751,
https://github.com/HTRomance-Project/medieval-italian, v1.0.1
3.
Baird, H.S.: Model-directed document image analysis. In: Proceedings of the Sym-
posium on Document Image Understanding Technology. vol. 1 (1999)
4.
Berg-Kirkpatrick, T., Durrett, G., Klein, D.: Unsupervised transcription of historical
documents. In: Proceedings of the 51st Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers). pp. 207–217 (2013)
5.
Bordier, J., Gille Levenson, M., Brisville-Fertin, O., Clérice, T., Chagué, A.:
HTRomance, Medieval Spain corpus of ground-truth for Handwritten Text
Recognition and Layout Segmentation [dataset] (2023), https://github.com/
HTRomance-Project/middle-ages-in-spain, v0.0.6
6.
Cencetti, G.: Lineamenti di Storia della scrittura latina: dalle lezioni di Paleografia
(Bologna a.a. 1953-54). Guerrini Ferri, G., Bologna (1997)
7.
Christlein, V., Bernecker, D., Maier, A., Angelopoulou, E.: Offline writer identifica-
tion using convolutional neural network activation features. In: Pattern Recognition:
37th German Conference, GCPR 2015, Aachen, Germany, October 7-10, 2015,
Proceedings 37. pp. 540–552. Springer (2015)
8.
Ciula, A.: Digital palaeography: using the digital representation of medieval script
to support palaeographic analysis. Digital Medievalist 1(2005)
9.
Cloppet, F., Daher, H., Églin, V., Emptoz, H., Exbrayat, M., Joutel, G., Lebourgeois,
F., Martin, L., Moalla, I., Siddiqi, I., Vincent, N.: New Tools for Exploring, Analysing
and Categorising Medieval Scripts. Digital Medievalist 7(Feb 2012). https://doi.
org/10.16995/dm.44
10.
Cloppet, F., Eglin, V., Helias-Baron, M., Kieu, V.C., Stutzmann, D., Vincent,
N.: ICDAR 2017 Competition on the Classification of Medieval Handwritings in
Latin Script. In: 14th IAPR International Conference on Document Analysis and
Recognition. ICDAR 2017. pp. 1371–1376. CPS, Kyoto (2017). https://doi.org/
DOI10.1109/ICDAR.2017.224, 00000
11.
Cloppet, F., Eglin, V., Kieu, V.C., Stutzmann, D., Vincent, N.: ICFHR2016 Com-
petition on the Classification of Medieval Handwritings in Latin Script. Proceedings
of International Conference on Frontiers in Handwriting Recognition pp. 590–595
(2016)
12.
Clérice, T., Chagué, A., Vlachou-Efstathiou, M.: CREMMA Medii Aevi [dataset]
(Oct 2023), https://github.com/HTR-United/CREMMA-Medieval-LAT, v0.1.2
13.
Clérice, T., Pinche, A.: Choco-Mufin, a tool for controlling characters used in
OCR and HTR projects (Sep 2021). https://doi.org/10.5281/zenodo.5356154,https:
//github.com/PonteIneptique/choco-mufin
14.
Daher, H., Églin, V., Brès, S., Vincent, N.: Étude de la dynamique des écritures
médiévales: analyse et classification des formes écrites. Gazette du livre médiéval
56(1), 21–41 (2011)
15.
Davis, L.F.: Towards an automated system of script classification. Manuscripta
42(3), 193–201 (1998)
16.
Derolez, A.: The palaeography of Gothic manuscript books: From the twelfth to
the early sixteenth century. Cambridge University Press (2003)

An Interpretable Deep Learning Approach for Script Type Analysis 17
17.
Djeddi, C., Meslati, L.S., Siddiqi, I., Ennaji, A., El Abed, H., Gattal, A.: Evaluation
of texture features for offline arabic writer identification. In: 2014 11th IAPR
international workshop on document analysis systems. pp. 106–110. IEEE (2014)
18.
Djeddi, C., Siddiqi, I., Souici-Meslati, L., Ennaji, A.: Codebook for writer char-
acterization: A vocabulary of patterns or a mere representation space? In: 2013
12th International Conference on Document Analysis and Recognition. pp. 423–427.
IEEE (2013)
19.
Gasparri, F.: Remarques sur la terminologie paléographique. Revue d’Histoire des
Textes 13(1964), 111–114 (1966)
20.
Gilissen, L.: L’expertise des écritures médiévales: recherche d’une méthode avec ap-
plication à un manuscrit du XIe siècle: le lectionnaire de Lobbes, Codex Bruxellensis
18018. Scriptorium/Les Publications de Scriptorium 6(1973)
21. Gilissen, L.: III. ductus et rapport modulaire. Scriptorium 29(2), 235–244 (1975)
22.
Gille Levenson, M.: Towards a general open dataset and model for late medieval
Castilian text recognition (HTR/OCR). Journal of Data Mining and Digital Hu-
manities (2023). https://doi.org/10.46298/jdmdh.10416
23.
Glaise, A., Clérice, T., Boschetti, F., Fischer, F., Chagué, A.: HTRomance, Medieval
Latin corpus of ground-truth for Handwritten Text Recognition and Layout Segmen-
tation [dataset] (2024), https://github.com/HTRomance-Project/medieval-latin,
v0.0.6
24. Goyal, K., Dyer, C., Warren, C., G’Sell, M., Berg-Kirkpatrick, T.: A probabilistic
generative model for typographical analysis of early modern printing. arXiv preprint
arXiv:2005.01646 (2020)
25.
Hannad, Y., Siddiqi, I., El Kettani, M.E.Y.: Writer identification using texture
descriptors of handwritten fragments. Expert Systems with Applications 47, 14–22
(2016)
26.
Hassner, T., Rehbein, M., Stokes, P.A., Wolf, L.: Computation and palaeography:
potentials and limits. Kodikologie und Paläographie im digitalen Zeitalter 3, 1–30
(2015)
27.
He, S., Schomaker, L.: Delta-n hinge: rotation-invariant features for writer identifica-
tion. In: 2014 22nd International conference on pattern recognition. pp. 2023–2028.
IEEE (2014)
28.
He, S., Schomaker, L.: Deep adaptive learning for writer identification based on
single handwritten word images. Pattern Recognition 88, 64–74 (2019)
29.
He, S., Wiering, M., Schomaker, L.: Junction detection in handwritten documents
and its application to writer identification. Pattern Recognition 48(12), 4036–4048
(2015)
30.
Hochberg, J., Kelly, P., Thomas, T., Kerns, L.: Automatic script identification from
document images using cluster-based templates. IEEE Transactions on Pattern
Analysis and Machine Intelligence 19(2), 176–181 (1997)
31.
Kestemont, M., Christlein, V., Stutzmann, D.: Artificial paleography: computational
approaches to identifying script types in medieval manuscripts. Speculum 92(S1),
S86–S109 (2017)
32.
Kopec, G.E., Lomelin, M.: Supervised template estimation for document image
decoding. IEEE Transactions on Pattern Analysis and Machine Intelligence 19(12),
1313–1324 (1997)
33.
Kordon, F., Weichselbaumer, N., Herz, R., Mossman, S., Potten, E., Seuret, M.,
Mayr, M., Christlein, V.: Classification of incunable glyphs and out-of-distribution
detection with joint energy-based models. International Journal on Document
Analysis and Recognition (IJDAR) 26(3), 223–240 (2023)

18 M. Vlachou-Efstathiou et al.
34.
Lebourgeois, F., Moalla, I.: Caractérisation des écritures médiévales par des méth-
odes statistiques basées sur les cooccurrences. Gazette du livre médiéval 56-57,
72–100 (2011)
35.
Leroy, N., Pinche, A., Camps, J.B., Clérice, T., Chagué, A.: HTRomance, Medieval
French corpus of ground-truth for Handwritten Text Recognition and Layout
Segmentation [dataset], https://github.com/HTRomance-Project/medieval-french,
v0.0.7
36.
Mamatsis, A.R., Mamatsi, E., Chalatsis, C., Arabadjis, D., Kampouri, P., Pa-
paodysseus, C.: A novel methodology for writer (hand) identification: establishing
rigas feraios wrote two important greek documents discovered in romania. Heritage
Science 11(1), 38 (2023)
37.
McGillivray, M.: Statistical analysis of digital paleographic data: what can it tell
us? Digital Studies/Le champ numérique 11 (2005)
38.
Moalla, I., Lebourgeois, F., Emptoz, H., Alimi, A.: Image analysis for palaeography
inspection. In: Second International Conference on Document Image Analysis for
Libraries (DIAL’06). pp. 8–pp. IEEE (2006)
39.
Muzerelle, D.: À la recherche d’algorithmes experts en écritures médiévales. Gazette
du livre médiéval 56(1), 5–20 (2011). https://doi.org/10.3406/galim.2011.1979
40.
Nigam, S., Verma, S., Nagabhushan, P.: Document analysis and recognition: A
survey. Authorea Preprints (2021)
41.
Oeser, W.: Das «a» als Grundlage für Schriftvarianten in der gotischen Buchschrift.
Scriptorium 25(1), 25–45 (1971)
42.
Ornato, E.: Ii. statistique et paléographie: peut-on utiliser le rapport modulaire
dans l’expertise des écritures médiévales? Scriptorium 29(2), 198–234 (1975)
43. Parkes, M.B.: English cursive book hands, 1250-1500. Oxford : Clarendon (1969)
44.
Pinche, A.: Cremma Medieval [dataset] (Oct 2023), https://github.com/
HTR-United/cremma-medieval
45.
Pinche, A., Clérice, T., Chagué, A., Camps, J.B., Vlachou-Efstathiou, M., Leven-
son, M.G., Brisville-Fertin, O., Boschetti, F., Fischer, F., Gervers, M., et al.:
Catmus-medieval: Consistent approaches to transcribing manuscripts (2023),
https://univ-lyon3.hal.science/hal-04453952v1
46.
Poulle, E.: Paléographie et méthodologie: vers l’analyse scientifique des écritures
médiévales. Bibliothèque de l’École des chartes 132(1), 101–110 (1974)
47.
Ramel, J.Y., Sidère, N., Rayar, F.: Interactive layout analysis, content extraction,
and transcription of historical printed books using pattern redundancy analysis.
Literary and Linguistic Computing 28(2), 301–314 (2013)
48.
Schomaker, L., Bulacu, M.: Automatic writer identification using connected-
component contours and edge-based features of uppercase western script. IEEE
transactions on pattern analysis and machine intelligence 26(6), 787–798 (2004)
49.
Schomaker, L., Franke, K., Bulacu, M.: Using codebooks of fragmented connected-
component contours in forensic and historic writer identification. Pattern Recogni-
tion Letters 28(6), 719–727 (2007)
50.
Siddiqi, I., Vincent, N.: Text independent writer recognition using redundant writing
patterns with contour-based orientation and curvature features. Pattern Recognition
43(11), 3853–3865 (2010)
51.
Siglidis, I., Gonthier, N., Gaubil, J., Monnier, T., Aubry, M.: The learnable type-
writer: A generative approach to text line analysis (2023), https://arxiv.org/abs/
2302.01660
52.
Sirat, C.: L’examen des écritures: l’œil et la machine: essai de méthodologie. Ed.
du Centre National de la Recherche Scientifique (1981)

An Interpretable Deep Learning Approach for Script Type Analysis 19
53.
Smith, M.: (review) derolez (albert), the palaeography of gothic manuscript books.
from the twelfth to the early sixteenth century, cambridge, 2003. Scriptorium 58(2),
274–279 (2004)
54.
Sommerschield, T., Assael, Y., Pavlopoulos, J., Stefanak, V., Senior, A., Dyer, C.,
Bodel, J., Prag, J., Androutsopoulos, I., de Freitas, N.: Machine learning for ancient
languages: A survey. Computational Linguistics 49(3), 703–747 (2023)
55.
Stansbury, M.: The computer and the classification of script. In: Kodikologie und
Paläographie im digitalen Zeitalter - Codicology and Palaeography in the Digital
Age. vol. 2, p. 238. BoD, Norderstedt (2009)
56.
Stokes, P.A.: Describing handwriting, part i-v. Blog Post (2011), https://digipal.
eu/blog/describing-handwriting-part-i/, last accessed on 15/03/2024
57.
Stutzmann, D.: Variability as a key factor for understanding medieval scripts: the
oriflamms project (anr-12-corp-0010). In: Brookes, S., Rehbein, M., Stokes, P. (eds.)
Digital Palaeography. Digital Research in the Arts and Humanities, Routledge,
https://halshs.archives-ouvertes.fr/halshs-01778620
58.
Stutzmann, D.: Nomenklatur der gotischen Buchschriften: Nennen? Systematisieren?
Wie und wozu? (Rezension über: Albert Derolez: The Palaeography of Gothic
Manuscript Books. From the Twelfth to the Early Sixteenth Century. Cambridge
u.a.: Cambridge University Press 2003.). IASLonline (2005), http://www.iaslonline.
de/index.php?vorgang_id=995
59.
Stutzmann, D.: Paléographie statistique pour décrire, identifier, dater... Normaliser
pour coopérer et aller plus loin ? In: Kodikologie und Paläographie im digitalen
Zeitalter 2 - Codicology and Palaeography in the Digital Age 2, pp. 247–277. No. 3
in Schriften des Instituts für Dokumentologie und Editorik, BoD, Norderstedt
(2010), https://kups.ub.uni-koeln.de/4353/
60.
Stutzmann, D.: Système graphique et normes sociales : pour une analyse électronique
des écritures médiévales. In: Medieval Autograph Manuscripts. Proceedings of the
XVIIth Colloquium of the Comité International de Paléographie Latine, held
in Ljubljana, 7-10 September 2010, pp. 429–434. No. 36 in Bibliologia, Brepols,
Turnhout (2013), https://www.brepolsonline.net/doi/10.1484/M.BIB.1.101494
61.
Stutzmann, D.: Clustering of medieval scripts through computer image analysis:
Towards an evaluation protocol. Digital Medievalist 10 (Jun 2016). https://doi.
org/10.16995/dm.61
62. Stutzmann, D.: Ecmen (2017), https://github.com/oriflamms/ECMEN
63.
Stutzmann, D.: Les «manuscrits datés», base de données sur l’écriture. In: De Rober-
tis, T., Giovè Marchioli, N. (eds.) Catalogazione, storia della scrittura, storia del
libro. I Manoscritti datati d’Italia vent’anni dopo, pp. 155–207. SISMEL - Edizioni
del Galluzzo, Firenze (2017)
64.
Stutzmann, D., Helias-Baron, M.: ICDAR 2017 Competition on the Classification
of Medieval Handwritings in Latin Script - Dataset (Nov 2017), https://zenodo.
org/record/5527690
65.
Tang, Y., Wu, X.: Text-independent writer identification via cnn features and
joint bayesian. In: 2016 15th International Conference on Frontiers in Handwriting
Recognition (ICFHR). pp. 566–571. IEEE (2016)
66.
Tomiello, A.: Dalla littera antiqua alla littera textualis. Gazette du livre médiéval
29(1), 1–6 (1996)
67.
Wolf, L., Dershowitz, N., Potikha, L., German, T., Shweka, R., Choueka, Y.:
Automatic paleographic exploration of genizah manuscripts. In: Fischer, F., Fritze,
C., Vogeler, G. (eds.) Kodikologie und Palaographie im Digitalen Zeitalter 2 -
Codicology and Palaeography in the Digital Age 2, Schriften des Instituts für

20 M. Vlachou-Efstathiou et al.
Dokumentologie und Editorik, vol. 3, pp. 157–179. BoD, Norderstedt, Germany
(2011)
68.
Xu, Y., Nagy, G.: Prototype extraction and adaptive ocr. IEEE Transactions on
Pattern Analysis and Machine Intelligence 21(12), 1280–1296 (1999)