=Paper=
{{Paper
|id=Vol-3834/paper104
|storemode=property
|title= Textual Transmission without Borders: Multiple Multilingual Alignment and Stemmatology of the ``Lancelot en prose'' (Medieval French, Castilian, Italian)
|pdfUrl=https://ceur-ws.org/Vol-3834/paper104.pdf
|volume=Vol-3834
|authors=Matthias Gille Levenson,Lucence Ing,Jean-Baptiste Camps
|dblpUrl=https://dblp.org/rec/conf/chr/LevensonIC24
}}
== Textual Transmission without Borders: Multiple Multilingual Alignment and Stemmatology of the ``Lancelot en prose'' (Medieval French, Castilian, Italian)==
https://ceur-ws.org/Vol-3834/paper104.pdf
Textual Transmission without Borders:
Multiple Multilingual Alignment and Stemmatology
of the “Lancelot en prose”
(Medieval French, Castilian, Italian)
Matthias Gille Levenson1,2,3,∗ , Lucence Ing1,3,∗ and Jean-Baptiste Camps1,3
1
Centre Jean Mabillon, École nationale des chartes, Paris Sciences & Lettres, France
2
CIHAM, UMR 5648, École Normale Supérieure de Lyon, France
3
ÉquipEx Biblissima+
Abstract
This study focuses on the problem of multilingual medieval text alignment, which presents specific chal-
lenges, due to the absence of modern punctuation in the texts and the non-standard forms of medieval
languages. In order to perform the alignment of several witnesses from the multilingual tradition of the
prose Lancelot, we first develop an automatic text segmenter based on BERT and then align the produced
segments using Bertalign. This alignment is then used to produce stemmatological hypotheses, using
phylogenetic methods. The aligned sequences are clustered independently by two human annotators
and a clustering algorithm (DBScan), and the resulting variant tables submitted to maximum parsimony
analysis, in order to produce trees. The trees are then compared and discussed in light of philological
knowledge. Results tend to show that automatically clustered sequences can provide results comparable
to those of human annotation.
Keywords
Multilingual alignment, Text segmentation, Medieval Arthurian literature, Stemmatology
1. Introduction
The production and transmission of written texts during Antiquity and the Middle Ages in-
volved a process of manual copying. During this process, the text was progressively trans-
formed by errors and innovations as it circulated, each copy introducing successive modifica-
tions. Since the 19th century, philologists have taken to study the transmission of texts, based
on innovations, using the genealogical tree (stemma codicum) as a metaphor to visually repre-
sent this transmission process. Yet, the study of textual transmission is still often limited to the
copies produced in a given language (e.g. Medieval French) and do not necessarily encompass
all the translations, in other medieval languages, that were part of the history of a given work.
This is due in part to stark difÏculties in aligning and analysing multilingual traditions, but
CHR 2024: Computational Humanities Research Conference, December 4–6, 2024, Aarhus, Denmark
∗
These authors contributed equally.
£ matthias.gille-levenson@ens-lyon.fr (M. Gille Levenson); lucence.ing@chartes.psl.eu (L. Ing);
jean-baptiste.camps@chartes.psl.eu (J. Camps)
ȉ 0000-0001-9488-5986 (M. Gille Levenson); 0000-0002-8742-3000 (L. Ing); 0000-0003-0385-7037 (J. Camps)
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
65
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
poses the risk of giving us a very limited and partial view of medieval cultures, constrained by
linguistic “frontiers” that were actually quite blurry and porous at the time.
1.1. The challenge of multilingual traditions
Multilingual collation of texts has attracted a few contributions, mostly focused on encoding or
displaying editions of widely circulated texts [24, 2, 20, 26]. Yet, due to technical and modelling
challenges, most critical editions are limited to a single language. National academic traditions,
and the relative rarity of comparative studies in, for instance, current Romance Philology might
be a factor, since research is often focused on editing either the source text or a specific lan-
guage version. This philological reality can also be explained by material issues, namely the
great difÏculty of taking into account a large number of witnesses in the editorial process, the
elaboration of a stemma, and the production of editions. Already true for large unilingual tra-
ditions alone, this is even more acute for multilingual traditions of widely circulated works,
since adding distinct language versions only increases the complexity of the task.
The use of computer tools can help to overcome some of these limitations and pave the
way for global studies of multilingual traditions, and even for the production of multilingual
editions of textual traditions in the longer term.
The interest of this approach lies in the history of the text, in a global romanistic approach:
the aim is to consider the textual tradition as a whole, and to contribute to the progress of
knowledge on the text in general. Connecting local traditions will be profitable in order to
produce global knowledge on the reception of these multiply translated texts, while possibly
improving local knowledge on a specific tradition. Multilingual collation could, for example,
help clarify the history of the text in cases where one translation is the direct archetype of
another.
This paper focuses on computational multilingual alignment and collation (subsection 2.2),
at “sentence” level and on similar fragments, on the classification of aligned sequences into
distinct variants (subsection 2.3), and on their subsequent stemmatological analysis (section 3).1
To perform the alignment, due to the nature of the textual data we use, a first step of automatic
segmentation of the text is necessary (subsection 2.1). The paper concentrates on the study of
the multilingual tradition of the Lancelot en prose, in order to understand the history of the
text and its translations. We take this text as a case study, which enables us to present a new
methodology that can help the study of various multilingual medieval traditions.
1.2. The multilingual tradition of the Lancelot
The study is based on a few witnesses of a complex medieval tradition, that of the Lancelot en
prose, an anonymous text composed in the first third of the 13th century which enjoyed great
success throughout the medieval period, with at least 126 manuscript witnesses, followed by
many printed editions from 1488 onwards [13, 4]. This great success is also evidenced by the
translations it has undergone (into Castilian, Catalan, Italian, High German, and Dutch). The
1
This article sets apart the identification of macroscopic displacement of large and medium-sized fragments (at
paragraph level, for example) that can occur in complex traditions (see subsection 1.2).
66
alignment of texts in a multilingual framework allows, through the recording of variants, to
establish the tradition of these translations.
In this paper, we focused solely on the Romance tradition, in particular the French source,
and the translations in Castilian and Italian (with the exclusion of the Catalan translation, of
which only small fragments remain). For both of these translations, only one complete witness
is preserved: in a manuscript produced in Florence in the last quarter of the 14th century, for the
Italian Lancellotto (Firenze, Biblioteca della Fondazione Ezio Franceschini 1); in a 16th century
manuscript, copied from a 1414 exemplar according to its colophon, for the Castilian Lanzarote
(Madrid, Biblioteca Nacional de España, 9611). Both texts have been edited [5, 9].
Given the extremely large number of witnesses of the French Lancelot, we have selected a
sample of five witnesses. They have been chosen on the basis of their supposed relationships
with the translations, according to existing philological knowledge: Paris, BnF, fr. 111 (15th c.)
is supposed to be close to the Lancellotto [6], while Paris, BnF, fr. 751 (13th c., 2nd half), or more
precisely the family of which it is a part, would be close to the Lanzarote [9]. In addition, due
to their easy availability and well known versions, we added several reference points, in the
form of the edition by Sommer, based on ms. London, BL, Add. 10293 (beg. 14th c.) [22, 23] and
Micha, based on mss Cambridge, Corpus Christi College, 45 (132/2 ) for volume 2 and Oxford,
Bodleian Library, Rawlinson D. 899 (14th c.) for volume 4 [17, 18]. Finally, as a representative of
the late French tradition, whose place in the genealogy remains to be elucidated, we included
the incunabula edition of 1488 (from exemplar Paris, BnF, RES-Y2-46 and RES-Y2-47; Rouen,
Jean le Bourgois et Paris, Jean Dupré, 1488). For a list of the witnesses, see Appendix A.
The Lancelot is a very long prose text, and therefore it can be highly unstable from one
witness to another. The witnesses containing the translations are not spared from this insta-
bility, and they are also fragmentary. Lancellotto is especially so, as it presents only three non-
consecutive episodes of the text. To enable alignment, the first step was therefore to identify
corresponding passages from one witness to another (Appendix B), and to retain only what
could be compared. This is why the studied text segments have identifiers: ii-48, ii-61, and
iv-75, corresponding to the sections of the text as they appear in Micha edition.2
1.3. State of the art on sentence alignment
While multilingual alignment is a fairly active field in NLP, relatively little work has been done
on the production and use of alignment methods for philological, stemmatological and ecdotic
purposes, in the field of heritage text with pre-orthographic languages.
Birnbaum and Eckhoff [3] are considering the creation of a multilingual alignment tool,
based on parts of speech only, which works in cases of literal translations, such as the Old
Church Slavonic and an original Greek version of the Codex Supralensis. The work of Meinecke,
Wrisley, and Jänicke [16] on French epic literature, explores the possibilities offered by the use
2
The first part of the identifier corresponds to the volume, the second to the number of the first segment according to
Micha [17, 18] Due to significant textual variation and the current capabilities of the aligner, we have subdivided
the longest sections, ii-61 and iv-75, into two and six parts respectively, so that none of the segments exceed
1,000 tokens. Despite this choice, the translations exhibit a significant number of omissions that are difÏcult to
detect automatically. For example, Lanzarote contains the episode of the sparrowhawk (segment ii-61-2), but in an
extremely shortened version, only a few sentences (Appendix C).
67
of semantic representations of words through embeddings, from a unilingual perspective only.
However, nothing is said about the definition and identification of “sentences” in the source
texts, as the unit chosen to compare the versions seems to be the verse. Yet, embedding-based
similarity calculations seem to be a promising venue for text alignment [27].
Recently, Liu and Zhu [15] have published a tool called Bertalign, a two-steps algorithm
that makes use of sentence-transformers and multilingual sentence embeddings, based on the
LaBSE model (“Language agnostic BERT Sentence Embeddings”) [1]. In particular, it allows to
align 1 to 𝑛 and 𝑛 to 1 texts fragments. Bertalign works with segments fixed upstream and is
designed for modern language states.3
2. Methodology
In this paper, we use the fixed fragments method designed by Liu and Zhu and we make use
of Bertalign to perform the global alignments, while at the same time proposing a specific seg-
mentation approach prior to the alignment.4 We describe our processing chain: segmentation,
pseudo-sentences alignment, classification into distinct variants and stemmatological analysis.
2.1. Segmentation
The alignment task requires fixed segments, as the tool we rely on, Bertalign, is designed for
contemporary languages for which sentence segmentation is not a problem (the period “.” is
enough to split the corpus, as modern translations tend to reproduce the same divisions sen-
tence by sentence). It can be an issue for medieval languages, where the notion of a modern
sentence, which begins with a capital letter and ends with a period, doesn’t really exist. Punc-
tuation is highly variable depending on the copyist, even if a global and comparative study has
yet to be carried out. Let’s take the following two sentences as examples:
BnF fr. 111 : ains sem part si tost quil leut commandee a dieu et cheuauche en telle maniere
iucqua tierce tant quil uient a lissue de la fourest. et il tourne a destre vers ung chemin uieil
et ancien.
BnF fr. 751 : Ains sen part si tost comme il lot comandee a dieu. et cheuauche en tel
maniere. tant quil uient a lissue de la forest et il torne a destre uers le chemin uiez et
ancien
Translation (for both passages): [But he goes away as soon as he commanded her
to God and rides [until 9 a.m.], until he comes to the entrance of the forest, and he
turns right towards an old pathway.]
The punctuation of these two transcriptions is original: it shows that it is not a reliable
marker for syntactic segmentation, as it can vary greatly depending on the source, and this
3
A more recent preprint picks up on this work and compares Bertalign to a new architecture [14], but we haven’t
had time to look at how the algorithm works and how good it performs on our data.
4
The word-alignment phase, because it requires its own methodology, will not be dealt with in this article: we’re
only interested in the “macroalignment” phase, at the sentence or syntagm level.
68
variability increases during the transcription and editing of the text. Moreover, original punc-
tuation is still very little taken into account by editors, who re-punctuate the text and can make
highly variable choices. It is therefore necessary to divide the texts into equivalent segments
that can then be compared and aligned. To do this, we have decided to base our approach on
sentence syntax. The assumption here is that there is a certain degree of correspondence be-
tween syntactic and semantic units, and that the syntactic tokenisation should be an efÏcient
way to approach the different semantic units of the text. Let’s take an example (Table 1): in the
second line, “si” is translated by “e” in Castilian, but the syntactic units are preserved, allowing
the alignment of segments that are formally quite different but share the same overall meaning.
Table 1
A correspondence between syntactic and semantic segments in fragment ii-48, on two following seg-
ments. The first one corresponds to the end of the preceding example. Each cell represents an alignment
segment
Micha Sommer Lanzarote Lancellotto
et il torne a destre et torne a destre en e dexo el camino Ed e torna a destra
vers un chemin viés vn petit chemin viez e tomo a mano inverso uno camino
et ancien; derecha vn camino vecchio e antico,
pequeño | e biejo
lleno de yerba
si ne demora gueres Si ne demora gaires e non andubo mucho si non dimora guari
2.1.1. Segmentation methods
To segment the texts, we try, compare and evaluate two methods: the first uses regular expres-
sions, the second one uses AutoModelForTokenClassification BERT models.
Using regular expressions The first method we tried was based on the identification of
function words, which required a language-by-language analysis of syntactic delimiters. Regu-
lar expressions have been developed to facilitate this identification and manage the high degree
of graphical variation. Here are a few examples of such markers in French:
"[Tt]ant que", "[Aa]nsi", "[sS]i est car", "[sS]i est que", "[sS]i",
"[Ee]t", "mais", "qu[ie] si", "[kc]ar", "[Qq]u?ant", "quar", "[Qq]u'",
"[Qq]u[ei]", "[Qq]uel", "[Dd]on[ct]", "[pP]uis"
In Spanish:
"[Pp]ues que", "[pP]or?que", "[pP]er", "[Qq]ue", "a que", "si",
"[Dd]o", "[Ee]", "commo quier que", "mas", "ass?i como", "como",
"[Aa]n?ss?i", "para", "aquel que"
And in Italian:
"[mM]a si", "[Ss]i", "tanto che", "che", "e", "donde", "ch'?"
69
This regular expression method is limited, however, as it doesn’t allow for fine segmentation
in the case of repeated markers (you have to think of all possible combinations), which leads
to over-segmentation of the text, as we’ll see below.
Creating an ad hoc segmenter Another problem with rule-based segmentation is that it
relies solely on the identification of character strings. Besides the fact that this leads to a
proliferation of rules for languages with significant graphical variation (ainsi, insi, einsi for
thus, for instance), it also identifies tokens that do not serve to delimit coherent syntactic units.
Indeed, for example, the coordinating conjunction et (and) is a good delimiter of clauses and
sentences, but it also serves to coordinate elements within the same syntactic units.
Let’s have a look at the following two sentences:
• “dont n’est a il mie merveille, se vos en estes couroucie et je vos pri que vos me diez que ce est”
[so it is not a surprise if you are angry and I beg you to tell me for what]: the conjunction
et (and) coordinates two propositions and can be used as a delimiter;
• “si me parti ier matin dolante et couroucie” [I left yesterday morning sad and angry]: the
conjunction coordinates two adjectives; it must not be used as a delimiter.
This is why the use of language models for identifying delimiters seemed relevant. Indeed,
these models allow for capturing the semantics of different tokens, and thus differentiating
the uses of the same words. It was decided to use AutoModelForTokenClassification BERT
models [28], with the aim of identifying tokens that can serve as delimiters of these textual
units. These models are usually trained on contemporary languages, but, as we will see, the
large amount of data used to train them allows for good generalization and good results on our
ancient language states.5
To produce specific training data for our corpus, tokens identified as delimiters were labeled
with the value 1, all other tokens with the value 0, based on human evaluation and a specific
set of grammatical rules.6
2.1.2. Segmenter train corpus
The models were trained on three languages: French, Castilian and Italian, in their medieval
forms. The medieval Castilian corpus comprises around 8,000 tokenised lines (about 36,200
words), half of which are various texts from the 13th to 15th centuries, retrieved from the
CORDE linguistic database [19]. The other half of the corpus consists of fragments taken
from the Lanzarote, which are not included in the corpus dealt with in this paper. The French
and Italian models were trained solely on data from the Lancelot, which is debatable from a
model generalization perspective but was motivated by the desire to quickly achieve effective
models for our texts. They were trained on 12,700 and 28,000 words (1,000 and 1,500 lines),
5
The models which were used are: for the Castilian, the BETO model [8]; for the French and for the Italian, the
models from the MDZ Digital Library [21].
6
We retain a certain number of specific tokens, such as coordinating conjunctions between two clauses, relative
pronouns, certain adverbs that begin sentences or clauses, etc., trying to segment appropriately without over-
segmenting, for example, when two delimiter tokens appear consecutively (in the clause “et qui trop bien se defren-
doient”, we only retain the conjunction et).
70
respectively. The difference in the number of words required to produce convincing models is
due to the size of the initially used BERT models (the French BERT model is much larger as it
was trained on a much larger dataset, making it more efÏcient for the given task).
2.1.3. Results
The results of the best segmentation models are presented in Table 2, Table 3 and Table 4, one
for each language, according to classic measures of accuracy, precision, recall, and F1 score7 .
Each table displays the results of a model established using regex and a BERT-based model.
The results show a substantial improvement in the text segmentation task, thanks to the use
of a BERT-based model compared to the use of simple regexes.
The best model for each language is chosen on the basis of a weighted average between
precision and recall, on a test corpus that has never been used for training. Given the whole
workflow, recall is the important metric. It’s more important to identify as many true delimiters
as possible than to be precise in identifying delimiters. This is because the alignment phase
that follows enables the alignment units to be merged, thereby compensating for false positives.
Conversely, false negatives (failure to identify a delimiter) will not be compensated for later on,
and will lead to an alignment of poorer quality. In view of this, it was decided to produce a
weighted average between recall and precision for the selection of the best model, and to assign
a weight of 2 to recall, and 1 to precision. Accuracy is high because it computes the total number
of correct label assignments, both for labels 0 (non-delimiter tokens) and 1 (delimiter tokens),
across the texts.
It is the second label, “Delimiter”, corresponding to the results of identifying tokens labeled
1, that is relevant. Between the F1 scores of the regex and BERT-based models, we observe a
significant improvement, with an increase of over 0.20 for French, 0.24 for Italian, and 0.23 for
Spanish. Our models thus successfully segment the text according to a syntactic (and semantic)
logic.8
Table 2
French model results of the segmentation models
Regexp BERT-based
Accuracy 0.954 0.984
None Delimiter None Delimiter
Precision 0.940 0.812 0.990 0.874
Recall 0.978 0.611 0.978 0.941
F1-score 0.959 0.670 0.984 0.906
We can see a short example of the segmentation produced by both methods:
BERT-based: et bien paroit a ses iauz [SEP] qu’ele avoit rouges et anflez [SEP] qu’ele eust
ploré
7
The evaluation was performed on test sets consisting of 4,300 words in Spanish, 1,340 words in French and 2,500
words in Italian.
8
Due to the small size of the training corpus, it was decided to produce adhoc models for each of the languages in
the corpus. A multilingual model with language metadata injection experiments will be produced and evaluated
in an article dedicated to segmentation. These results provide a baseline for future in-depth studies.
71
Table 3
Italian model results
Regexp BERT-based
Accuracy 0.961 0.983
None Delimiter None Delimiter
Precision 0.937 0.711 0.981 0.827
Recall 0.971 0.523 0.976 0.866
F1-score 0.954 0.602 0.978 0.846
Table 4
Castilian model results
Regexp BERT-based
Accuracy 0.951 0.981
None Delimiter None Delimiter
Precision 0.942 0.688 0.981 0.869
Recall 0.962 0.584 0.982 0.863
F1-score 0.952 0.632 0.981 0.866
Regex: et bien paroit a ses iauz qu’ele avoit rouges [SEP] et anflez qu’ele eust ploré
The regex segmentation cut the sentence once, in the middle of a phrase (“rouges et an-
flez” that is “red and swollen”), whereas the BERT-based segmentation correctly segments the
sentence twice, in accordance with the grammatical structure, based on relative pronoun and
subordinative conjunction (qu’). The method chosen also has an influence on segment size,
which is not shown in the evaluation above, and which has an impact on the quality of the
alignments produced, in favor of the BERT-based method, as we’ll see below.
2.2. Alignment
Once the BERT-based models are trained and selected, they are integrated into the alignment
workflow, producing text segmentation based on the recognition of the good delimiters, before
the actual alignment phase. As stated above, Bertalign is used to perform the alignments9 . The
alignment is carried out on the basis of a main/pivot witness chosen in advance on the basis
of philological knowledge, that is, Micha [17]. This pivot witness is compared with each of the
others, and the alignments are merged. To create the merged alignments table we make use of
a graph method to connect each pair of aligned segment and create the final alignment unit,
while conserving the 1 to many and many to one alignments. Considering each aligned pair as
connected nodes in a graph, it is possible to build up complete alignment units by connecting
all nodes together thanks to the pivot fragment10 (figures 1 and 2).
9
The parameters we used are the following: max-align=3, window=5, skip=-.2, margin=True,
len_penality=True.
10
As with all methods using a pivot witness, alignment units in which the pivot omits text (omission or innovation
of other branches) must be handled subsequently. Indeed, if the base witness is omitted, the link between the
other witnesses is unknown. They must therefore be re-injected, which would require to design a second round
of alignment that would change the base witness on the omitted parts. This work has yet to be been produced,
72
Segm1Pivot
Segm1Pivot Segm1Pivot
Segm1WitB
Segm4WitC Segm1WitD
Figure 1: Getting pairs of aligned fragments
Segm4WitC Segm1WitD
Segm1Pivot
Segm1WitB
Figure 2: Merging the pairs into a single alignment unit by connecting all the nodes
2.2.1. Alignment evaluation
To assess the impact of the chosen segmentation method on alignment, we conducted an eval-
uation of the alignment produced by each of the two methods (regex and BERT-based segmen-
tation).
Description of alignment results The evaluation is based on correcting the alignments ob-
tained (correcting the alignment table of indices). Given that segmentation varies between dif-
ferent segmenters, the correction process must be repeated for results from both BERT-based
segmentation and regex segmentation. Table 5 shows some alignment results with the frag-
ment iv-75-1 (see also Appendix D, Table 11).
These results present the same alignment error in Lancelloto and BnF fr. 111, but due to two
different causes: in the case of the Lancelloto, the error can be attributed to the segmentation
phase, where the segmenter saw a single unit and did not separate between “consiglio” and
“chi”. In the case of BnF fr. 111, segmentation was done properly, but the error was produced
in the alignment phase, where the two units were regrouped.
Evaluation Alignments are evaluated on the basis of the Alignment Error Rate (AER) [25,
2.6, p. 21]:
2 × |𝐿 ∩ 𝐿gold |
𝐴𝐸𝑅 = 1 −
|𝐿| + |𝐿gold |
but it concerns a small number of alignment units only.
73
Table 5
Example of alignment automatically produced, with regex segmentation on the segment iv-75-1 (small
extract). The meaning of the passage is “…that I would remedy it as I can”. We can seen an alignment
error with witness BnF fr. 111.
micha lancellotto fr333 inc fr111 fr751 sommer lanzarote
que j’i e tutto il que ie i que ie y met- car ie y et uolentiers que ie y me- que os
metroie buon con- metroie troie tout le mectroye i metroie teroie volen- porne todo
volentiers siglio ch’i uolenters meilleur con- uoulentiers tout le bon tiers tout le el mejor
tout le bon potrei tot le bon seil tout le con- conseil bon conseil consejo
conseil conseil seil|que ie
pourroye
que je por- que ie por- que ie pour- que ie i por- que iou que yo
roie roie roie roie mestre y porroie pudiere
mettre
Where L and 𝐿gold consist of predicted and hand-corrected links. The evaluation is conducted
on indices by comparing automatically generated indices with corrected ones. As the align-
ment is produced for each witness compared to the pivot witness, an AER is produced for
each pair, and Table 6 presents the average of those in order to have an overall value for each
evaluated segment of text.
We distinguish between two types of results: results that consider incorrect alignments due
to segmentation issues, and those that do not. The evaluation is performed on different parts
of the text, each with varying numbers of alignment units.
Table 6
Alignment Error Rate with LaBSE, excluding or including segments that were not properly segmented.
The numbers of evaluated segments are in the ”Nb of segm.” columns. Numbers in bold are detailed
infra (improvement of the results with BERT-based segmentation for ii-48 and ii-61-1 sections but not
for the iv-75-1 one).
fragment Regex-segm. Nb of segm. Bert-segm. Nb of segm.
well segmented all well segmented all
ii-48 0.13 0.16 71 0.11 0.15 112
ii-61-1 0.30 0.38 136 0.26 0.32 152
ii-61-2 0.15 0.18 101
iv-75-1 0.24 0.33 120 0.27 0.32 169
iv-75-2 0.25 0.26 102
We observe a systematic improvement between the results that consider erroneous segmen-
tation and those that do not, both for regex segmentation and BERT-based segmentation. Sev-
eral sections of text were evaluated using both methods, revealing a slight overall improvement
in results with BERT-based segmentation.
It is important to note that the quality of alignment varies depending on the textual proximity
among versions in each witness. For example, there is a gap (0.15) between the results for
sections ii-48 and ii-61-1, which can be attributed to varying lacunas and omissions observed
in several witnesses for the latter.11
11
For example, Lancellotto doesn’t present the text for the beginning of Agloval episode, that leads to a poor align-
ment, because around 20 alignment units are not present in this witness.
74
Moreover, the poor results observed for the iv-75-1 section (better results are observed for
regex segmentation than for BERT-based segmentation) can be attributed to numerous in-
stances of missing correspondence among several witnesses, reflecting divergent traditions
with frequent omissions. In this context, regex segmentation, generating fewer segments (669
compared to 797 in this case), yields better alignment. This is because these segments consol-
idate more units, potentially accommodating omissions that might affect other segments. In
contrast, BERT-based segmentation, which offers finer granularity, tends to make more errors.
There is indeed a notable difference in alignment results due to the size of the produced data:
regex segmentation generates less segments, i.e., aligned units, as illustrated in Figure 3.
Figure 3: Number of tokens per segment (aligned units, left) and unit (small units automatically pro-
duced by the segmenter, right). BERT-based segmentation produces units with a more stable number
of tokens and thus yields more semantically coherent aligned segments.
BERT-based segmentation results in a significantly higher number of segments containing 3
to 14 tokens (even until 21 tokens), which correlates with the higher total number of segments
obtained with this method. Conversely, regex segmentation produces more segments with a
higher number of tokens (more than 30), as well as segments with one or two token(s). These
segments are either too large to facilitate accurate alignment or too small to maintain semantic
coherence. Thus, alignment tends to be of better quality with the BERT-based method for
stemmatologic purposes.
Correlatively, the size of the units (the units produced by the segmentation) within the seg-
ments changes, as we can see from the figure 3. With the regex segmentation, there is a high
number of very small units containing only one, two, or three tokens. In contrast, there are
more units containing 5 to 8 tokens with BERT-based segmentation. As the units in regex seg-
mentation are very short, they are grouped easily into segments. If this facilitates better results
for the highly variable parts of text we described above, this approach results in a generally
poor alignment, whereas BERT-based segmentation produces fewer segments with more se-
mantically consistent units. If the results from the AER are not very promising, we can assume
that the BERT-based segmentation produces a more coherent segmentation that positively im-
pacts the alignment task. It is important to note that AER is a good metric for evaluating the
quality of an alignment, but that it is not sufÏcient. Indeed, a text can be split into only three
75
parts and have a low AER, if the splits are correct, whereas this kind of result would not allow
for any meaningful exploitation of the data. On the contrary, the evaluation of the distribution
of units and segments lengths shows that BERT-based segmentation provides more coherent
elements to align.
2.3. Classification in variants
2.3.1. Method
Most stemmatological and phylogenetic algorithms needs features (i.e. variants) that are arbi-
trarily coded (usually, using letters or numbers), e.g. variant 1, variant 2, etc. [7] In our case,
this necessitates, for each alignment unit, to cluster the readings from every witnesses into sets
of relevant clusters (Table 7). The number of these clusters can vary from 1 (all witnesses have
the same variant) to a maximum number equal to the total number of witnesses (each witness
bears a different variant).
Table 7
Table showing an aligned portion of text for each witness, and its numerically coded classification in
variants by both annotators (here in agreement) and a clustering algorithm. The main opposition here
is between “heard” (oy, udito,…) and “done” (fait), and an omission. An OCR mistake (“or” for “oï ”) in
part prevents a correct clustering by the algorithm.
Witness Aligned text Ann. 1 Ann. 2 DBScan
Micha qu’il avoit or 1 1 4
Sommer quil auoit oy 1 1 1
fr751 quil auoit fait: 2 2 2
fr111 quil auoit ouy 1 1 1
inc 0 0 0
lanzarote 0 0 0
lancellotto ch’elli avea udito 1 1 3
As such, this task is an unsupervised clustering task, where the number of desired clusters
cannot be known in advance, which excludes most clustering methods (such as, for instance,
k-means or k-medoids). For this reason, we choose an unsupervised density-based clustering
method, DBScan (Density-based spatial clustering of applications with noise [11]), that we
apply to the cosine distances between each witness, for each alignment unit, in the embedding.
DBScan tries to find dense clusters of points, i.e., points with a given minimum number of
neighbours (MinPts) situated at a given maximum distance (𝜖). If DBscan does not necessi-
tate to set the number of clusters, it requires the setting of the two parameters MinPts and 𝜖.
These parameters can only be chosen contextually, in relation to the number of individuals
(witnesses) and the distances between them. For the MinPts, we set it at 1, given the small
number of witnesses. For 𝜖, we follow a methodology called DMDBScan, that tries to identify
relevant values, by first computing the minimum distances to the MinPts nearest points, and
then plotting them by ascending order to identify values for at which there are sudden sharp
changes in the minimum distances, possibly revealing different densities [10]. Following this
methodology, we set 𝜖 at 0.2 (Appendix F).
76
Table 8
Mean Adjusted Rand Index
Ann. 1 Ann. 2 DBScan
Annotator 1 1
Annotator 2 0.70 1
DBscan 0.04 0.04 1
DBScan is applied to the cosine distances between the readings of the witnesses in the em-
bedding. Yet, due to the multilingualism of the corpus, and the important spelling variation
characteristic of medieval language, some distances are artificially increased not for seman-
tic reasons but because of formal bias. To correct for this bias, a weighting is applied to all
distances prior to performing clustering, according to:
dist(𝑎𝑖 , 𝑏𝑖 )
weightedDist(𝑎𝑖 , 𝑏𝑖 ) = 1
2
(𝜇(dist(𝑎, 𝑛)) + 𝜇(dist(𝑏, 𝑛)))
where 𝑎 and 𝑏 are two witnesses, and 𝑖 a given alignment unit, and where 𝜇(dist(𝑎, 𝑛)) is the
arithmetic mean of the distance between 𝑎 and all other witnesses for all segmentation units.
2.3.2. Results and evaluation
In order to evaluate the results of clustering, we compare the clustering performed by DBScan
to that of the two human annotators (the first two authors of this paper), and compute the mean
Adjusted Rand Index (ARI), whose value is contained between -1 and +1, where -1 denotes a
complete opposition, 1 a complete agreement, and 0 a case where both clustering appear to
have been independently randomly labelled.
We compare the mean ARI between the two human annotators and the DBScan result. The
results are presented in Table 8. The score is very low for the DBscan results, especially com-
pared to the correspondance between the human annotators. Yet, it does not necessarily mean
that the clustering results does not yield significant genealogical information to be processed
by the stemmatological algorithms, as will be seen in the next section.
3. Results of the stemmatological analysis
The alignment tables we obtained through the BERT-based segmentation and the alignment
phase enable philological analysis. It was performed in two distinct ways: first, a traditional
stemmatic analysis, based on human expertise, was performed. Then, the variants resulting
from the human annotations and the DBScan clustering were subjected to stemmatological
algorithms, and the results were compared.
3.1. Human-identified witness groups
We decided first to evaluate the links between the witnesses, based on the alignment tables
and human close reading expertise (ours, and that of previous philologists having worked on
77
this topic). Each one of the witnesses we chose presents particular readings (subsection E.1).
However, we can distinguish two main groups within the tradition. On one hand, the witnesses
Micha, BnF fr. 751, BnF fr. 111 and Lancellotto, and on, the other hand, Sommer, the incunable,
and Lanzarote. They can be determined on the basis of omissions or distinct reading groups
(subsection E.2). The oppositions between the two groups are quite frequent and stable.
Nevertheless, Lanzarote has common readings with the first group and sometimes oscillates
between the readings of the two groups (subsection E.3). In fact, the text in the Lanzarote
presents a quite original version, especially due to the shortened passages it contains (subsec-
tion 1.2). It also often offers specific readings and innovations, particularly when the translation
presents cases of direct speech instead of basic narration (subsection E.4).
The text of the Lancellotto doesn’t show such originality. On the contrary, it is very close to
the French text, using many gallicisms [5, p. 43-44]. It also stays very close to its own group.
It shares particular common readings with BnF fr. 111, as well as several omissions. How-
ever, sometimes, Lancellotto knows a common reading with Micha against BnF fr. 111 (subsec-
tion E.5). The Lancellotto also shows some innovations that prove that the very model of the
translation is one not part of our selection (subsection E.6).
We can assume that the witnesses we studied belong to two main groups, and that the wit-
nesses that interest us, the Lanzarote and the Lancellotto, each fall into one of these groups.
However, the specific variants and innovations found in each witness do not allow the identi-
fication of a specific model or the determination of a more precise afÏliation.
The identification of groups confirms the filiation between BnF fr. 111 and Lancellotto [5],
but does not confirm the specific link between BnF fr. 751 and Lanzarote. It is important to
note that BnF fr. 751 is really unstable. Indeed, if it shares most of its readings with the first
group, in the variants we studied, it also has its own unique readings.
3.2. Results of the stemmatological algorithm
The second approach applies a well known phylogenetic algorithms to the table of variants
produced by the two human annotators, and the one produced by the DBScan clustering. The
chosen algorithm is the Branch and bound algorithm, whose goal is to determine minimal
evolutionary trees by the criterion of maximum parsimony [12]. This algorithm can produce
one or several trees, if different configurations achieve the same maximum parsimony value.
The trees resulting from the analyses are presented in Figure 4.
The trees resulting from the tables produced by the human annotators present, unsurpris-
ingly, strong similarities with the result of the human close reading expertise on the genealogy.
In particular, they all display the groupings {Lanzarote, Sommer, Incunabula} versus {Lancel-
loto, Micha, fr. 751 and fr. 111}. The second and third alternative trees from Annot. 2 also show
the grouping of Lancelloto and fr. 111, already observed by the editor of the Lancelloto [6], but
only the third tree from Annot. 2 shows the stronger similarity between fr. 751 and Lanzarote,
that was hypothesised by its editor [9].
In comparison, the results based on the DBScan do not show fully the opposition {Lanzarote,
Sommer, Incunabula} versus {Lancelloto, Micha, fr. 751 and fr. 111}, because Sommer switches
families. Yet, it provides results more consistent with the analyses of the editors of the Lan-
zarote, that is positioned close to the fr. 751 [9], while Lancelloto remains relatively close to
78
Figure 4: Results of the Branch and bound algorithm on the variant tables produced by the two human
annotators and the DBSsan clustering The parsimony values are, for the first annotator, 104; for the
second, 135; for the DBScan results, 292.
79
fr. 111.
If the results obtained from the automatic clustering of variants differ from the expertise of
the human annotators, on the other hand the genealogical results do not necessarily contradict
previous expertise, and seem to remain significant up to a point. Deciding whether the human
annotators or the clustering have provided the most useful data remains to be fully assessed, but
current results indicate that, even if the clustering differs in terms of classification of readings
from human annotation (as shown by the mean Adjusted Rand Index), the results obtained
through stemmatological analysis still remain indicative.
4. Conclusion
We’ve designed a tool that can facilitate the study of unilingual and multilingual variant tradi-
tions, and its stemmatological analysis. The results obtained through clustering and stemma-
tological analysis tend to show that an automatic clustering of variants and the application of
a phylogenetic algorithm can quickly produce results that are indicative of the genealogy of
witnesses, and comparable to some extent to a traditional human expertise.
Yet, in the current state, many aspects of the processing workflow could still be improved,
especially if the results of the automatic alignement and collation were to be used for scholarly
editing purposes.
As far as segmentation is concerned, the delimiter tokens are those that reliably identify co-
herent grammatical segments (relative pronouns, coordinating conjunctions between clauses,
and subordinating conjunctions, etc.). This is a problem for all sentences that begin with per-
sonal pronouns, articles, or nouns and not with the markers listed above. Therefore, we need
to produce more segmented data to be able to distinguish all these complex cases.
As for the alignment step, we show that LaBSE’s sentence embeddings model works sur-
prisingly well on medieval language states. Still, there is room for improvement and some
fine-tuning of LaBSE model on aligned medieval texts should improve the results. Variant
clustering should also benefit from this fine-tuning. However, it will be a complex and time-
consuming task, because of the required data, that is aligned texts: this step is planned for the
future. Another important step will be microscopic alignment, which involves further work
on modeling the alignment process, because it differs quite significantly from the alignment
of monolingual texts, particularly with regard to the variations in the word order in the tar-
get sentence and in the source sentence. The micro-alignment could allow also more precise
results in the stemmatological step.
Materials and code availability
The tool for alignment and collation is freely available on Zenodo:10.5281/zenodo.12732533
and is maintained on github: https://github.com/ProMeText/Aquilign/. The result data can be
found in the results_dir directory. The segmenter train corpus is available in the subfolder:
data/tokenisation. The data and scripts for clustering and stemmatological analysis are
available on a specific Zenodo repository: 10.5281/zenodo.12728282.
80
Acknowledgments
Co-funded by the Agence Nationale de la Recherche under the Équipex Biblissima+ (ANR, 21-
ESRE-0005).
Co-funded by the European Union (ERC, LostMA, 101117408). Views and opin-
ions expressed are however those of the author(s) only and do not necessarily reflect those of
the European Union or the European Research Council. Neither the European Union nor the
granting authority can be held responsible for them.
The authors additionally wish to thank Florian Cafiero for his thoughts on this research.
References
[1] M. Artetxe and H. Schwenk. “Massively Multilingual Sentence Embeddings for Zero-
Shot Cross-Lingual Transfer and Beyond”. In: Transactions of the Association for Compu-
tational Linguistics 7 (2019), pp. 597–610. doi: 10.1162/tacl\_a\_00288.
[2] L. Bambaci, F. Boschetti, and R. Del Gratta. “Qohelet Euporia: A Domain Specific Lan-
guage to Annotate Multilingual Variant Readings”. In: 2018 IEEE 5th International Congress
on Information Science and Technology (CiSt). 2018, pp. 266–269. doi: 10.1109/cist.2018.8
596332.
[3] D. J. Birnbaum and H. M. Eckhoff. “Machine-Assisted Multilingual Alignment of the Old
Church Slavonic Codex Suprasliensis”. In: V Zeleni Drželi Zeleni Breg. Ed. by S. M. Dickey
and M. R. Lauersdorf. Indiana: Bloomington, 2018, pp. 1–15. url: https://www.research
gate.net/profile/Nada-Sabec/publication/371120502%5C%5FToward%5C%5FLess%5C%5
FFormal%5C%5FWays%5C%5Fof%5C%5FAddressing%5C%5Fthe%5C%5FOther%5C%5Fi
n%5C%5FSlovene/links/6473967b59d5ad5f9c803dac/Toward-Less-Formal-Ways-of-Ad
dressing-the-Other-in-Slovene.pdf.
[4] L. Brun. Arlima - Archives de littérature du Moyen Âge. Ottawa, 2005. url: https://www
.arlima.net/.
[5] L. Cadioli, ed. Lancellotto: versione italiana inedita del Lancelot en prose. Firenze: Edizioni
del Galluzzo, 2016.
[6] L. Cadioli. “Le Lancellotto italien”. In: La matière arthurienne tardive en Europe, 1270-
1530 = Late Arthurian Tradition in Europe. Rennes: Presses universitaires de Rennes, 2020,
pp. 501–510.
[7] F. Cafiero and J.-B. Camps. “Stemmatology: An R Package for the Computer-Assisted
Analysis of Textual Traditions”. In: Proceedings of the Second Workshop on Corpus-Based
Research in the Humanities CRH-2, 25-26 January 2018, Vienna, Austria, ed. Andrew U.
Frank; Christine Ivanovic; Francesco Mambrini; Marco Passarotti; Caroline Sporleder (2018).
[8] J. Cañete, G. Chaperon, R. Fuentes, J.-H. Ho, H. Kang, and J. Pérez. “Spanish Pre-Trained
BERT Model and Evaluation Data”. In: PML4DC at ICLR 2020. 2020, pp. 1–9. url: https:
//arxiv.org/abs/2308.02976.
81
[9] A. Contreras Martı́n and H. L. Sharrer, eds. Lanzarote del Lago. Alcalá de Henares: Centro
de estudios cervantinos, 2006.
[10] M. T. Elbatta and W. M. Ashour. “A dynamic method for discovering density varied clus-
ters”. In: International Journal of Signal Processing, Image Processing and Pattern Recogni-
tion 6.1.1 (2013), pp. 123–134.
[11] M. Ester, H.-P. Kriegel, J. Sander, X. Xu, et al. “A density-based algorithm for discovering
clusters in large spatial databases with noise”. In: kdd. Vol. 96-34. 1996, pp. 226–231.
[12] M. D. Hendy and D. Penny. “Branch and bound algorithms to determine minimal evolu-
tionary trees”. In: Mathematical biosciences 59.2.2 (1982), pp. 277–290.
[13] Irht. Jonas: Répertoire des textes et des manuscrits médiévaux d’oc et d’oı̈l. Paris et Orléans,
2016. url: http://jonas.irht.cnrs.fr/.
[14] O. Kraif. Adaptative Bilingual Aligning Using Multilingual Sentence Embedding. preprint.
arXiv, 2024. url: http://arxiv.org/abs/2403.11921.
[15] L. Liu and M. Zhu. “Bertalign: Improved Word Embedding-Based Sentence Alignment
for Chinese–English Parallel Corpora of Literary Texts”. In: Digital Scholarship in the
Humanities 38.2.2 (2023), pp. 621–634. doi: 10.1093/llc/fqac089.
[16] C. Meinecke, D. J. Wrisley, and S. Jänicke. Automated Alignment of Medieval Text Versions
Based on Word Embeddings. preprint. Open Science Framework, 2020. doi: 10.31219/osf
.io/tah3y.
[17] A. Micha, ed. Lancelot: roman en prose du XIIIe siècle. Tome II. Genève: Droz, 1978.
[18] A. Micha, ed. Lancelot: roman en prose du XIIIe siècle. Tome IV. Genève: Droz, 1979.
[19] M. Sánchez Sánchez and C. Domı́nguez Cintas. “El banco de datos de la RAE: CREA y
CORDE”. In: Per Abbat: boletı́n filológico de actualización académica y didáctica 2 (2007),
pp. 137–148. url: https://dialnet.unirioja.es/servlet/articulo?codigo=2210249.
[20] G. Scala. “La Tradizione Manoscritta Del ”Livre Du Gouvernement Des Roys et Des
Princes” Di Henri de Gauchy. Studio Filologico e Saggio Di Edizione”. PhD thesis. Uni-
versity of Zurich, 2021. doi: 10.5167/uzh-202620.
[21] S. Schweter. Europeana BERT and ELECTRA models. 2020. doi: 10.5281/zenodo.4275044.
[22] H. O. Sommer, ed. The Vulgate Version of the Arthurian Romances. Vol. 4, Le Livre de
Lancelot del Lac, part II. Washington: Carnegie Institution of Washington, 1911.
[23] H. O. Sommer, ed. The Vulgate Version of the Arthurian Romances. Vol. 5, Le Livre de
Lancelot del Lac, part III. Washington: Carnegie Institution of Washington, 1912.
[24] F. Steven R. “The Complete Medieval Dreambook. A Multilingual, Alphabetical Somnia
Danielis Collation”. PhD thesis. University of Michigan, 1978.
[25] J. Tiedemann. Bitext Alignment. 1st ed. Synthesis Lectures on Human Language Tech-
nologies. Morgan & Claypool Publishers, 2011. url: http://gen.lib.rus.ec/book/index.ph
p?md5=7fb8c6d1d4e8924f79a03026e23b6517.
82
[26] J. Tronch. “Displaying Textual and Translational Variants in a Hypertextual and Multilin-
gual Edition of Shakespeare’s Multi-text Plays”. In: Early Modern Studies after the Digital
Turn. Ed. by L. Estill, D. K. Jakacki, and M. Ullyot. Medieval and Renaissance Texts and
Studies 502. Toronto, Ontario: Iter Press, 2016, pp. 92–116.
[27] D. Witschard, I. Jusufi, R. M. Martins, K. Kucher, and A. Kerren. “Interactive Optimiza-
tion of Embedding-Based Text Similarity Calculations”. In: Information Visualization 21.4
(2022). doi: 10.1177/14738716221114372.
[28] T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R.
Louf, M. Funtowicz, J. Davison, S. Shleifer, P. v. Platen, C. Ma, Y. Jernite, J. Plu, C. Xu,
T. L. Scao, S. Gugger, M. Drame, Q. Lhoest, and A. M. Rush. “Transformers: State-of-the-
Art Natural Language Processing”. In: Proceedings of the 2020 Conference on Empirical
Methods in Natural Language Processing: System Demonstrations. Online: Association for
Computational Linguistics, 2020, pp. 38–45. url: https://www.aclweb.org/anthology/20
20.emnlp-demos.6.
Appendices
A. Witnesses
• lancellotto: Lancellotto, italian witness, edited by Luca Cadioli (cf. bibliography).
Firenze, Biblioteca della Fondazione Ezio Franceschini, 1 (last quarter of the 14th century).
• lanzarote: Lanzarote, castillan witness, edited by Antonio Contreras Martín and Harvey
L. Sharrer (cf. bibliography) .
Madrid, Bibioteca Nacional de España, 9611 (16th , copy of a manuscript dated from 1414).
• sommer: edition of H. Oscar Sommer (cf. bibliography).
Tomes IV and V: London, British Library, Additional 10293 (beginnings of the 14th cen-
tury).
• micha: edition of Alexandre Micha (cf. bibliography).
Tome II: Cambridge, Corpus Christi College Library, 45 (2nd half of the 13th century).
Tome IV: Oxford, Bodleian Library, Rawlinson D. 899 (14th century).
• fr111:
BnF, français 111 (15th century).
• fr333:
BnF, français 333 (first quarter of the 14th ).
• fr751:
BnF, français 751 (2nd part of the 13th century).
• inc:
BnF, RES-Y2-46 (vol. 1) and RES-Y2-47 (vol. 2) (Printed in Rouen [vol. 1, by Jean le
Bourgois] and Paris [vol. 2, by Jean Dupré] in 1488).
83
B. Witnesses and segments of text
The Lancellotto is particularly fragmentary. Only the sections of the text corresponding to
these fragments were studied.
According to Micha, its corresponds to the sections: II, XLVIII 29 – L 11; II, LXI 26 - LXIX 23;
IV, LXX 1 - LXXXI 17. But, for the last, the Lanzarote omitted the text and starts at IV, LXXV,
with a very reduced portion in LXXX 16 to 43. Thus, the studied section of text starts at IV,
LXXV.
We indicated here the witnesses and the folios where the segments of the text can be found.
Lancellotto Lanzarote Sommer Micha proper id status
name
f. 222r-226v f. 203r-218v IV, 271-285 II, XLVIII 29 suite ii-48 full
– L 11 Char-
rette
I
f. 243r- f. 265r- IV, 327-362 II, LXI suite ii-61 without the
254vb 286v to 26-LXIX 23 Char- end in the
LXVIII, with rette Lanzarote
reduced 8-12 II
f. 254vb- f. 286v-320v V, 3-143 IV, LXX part of iv-75 without the
297v IV, LXXV- 1-LXXXI 17 Agravain beginning
LXXX, with and the
LXXX 16 end in the
à 43 very Lanzarote
reduced
These passages correspond, in the other French witnesses, to:
BnF fr. 751 BnF fr. 333 BnF fr. 111 incunable id
227va-232va 138vb-141vb Diii.v-Dviii.r ii-48
248va-259va 151rb-157vb Fiiii.r-Gv.v ii-61
270vb-291rb 23rb-47va 170va-183va Iviii.v- end vol. 1 iv-75
(Liiii.v) + vol. 2 be-
ginning to Bv.v
We can note that, as we indicated in the introduction, the fr. 333 only contains the third
studied segment and that the text corresponding to the incunabula is contained by two different
volumes.
84
C. Example of missed section
As example of missed sections of text that can have an impact on the alignment process, we
can take the fight in the episode of the hawk (corresponding to the segment ii-61-2).
In Micha’s edition, the text is the following (the whole episode is given by all the other
studied witnesses):
… tant qu’il vienent en une valee. Et lors li mostre un poi en sus del chemin a senestre la loge dont
li chevaliers issi, dont ele se plaint. « Venés avant, fet il, seurement, et se vos veez l’esprevier, si le
prenés, ja por nului nel laissiés. Et je vos creant loialment que je le vos garantirai a mon pooir contre
tos cels qui le voldront contredire. Et se li espreviers n’i est, si me mostrez le chevalier qui le vos toli
et jel vos ferai amender tot a vostre volenté. — Sire, fet ele, de Dieu aiés vos bone aventure ! Mais
je voldroie miels que vos le me poissiés rendre a pes que a guerre. — Par Dieu, fet il, se ie ne le puis
avoir par debonaireté, si l’avrai je par force. » Lors sont venu a la loge ; si entre mesire Yvain tos
premiers et la damoisele aprés. Et mesire Yvain ne salue nul de cels de laiens, ains dist si haut que
tuit le porent oïr : « Damoisele, venés avant et si prenés vostre esprevier, se vos saiens le poez veoir ;
si l’enportez ausi a droit com il en fu portez a tort. — Sire, fet ele, volentiers, ausi le voi je la. » Et
ele vient a une perche ou il seoit, si li deslie les giés let l’en volt porter, quant uns chevaliers saut
avant qui li dist : « Fuiés. damoisele ! Ne le remués, que par mon chief vos n’en porterés point. Et
de tant com vos i estes retornee, avés vos del tot vos pas perdus, que vos ne l’enporterés n’en l’une
main n’en l’autre ; et se vos volés avoir oisel, si querés autre, kar a cestui ne vos deduirés vos jamés.
— Laissiés li, dans chevaliers, fet mesire Yvain, qu’ele l’enportera, et se vos li volés fere force, vos en
serés tart al repentir. — Comment ? fet cil. Estes vos ci venus por le deffendre ? — Ce verrois vos
bien, fet mesire Yvain, se vos ti tolés. » Et cil giete maintenant le main por le tolir ; et mesire Yvain
a trait l’espee et li dist qu’il li coupera le bras, s’il toche plus ne a lui ne a la damoisele. « Voire, fet
cil, par mon chief mal le deïstes ! » Lors cort a son hialme et le met en sa teste, et il estoit molt bien
armés de tote autre armeure. Maintenant saut en son cheval, kar tos estoit prest, et prent son escu
et son glaive et dist a mon seignor Yvain qu’il se gart de lui ; si li laisse corre, son glaive aloigné et
mesire Yvain a lui ; si s’entredonent si grans cops sor les escus qu’il les font fendre et percier et les
haubers desmaillier et derompre ; si se metent es chars nues les glaives trenchans, si s’entrehurtent
des escus et des cors et des visages, si s’entreportent a terre tot enferré. Mesire Yvain est navrés el
costé destre et li chevaliers fu ferus par mi le cors si durement qu’il n’a pooir de soi relever de la ou
il gist. Et mesire Yvain se redrece a tot le tronçon qui demi li est el costé ; si trait l’espee et s’apareille
d’assaillir le chevalier qui le meillor cop li a doné qu’il receust pieça ; et il le cuide tot prest trover
de deffendre, si voitqu’il ne se remue, et lors li cort sus et li esrache le hialme de la teste et dist qu’il
li coupera le chief sans arest, s’il ne se tient por outré. Et cil parole a grant paine com cil qui molt
estoit bleciés, si crie merci et dist : « Ha, frans chevaliers, ne m’ocie mie, mais laisse moi vivretant
que j’aie mon Salveor receu, kar je sai bien que je sui navrés a mort ; si vos pri por Dieu que vos alés
ci pres desus cest tertre querre un saint home mire qui i maint, et li faites avec lui aporter a corpus
Domini. Et il dist que si fera il volentiers ; si commande la damoisele qu’ele s’en aut et ele le fet. Mais
ele fet assés greignor duel que devant, kar ele voit .I. chevalier ocis et .I. autre navré, et por si petit
d’acheison. Et mesire Yvain vet querre l’ermite, ensi com li chevaliers li ot dit, et li amaine. Et quant
il fu revenus arrieres, si trueve iluec .I. escuier et une damoisele qui estoit amie al chevalier et faisoit
le greignor duel del monde. Et quant li chevaliers fu confés et il ot receu son Salveor, si le coucha
l’en en la loge. Et mesire Yvain s’en vet avec l’ermite et enmaine son cheval en destre, kar a cheval
n’i alast il mie delés si haut saintuaire comme Nostre Salveor. Quant il furent venu a l’ermitage que
l’en apeloit l’ermitage del Mont, si desarment troi frere qui laiens estoient mon seignor Yvain, et il
en i avoit un qui molt savoit de plaies garir; si s’entremist de mon seignor Yvain et s’en prist garde
erraument et il osta le tronçon qu’il avoit el costé et s’estanchaé a sainier ; et de cele plaie demora
mesire Yvain .XV. jors laiens.
In the Lanzarote, the episode is really reduced, since it presents only the following text:
85
… y andubieron tanto que llegaron ala Ramada ado estaua el cauallero y la donzella se lo mostro don
yban quando le vio dixole señor cauallero yo vos Ruego que dedes su gauilan aesta donzella quele
tomastes o sino enla vatalla sodes el gauilan no se lo dare yo dixo el Cauallero mas dela batalla presto
so e luego se dexaron correr el vno contrael otro y el cauallero quebro su lança en don yban y don
yban lo firio tan de Reçio quelo derribo en tierra todo atordido que no se pudo leuantar e don yban
desçendio porle cortar la cabeza mas el le Rogo quele perdonase y que faria todo su mandado y don
yban lo dexo y el Cauallero selo agradesçio mucho y dio luego el gauilan ala donzella y ella se fue
muy alegre conel y don yban se fue ensu demanda
This example shows the need to define what a collatable example is. An abstract is not
really alignable with its source, and we need to go through another processing step and not try
alignment. This shows the need to produce solid textual relationship typologies before making
text processing decisions.
D. Example of alignment table
Table 11: Example of good quality alignment taken from ii-48 fragment (selected witnesses, no
correction performed), with the BERT-based segmentation. Segments are separated
with pipe “|” characters.
micha fr751 inc lanzarote lancellotto
mais il voit l’eve noire Mais il uoit leue mes il uoit Mas el agua ma e vede
et parfonde |et si peril- si parfonde et si la riuiere hera muy fonda l’acqua nera
luse perilleuse parfonde et e peligrosa e e profonda e
dangereuse a negra e bien perigliosa
passer. cuidaua morir
qu’il cuide bien noier, que il cuide bien et scait bien che crede bene
perir morire
s’il se met dedens; se il se met sil entre de- si se y Metiese s’elli|si mette di
dedans. dens|qu il se dentro;
met en peril de
mort.
et d’autre part il voit et dautre part il Et daultre part il e dela otra parte e d’altra parte e
cele uoit cele uoit celle veya la donçella vede colei
qui si durement crie qui si docemet li qui si piteuse- que muy afin- che|si dura-
merci; prie [mer]ci. ment fui crie cada mente le mente gli grida
mercy. pedia merçed mercé,
si l’em prent tels pitiés Si len prent tes Si luy en prent e ovo tal piedad si ne gli prende
pitiez telle pitie della tale piatà
qu’il en laisse totes quil en laisse quil en laisse que le fiço todo che ne lascia
poors totes paours. toute paour el miedo perder tutte paure
86
et fet le signe de la Si fait le signe et fait le signe e fiço la señal e si fa el segno
crois en mi son vis, de la crois enmi de la croix dela cruz sobre della santa
son uis. deuant son si croce nel milu-
uisaige ogo del suo
viso,
puis enbrace l’escu |et puis enbrace puys embrache e enbraço el poscia imbrac-
broche le cheval des lescu. |et son escu |et escudo |e firio cia lo scudo
esperons |et se fiert en broche le fiert son cheual al cauallo de |e broca il
l’eve. |Et li chevals cheual des des esperens las espuelas cavallo degli
fu fors,|si commence a esperons|et se |et se lanche |e lanço se sproni, |si fiede
noer fiert en leue tot dedens leaue en el agua nell’acqua is-
errantment. |Et tout erraument: e el cauallo nellamente. El
li cheuaus fu |et le cheual començo de cavallo fu forte,
fors|si comance commence a nadar luego incominciò a
a noer noer notare
si tost com il ot terre si tost com il ot si tost |quil eust e perdio tierra si tosto com’elli
perdue, terre perdue terre perdue. en tal manera ebbe terra per-
duta,
si l’enporte d’autre si lenporte et sen passe le que paso dela si ne l portò
part de la rive a dautre part la cheual iusques otra parte del d’altra parte
quelque paine, |mais riue a quelque de laultre part agua |mas ante della riva a
ançois ot beu de l’ève painne. de leaue |mes beuio el cauallo qualche pena,
li uns et li autre, ce fust a grant del agua |’ma inanzi
paine |car ains furono amen-
quilz feussent duni tutti
passes beurent molli,
de leaue lun et
laultre
et se li chevals ne fust et se le cheual e si el caballo e se l cavallo
si buens, |noié fuissent neust este tan bueno no non fosse
ambedui, fort et bon ilz fuera sin dubda stato si forte,
feussent noyes |el moriera enel anegati fossono
et lun et laultre agua amenduni,
kar li chevaliers estoit car le cheualier e boores |porel ché l cavaliere
pesans por les armes estoit fort pe- peso delas ar- era pesante per
sant pour les mas l’arme
armes
qu’il avoit vestues. quil auoit que llebaua ch’elli avea
uestues. vestite.
87
Quant il fu de l’autre Si ne descendi Quant il fut e desque fue Quand’e fu da
part de l’eve, |si ne de- pas de lautre part dela otra Parte l’altra parte
scendi pas, de leaue |si del agua dexo se dell’acqua, |si
ne descendist correr aquellos non discende
oncques que la donçella passo,
tenian
88
E. Readings and groups of witnesses
E.1. Specific readings
In some sections, every witness has its own reading (fragment ii-61-1):
micha sommer fr751 fr111 inc lanzarote lancellotto
Puis traist puis traist Puis trait puis trait apres trait e despues Poscia trae
hors de hors de sa la piece de hors de hors de tiro la fuori di
sa char quisse le lespee qui sa char la sa cuisse pieça dela sua carne
la piece piece de en sa char piece de lespee espada la spada e
de l’espee lespee estoit. lespee desu pierna l’apicca:
qui dedens
estoit
E.2. Examples of readings specific to two main groups
The two main groups can be established from common omissions, for example (fragment ii-48):
micha sommer fr751 fr111 inc lanzarote lancellotto
si le font de- Si le font si le font si l fanno
sarmer, ou desarmer desarmer disarmare
il volsist ou uossist ou uoulsist il o volesse O
non, et il le non. et il le ou non. et non, ed elli
fist a envis, fait mout il le fist a il fece ad
enuiz. peine. invidia,
We can also identify common variants that systematically oppose group 1 to group 2. In the
following table (fragment ii-48), the two variants oppose shield to weapons, castle to forest.
micha sommer fr751 fr111 inc lanzarote lancellotto
mes il ne mais il ne mais il ne mais il ne mes il ne mas no ma e’ non
portoit portoit portoit portoit pas portoit traia el porta mica
mie tel mie tels mie tel tel escu mye telles tales tale scudo
escu armes escu armes armas
Al chastel, En la Au chas- ou chastel En la for- enla flo- Al
fet ele, de forest fait tel fait ele de fleago est de flo- resta de castello,
Floego elle de de floego. fait ella rega: donseglo- diss’ella,
floregas. rega dixo di Fleego,
la donçella
89
E.3. Oscillation of Lanzarote between two groups
Lanzarote belongs to the second group, but presents sometimes readings from the first one, or
even a mixed reading, as we can see in the following table (fragment ii-48):
lanzarote micha sommer
e el fue alla por aluergar e et il torne cele part por her- si torne cele part pour her-
fallo ala puerta dos frailes bergier. Et quant il i vint, si bergier. Et quant il vint la
trueve a la porte IIII. des si trouua .ii. freres
freres
The first group, with Micha, is characterized by the mention of the porte and the presence of
four freres whereas the second one, the Sommer’s one, is characterized by the absence of the
mention of the porte and the presence of only two freres. The text in the Lanzarote presents
the mention of a porte and only two freres.
E.4. Example of innovation in Lanzarote
Lanzarote presents some specific innovations compared to the other witnesses in the following
table (fragment ii-48):
micha lanzarote
puis voit delés les chevaliers une e el asi Catando vio dentro enlos arcos
tombe, la plus riche vna muy Rica tumba
qui onques fust fete par home, kar qual nunca tal viera el ni ome del
ele ert tote d’or fin a chieres pierres mundo que hera toda de oro fino e
precioses qui molt valoient miels d’un de piedras preçiosas que estaua toda
grant roialme. Se la tumbe fu de grant labrada de diuersas maneras e de muy
bialté, nient ne monte la bialté envers muchas cosas de figuras e de otras
la richece dont ele estoit et avec ce es- cosas que tal tumba de tal manera
toit ele la greignor que Lancelos eust nunca viera ni ome del mundo que la
onques veue: non ouiese visto no podria asmar la fer-
mosura que enella auia de tales cosas
como enella estauan
si se merveille molt qui puet estre li e don lançarote se marauillo e dixo
princes entresi mesmo Quien podria ser el
prinçipe
E.5. Opposition between Lancellotto and Micha against BnF fr. 111
We can find a multitude of readings that oppose the group Lancellotto/Micha against BnF fr.
111 (fragment ii-61-1), such as the ones in the following table:
90
lancellotto micha fr111
se noi abiamo per miscre- se nos avons par mescre- se nous auons foloye par
denza fallato qua in adi- ance foloié ça en arieres mescreance
etro
… che trastutti cristinia- … que tuit se crestienoient, … que tous se
vano, si confessò, udendo si reconuit oiant tot le c[re]stienneroit. si cogneut
tutto el popolo, pueple deuant tout le peuple
Some variants are more important (fragment ii-48):
lancellotto micha fr111
e quelli iscioglie im- et cil se destrosse main- et cil baille sa uenoison a
mantanente sua caccia- tenant de sa venoison, porter a son compaiguon.
gione e la dona a suo si la baille a son com-
compagnone per portar- paignon por porter,
nela,
E.6. Example of innovation in Lancellotto
Lancellotto presents in some passages innovations (fragment ii-61-2):
micha lancellotto
et de ceu est il tos esbahis. Aprés re- e di ciò si maraviglia egli molto, si nè
garde la pucele, si se merveille plus as- molto isbigotito.
sés de sa bialté que del vaissel,
Ma di cosa che veggia non si mar-
aviglia egli niente inverso della
trasgran biltà della damigella che
tanto gli sembia esser bella ch’a pena
l’osa riguardare, ché più gli è aviso
che sua faccia risprenda che sole, e
suoi occhi rilucenti e sua capellatura
che gli sembia ad essere di fine oro; ve
di persona si, come dice in suo cuore,
kar onques mes ne vit il feme non ne vide mai niuna
91
F. DMDbscan
Figure 5: Minimum distances to the MinPts nearest points, sorted by ascending order. Two different
densities levels are possible, the first one breaking around 0.2 distance, and the other around 1.2.
92