This study focuses on the problem of multilingual medieval text alignment, which presents specific challenges, 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 proseLancelot, 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.
The production and transmission of written texts during Antiquity and the Middle Ages involved a process of manual copying. During this process, the text was progressively transformed by errors and innovations as it circulated, each copy introducing successive modifications. Since the 19th century, philologists have taken to study the transmission of texts, based on innovations, using the genealogical treest(emma codicum) as a metaphor to visually represent 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 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.
Multilingual collation of texts has attracted a few contributions, mostly focused on encoding or
displaying editions of widely circulated tex2ts4,[
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 collatsiuobnse(ction 2.2), at “sentence” level and onsimilar fragments, on the classification of aligned sequences into distinct variants s(ubsection 2.3), and on their subsequent stemmatological analysisse(ction 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 s(ubsection 2.1). The paper concentrates on the study of the multilingual tradition of thLeancelot 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.
The study is based on a few witnesses of a complex medieval tradition, that of tLhaencelot 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 onwards13[
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 ththec1e4ntury, for the
ItalianLancellotto (Firenze, Biblioteca della Fondazione Ezio Franceschini 1); in ath1c6entury
manuscript, copied from a 1414 exemplar according to its colophon, for the CastiLlaiannzarote
(Madrid, Biblioteca Nacional de España, 9611). Both texts have been edited5, [
Given the extremely large number of witnesses of the FrencLhancelot, 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. 11th1 (c1.)5
is supposed to be close to theLancellotto [
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 instability, and they are also fragmentaryL.ancellotto is especially so, as it presents only three nonconsecutive episodes of the text. To enable alignment, the first step was therefore to identify corresponding passages from one witness to anotherAp(pendix 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 editio2n.
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 Eckhof [
Recently, Liu and Zhu [
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 segmentation approach prior to the alignmen4t.We describe our processing chain: segmentation, pseudo-sentences alignment, classification into distinct variants and stemmatological analysis.
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 sentence 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. Punctuation 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 3A more recent preprint picks up on this work and compares Bertalign to a new architectu14r]e, [but we haven’t had time to look at how the algorithm works and how good it performs on our data. 4The 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. variability increases during the transcription and editing of the text. Moreover, original punctuation 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 between syntactic and semantic units, and that the syntactic tokenisation should be an efÏcient way to approach the diferent semantic units of the text. Let’s take an example (Table1): 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 diferent but share the same overall meaning. si ne demora gueres
2.1.1. Segmentation methods To segment the texts, we try, compare and evaluate two methods: the first uses regular expressions, 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. Regular 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: And in Italian: "[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" "[mM]a si", "[Ss]i", "tanto che", "che", "e", "donde", "ch'?"
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 variatioainns(i, insi, einsi for thus, for instance), it also identifies tokens that do not serve to delimit coherent syntactic units. Indeed, for example, the coordinating conjunctioent (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 diferent tokens, and thus diferentiating
the uses of the same words. It was decided to use AutoModelForTokenClassification BERT
models [
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 rule6s.
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 th13to 15th centuries, retrieved from the
CORDE linguistic database 1[
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) logic8. 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 7The evaluation was performed on test sets consisting of 4,300 words in Spanish, 1,340 words in French and 2,500 words in Italian. 8Due 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.
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 r(o“uges et anlfez ” 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 conjunctionq(u’). 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.
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 alignme9n.tTshe alignment is carried out on the basis of a main/pivot witness chosen in advance on the basis of philological knowledge, that is, Mich1a7[]. 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 fragme1n0t(figures 1 and 2). 9The parameters we used are the following: max-align=3, window=5, skip=-.2, margin=True, len_penality=True. 10As 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,
2.2.1. Alignment evaluation To assess the impact of the chosen segmentation method on alignment, we conducted an evaluation of the alignment produced by each of the two methods (regex and BERT-based segmentation).
Description of alignment results The evaluation is based on correcting the alignments obtained (correcting the alignment table of indices). Given that segmentation varies between different segmenters, the correction process must be repeated for results from both BERT-based segmentation and regex segmentation. Table5 shows some alignment results with the fragment iv-75-1 (see alsoAppendix D, Table 11).
These results present the same alignment error Linancelloto and BnF fr. 111, but due to two diferent 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 betweenco“nsiglio” 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 2.6, p. 21]:
Alignments are evaluated on the basis of the Alignment Error Rate (AER25),[ but it concerns a small number of alignment units only.
= 1 − 2 × | ∩ || + | gold| gold| Where L and goldconsist of predicted and hand-corrected links. The evaluation is conducted on indices by comparing automatically generated indices with corrected ones. As the alignment is produced for each witness compared to the pivot witness, an AER is produced for each pair, and Table6 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 diferent parts of the text, each with varying numbers of alignment units.
We observe a systematic improvement between the results that consider erroneous segmentation and those that do not, both for regex segmentation and BERT-based segmentation. Several 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 latte1r1. 11For example,Lancellotto doesn’t present the text for the beginning of Agloval episode, that leads to a poor alignment, because around 20 alignment units are not present in this witness.
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 instances 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 consolidate more units, potentially accommodating omissions that might afect other segments. In contrast, BERT-based segmentation, which ofers finer granularity, tends to make more errors.
There is indeed a notable diference in alignment results due to the size of the produced data: regex segmentation generates less segments, i.e., aligned units, as illustrated in Fig3u.re
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 segments 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 segmentation 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 semantically 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 impacts 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 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.
2.3.1. Method Most stemmatological and phylogenetic algorithms needs features (i.e. variants) that are arbitrarily coded (usually, using letters or numbers), e.g. variant 1, variant 2, 7e]tcI.n[our case, this necessitates, for each alignment unit, to cluster the readings from every witnesses into sets of relevant clustersT(able 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 diferent variant).
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 nois1e1][), 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 (). If DBscan does not necessitate to set the number of clusters, it requires the setting of the two parameteMrisnPts and . These parameters can only be chosen contextually, in relation to the number of individuals (witnesses) and the distances between them. For theMinPts, we set it at 1, given the small number of witnesses. For , we follow a methodology called DMDBScan, that tries to identify relevant values, by first computing the minimum distances to theMinPts 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 diferent densities1[0]. Following this methodology, we set at 0.2 (Appendix F).
DBScan is applied to the cosine distances between the readings of the witnesses in the embedding. Yet, due to the multilingualism of the corpus, and the important spelling variation characteristic of medieval language, some distances are artificially increased not for semantic 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( , ) weightedDist( , ) = 1 (( dist(, )) + ( 2 dist(, ))) where and are two witnesses, and a given alignment unit, and wher e( dist(, )) is the arithmetic mean of the distance between 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 inTable 8. The score is very low for the DBscan results, especially compared 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.
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.
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 this topic). Each one of the witnesses we chose presents particular readingssub(section E.1). However, we can distinguish two main groups within the tradition. On one hand, the witnesses Micha, BnF fr. 751, BnF fr. 111 andLancellotto, 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 groupss(ubsection E.3). In fact, the text in the Lanzarote presents a quite original version, especially due to the shortened passages it contasinusbs(ection 1.2). It also often ofers specific readings and innovations, particularly when the translation presents cases of direct speech instead of basic narrationsu(bsection 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 gallicisms5[, 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. However, sometimes,Lancellotto knows a common reading with Micha against BnF fr. 111su(bsection E.5). The Lancellotto also shows some innovations that prove that the very model of the translation is one not part of our selectiosnub(section E.6).
We can assume that the witnesses we studied belong to two main groups, and that the witnesses 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 identiifcation 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 [
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 parsimon1y2][. This algorithm can produce one or several trees, if diferent configurations achieve the same maximum parsimony value. The trees resulting from the analyses are presentedFiingure 4.
The trees resulting from the tables produced by the human annotators present,
unsurprisingly, strong similarities with the result of the human close reading expertise on the genealogy.
In particular, they all display the groupingLsan{ zarote, Sommer, Incunabula} versus
{Lancelloto, Micha, fr. 751 and fr. 111}. The second and third alternative trees from Annot. 2 also show
the grouping ofLancelloto and fr. 111, already observed by the editor of thLeancelloto [
In comparison, the results based on the DBScan do not show fully the oppositiLoann{zarote, 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 ofLtahnezarote, that is positioned close to the fr. 7519[], while Lancelloto remains relatively close to fr. 111.
If the results obtained from the automatic clustering of variants difer 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 difers 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.
We’ve designed a tool that can facilitate the study of unilingual and multilingual variant traditions, and its stemmatological analysis. The results obtained through clustering and stemmatological 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 coherent grammatical segments (relative pronouns, coordinating conjunctions between clauses, and subordinating conjunctions, etc.). This is a problem for all sentences that begin with personal 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 surprisingly well on medieval language states. Still, there is room for improvement and some ifne-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 timeconsuming 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 difers quite significantly from the alignment of monolingual texts, particularly with regard to the variations in the word order in the target sentence and in the source sentence. The micro-alignment could allow also more precise results in the stemmatological step.
The tool for alignment and collation is freely available on Zen1o0d.5o2:81/zenodo.12732533 and is maintained on github:https://github.com/ProMeText/Aquilign./The result data can be found in theresults_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.
Co-funded by theAgence Nationale de la Recherche under the Équipex Biblissima+ (ANR, 21ESRE-0005).
Co-funded by the European Union (ERC, LostMA, 101117408). Views and opinions 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.
• lancellotto: Lancellotto, italian witness, edited by Luca Cadioli (cf. bibliography).
Firenze, Biblioteca della Fondazione Ezio Franceschini, 1 (last quarter of t hthec1e4ntury). • lanzarote: Lanzarote, castillan witness, edited by Antonio Contreras Martín and Harvey L. Sharrer (cf. bibliography) .
Madrid, Bibioteca Nacional de España, 9611 (1th6, copy of a manuscript dated from 1414). • sommer: edition of H. OscarSommer (cf. bibliography).
Tomes IV and V: London, British Library, Additional 10293 (beginnings of theth14century). • micha: edition of Alexandre Micha (cf. bibliography).
Tome II: Cambridge, Corpus Christi College Library, 45nd(2half of the 13th century).
Tome IV: Oxford, Bodleian Library, Rawlinson D. 899 t(h14century). • 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).
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, theLanzarote 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. These passages correspond, in the other French witnesses, to:
BnF fr. 751
BnF fr. 333
BnF fr. 111 incunable Lancellotto
Lanzarote
Sommer
Micha f. 222r-226v f. 203r-218v
IV, 271-285 f. 254vb f. 297v 243r- f. 265r- IV, 327-362 286v to LXVIII, with reduced 8-12 254vb- f. 286v-320v V, 3-143
IV, LXXVLXXX, with LXXX 16 à 43 very reduced
proper id name II, XLVIII 29 suite – L 11 Charrette
I II, LXI suite 26-LXIX 23 Charrette
II IV, LXX part of iv-75 1-LXXXI 17 Agravain ii-48 ii-61
id ii-48 ii-61 status full without the end in the Lanzarote without the beginning and the end in the Lanzarote 227va-232va 248va-259va 270vb-291rb 23rb-47va 138vb-141vb
Diii.v-Dviii.r 151rb-157vb Fiiii.r-Gv.v 170va-183va Iviii.v- end vol. 1 iv-75 (Liiii.v) + vol. 2 beginning 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 diferent volumes.
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 defendre ? — 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 defendre, 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: … 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.
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 afinmerci; prie [mer]ci. ment fui crie cada mente le mercy. pedia merçed che|si duramente gli grida 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 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 miluuisaige ogo del suo viso, puis enbrace l’escu |et puis enbrace puys embrache e enbraço el poscia imbracbroche 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 ifert en leue tot dedens leaue en el agua nell’acqua iserrantment. |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 perduta, 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 amenquilz feussent duni tutti passes beurent molli, de leaue lun et laultre et se le cheual e si el caballo e se l cavallo neust este tan bueno no non fosse fort et bon ilz fuera sin dubda stato si forte, feussent noyes |el moriera enel anegati fossono et lun et laultre agua amenduni, car le cheualier e boores |porel ché l cavaliere estoit fort pe- peso delas ar- era pesante per sant pour les mas l’arme armes et se li chevals ne fust si buens, |noié fuissent ambedui, kar li chevaliers estoit pesans por les armes qu’il avoit vestues.
quil uestues. auoit que llebaua ch’elli vestite.
avea 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
In some sections, every witness has its own reading (fragment ii-61-1): sommer puis traist hors de sa quisse le piece de lespee fr751 Puis trait la piece de lespee qui en sa char estoit.
fr111
inc puis trait apres trait hors de hors de sa char la sa cuisse piece de lespee lespee lanzarote e despues Poscia trae tiro la fuori di pieça dela sua carne espada la spada e desu pierna l’apicca: micha Puis traist hors de sa char la piece de l’espee qui dedens estoit micha si le font desarmer, ou il volsist ou non, et il le fist a envis, micha mes il ne portoit mie tel escu
The two main groups can be established from common omissions, for example (fragment ii-48): sommer fr751 fr111 inc
lanzarote Si le font si le font desarmer desarmer uossist ou uoulsist il non. et il le ou non. et fait mout il le fist a enuiz. peine. si l fanno disarmare o volesse O non, ed elli il fece ad 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 oppossheield to weapons, castle to forest. sommer fr751 fr111 inc lanzarote lancellotto mais il ne mais il ne mais il ne mes il ne mas portoit portoit portoit pas portoit traia mie tels mie tel telescu mye telles tales armes escu armes armas no ma e’ non el porta mica talescudo Alchastel, En la Au chas- ou chastel fet ele, de forest fait tel fait ele de fleago Floego elle de de floego. fait ella lforegas.
En la for- enla flo- Al est de flo- resta de castello, rega: donseglo- diss’ella, rega dixo di Fleego, la donçella
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 e el fue alla por aluergar e fallo ala puerta dos frailes micha et il torne cele part por herbergier. Et quant il i vint, si trueve a la porte IIII. des freres sommer si torne cele part pour herbergier. Et quant il vint la si trouua .ii. freres
The first group, with Micha, is characterized by the mention of theporte and the presence of fourfreres whereas the second one, the Sommer’s one, is characterized by the absence of the mention of the porte and the presence of only twofreres. The text in the Lanzarote presents the mention of aporte and only twofreres.
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’unde piedras preçiosas que estaua toda grant roialme. Se la tumbe fu de grantlabrada 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 eustnunca viera ni ome del mundo que la onques veue: non ouiese visto no podria asmar la fermosura que enella auia de tales cosas como enella estauan si se merveille molt qui puet estre lie don lançarote se marauillo e dixo princes entresi mesmo Quien podria ser el prinçipe
We can find a multitude of readings that oppose the groupLancellotto/Micha against BnF fr. 111 (fragment ii-61-1), such as the ones in the following table: micha 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): 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 comcompagnone per portar- paignon por porter, nela,
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 asm-olto isbigotito. sés de sa bialté que del vaissel,
Ma di cosa che veggia non si maraviglia 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