=Paper=
{{Paper
|id=Vol-2989/long_paper51
|storemode=property
|title='Psyché' as a Rosetta Stone? Assessing Collaborative
Authorship in the French 17th Century Theatre
|pdfUrl=https://ceur-ws.org/Vol-2989/long_paper51.pdf
|volume=Vol-2989
|authors=Florian Cafiero,Jean-Baptiste Camps
|dblpUrl=https://dblp.org/rec/conf/chr/CafieroC21
}}
=='Psyché' as a Rosetta Stone? Assessing Collaborative
Authorship in the French 17th Century Theatre==
https://ceur-ws.org/Vol-2989/long_paper51.pdf
‘Psyché’ as a Rosetta Stone? Assessing Collaborative
Authorship in the French 17th Century Theatre
Florian Cafiero1 , Jean-Baptiste Camps2
1
GEMASS | CNRS / Université Paris-Sorbonne, 59-61 rue Pouchet, 75017 Paris, France
2
École nationale des chartes | Université PSL, 65 rue de Richelieu, 75002 Paris, France
Abstract
During the 17th century, a significant number of collaborations emerged between playwrights, among
which authors as famous as Pierre Corneille, Thomas Corneille or Molière, as well as Philippe Quin-
ault or Jean Donneau de visé. The actual division of labour between authors can sometimes be
deduced from historical documents, but is most of the time uncertain. In this paper, we try to
address this question by using the information we got from one specific instance of collaboration:
Psyché (1671). We first try to assess the accuracy of the notice to the reader of the printed edition of
the play, where each author’s involvement is clearly claimed, using machine learning and “rolling sty-
lometry” methodology. We then use the optimal parameters already applied to this play to analyse
other collaborative works of the time, in particular cases of potential collaboration between Thomas
Corneille and Jean Donneau de Visé in Circé and L’Inconnu.
Keywords
authorship attribution, French literature, 17th century, rolling stylometry, collaborative authorship
1. Introduction
1.1. Collaborative authorship and stylometry: the challenges of the Théâtre
classique
Stylometry, and notably ‘rolling stylometry’, has been successfully used to identify the co-
authors of a literary work in several cases. With Burrows’ delta, it has for instance been
used to assess Ford’s claims about his implications in collaborations worth Joseph Conrad
[25], to determine the beginning of Vostaert’s intervention on Dutch Arthurian novel Roman
van Walewein [5], or to understand Lovecraft’s and Eddy’s implication in The Loved Dead
[13]. A distance-based approach has also advanced our understanding of the collaboration
between Julius Caesar and General Hirtius, and confirmed that pseudo-Caesar texts had been
written by an anonymous writer [16]. Principal Components Analysis was used to visualise
the importance of Hildegard of Bingen’s last secretary Guibert-Martin de Gembloux in her
late production [15]. Using support-vector machines, rolling stylometry more recently helped
to confirm John Fletcher and William Shakespeare’s collaboration for Henry VIII [22]. Co-
authorship between Nobel Prize winner Yasunari Kawabata and one alleged ghostwriter was
detected using in parallel various supervised machine-learning settings [29].
CHR 2021: Computational Humanities Research Conference, November 17–19, 2021, Amsterdam, The
Netherlands
£ florian.cafiero@cnrs.fr (F. Cafiero); Jean-Baptiste.Camps@chartes.psl.eu (J. Camps)
DZ 0000-0002-1951-6942 (F. Cafiero); 0000-0003-0385-7037 (J. Camps)
© 2021 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Wor
Pr
ks
hop
oceedi
ngs
ht
I
tp:
//
ceur
-
SSN1613-
ws
.or
0073
g
CEUR Workshop Proceedings (CEUR-WS.org)
377
Yet, the task raises a number of difficulties. First, style change detection is not a completely
solved problem, and even recent competitions [30] have shown that precisely determining style
breaches, i.e. places where the authorship switches in a collaborative text, was still a complex
task.
The task is all the more complicated for French 17th century plays, known as “théâtre
classique”. The 17th century is a time when the notion of authorship settles in France [31].
Yet, as stated by Quiel, the need for originality or the fear of being too derivative were not yet
a major worry for the authors of the time:
French playwrights do not refrain from imitating or adapting whichever text is likely
to be presented on a stage, sometimes without even bothering to make significant
additions or to add characteristics from their own literary style [23].1
Strongly codified, building their plots on the same Spanish, Italian, Greek or Latin models,
these plays are often very homogeneous, which makes it more difficult to properly attribute
texts . In particular, similarities induced by the literary genre or subgenre can be as strong as
similarities induced by the authors’ idiolect [27, 3, 4].
Imitation could even go to the point of piracy or plagiarism, such as attempting to steal
another author play, or led to disputes about the true source of a story. The polygraph and
literary “entrepreneur” Donneau de Visé [28] – famous later on as the as the powerful founder
and editor of the monthly Mercure Galant, a collective literary periodical that published a
mix of recycled manuscript or printed pieces, reader contributions (poems, etc.) and his
own original material – was familiar of such practices at least in his early career. He more
or less started his career by trying to steal Sganarelle ou le cocu imaginaire from Molière:
before Molière could publish himself his own play, Donneau published in 1660, with the help
of the printer Jean Ribou, both a pirate edition of Molière’s Sganarelle, in which he added
his commentaries and that he went to the point of dedicating to Molière himself (!), and a
plagiarised play, La Cocue imaginaire, where he reversed masculine and feminine roles [8]. He
also had an important dispute with Quinault around the Mère Coquette, two plays with this
name being published, in 1665 by Quinault and 1666 by Donneau.
A final challenge is related to the diverse potential nature of collaborative writing. Ac-
cording to Pennebaker and Ireland [21], three main hypotheses can be made on the result of
collaborative writing: the “Just-like-another-member-of-the-team hypothesis” were collabora-
tive writing can be distributed in portions successively attributable to the idiolect of one of
the authors; “The average person hypothesis” were the resulting style is an average of the
authors’ idiolects; and finally the “synergy hypothesis” were the contact situation and inter-
actions between different indivual idiolects create a resulting singular style, different of each
individual one. This last hypothesis tends to be verified in famous cases such as the Lennon
and McCartney collaboration or the one between Hamilton and Madison.
As we well see, suspected collaborative writing cases in 17th century French theatre – even
though we might suspect them on declarative grounds to fall mainly in the first “Just-like-
another-member-of-the-team hypothesis” – also presents clue of a division of the work on
different authorial levels, for instance content versus form, narrative versus versification.
To help us address the various collaborative authorship problems raised by the writings of
this century, we thus try to work on one of the best documented collaboration of the time:
Psyché.
1
All translations are our own.
378
1.2. ‘Psyché’ and its notice ‘to the reader’
Psyché is a tragedy-ballet in five acts, written in free verse, and created in 1671, during the
very long festivities following the peace of Aix-la-Chapelle in 1668. It originates in Louis XIV’s
desire to give a new show in the “Salle des Machines” of the Tuileries Palace. Built in 1660
by renowned architect Louis Le Vau, this theatre took its name from the machinery designed
by Gaspare, Carlo and Lodovico Vigarani, allowing for spectacular effects and complex set
changes [6]. The acoustics of this theatre were poor. It was thus abandoned, not being used
since 1662. But its large capacity, and the existing impressive sets from Cavalli’s opera Ercole
Amante, would have drawn the French king to commission a new play specifically designed for
this place.
In 1758, Lagrange-Chancel reported in the preface to his own Orphée that several authors
would have proposed a project for this occasion.
The late King having resolved to give to all his court one of these great celebra-
tions in which he liked to have a rest from his works, wanted to take advice from
Racine, Quinault, and Molière, which, among the the great geniuses of this century,
he regarded as the most capable of contributing, by their talent, to the magnifi-
cence of his pleasures. To that effect, he asked them to pick a subject for which
they could use an excellent decor representing the underworld, kept safe and sound
in the furniture storage unit. Racine proposed the subject of Orphée; Quinault,
the abduction of Proserpine, which he subsequently turned into one of his most
beautiful operas; and Molière, with the help of the great Corneille, championed the
subject of Psyché, which prevailed over the two others [17].
Through the correspondence of the Vigaranis, we know that the decision to give this Psyché
at the Salle des Machines was made only a few weeks before it was played. In a letter written
on December 12th, 1670, Vigarani explained that they were “preparing a great show, to be
performed for the Epiphany at the Tuileries theatre” [24], The imminence of the deadline forced
everyone to rush to get the work done. As stated in a letter written on December 15th, 1670,
“Carlo is very busy because of the show prepared for the Epiphany. He is very tired. He is
doing his best to please the King, but he doubts he will be strong enough to continue.”
The lack of time apparently had consequences on Molière’s ability to finish the play. This
led to a singularity: an official account of each author’s implication. In a notice from the
publisher “au lecteur” (to the reader), we find this explanation on how the work is supposed
to have been divided:
This work is not written by a single hand. Mr Quinault wrote all the poetry
of the parts set to music, except the Italian Complaint. Mr de Molière wrote the
outline of the play, set its arrangement - he focused more on the beauties and the
pump of the show than on its strict observance of the rules. Regarding versification,
he did not get the time to execute it in its entirety. Carnival was approaching, and
the insisting demands of the King, who wanted to entertain himself several time
before Lent, forced him into accepting some assistance. Thus, only verses from the
Prologue, the First Act, the first scene of the Second Act, and the first scene from
the third Act are his work. Mr. Corneille used two weeks to versify the rest; and
this way, His Majesty’s orders were satisfied in time [20].
379
Does this notice to the reader seem plausible? The situation where the King’s urgent de-
mands change the author’s agenda was not unprecedented. For instance, when in 1664, Louis
XIV commissioned a new play with ballet to Molière for the “Plaisirs de l’île enchantée” fes-
tivity in Versailles, the latter could not finish in time the versification of his play. As stated
in its first edition [19], “an order from the King, who pressed this matter, forced [the author]
to finish all the rest in prose.” Only thirty percent of La Princesse d’Elide is thus in verse, the
rest being in prose.
Stylistic studies also seem to confirm the plausibility of this notice. Psyché is one of the
rare examples of mixed verses by their authors. Before that, Pierre Corneille had only used it
once, in Agésilas in 1666. The same goes for Molière, who also mixed various types of verses in
his Amphitryon in 1668. And the way the two authors use mixed verses is different [2]. While
Molière did not hesitate to use heptasyllables in Amphitryon, Corneille never used a single one
of them in Agésilas. This distinction is still observable in Psyché: in the part attributed to
Molière, we find 37 heptasyllables - which fits the proportion observed in Amphitryon; in the
part attributed to Corneille, we do not find any heptasyllable.
Without taking this notice “to the reader” for granted, we can consider that it gives poten-
tially truthful information regarding the play, that we will first try to disprove or verify.
1.3. Hollywood in the 17th century: Special effects, music and the “Pièces à
machines”
Psyché raises a few specific concerns because of its genre. This play is indeed a rare instance
of pièce à machines, a subgenre very much in favour between the 1650s and the 1670s, but
of which only 15 plays or so have been composed [33, 32]. Changes of scenery in each act,
flying characters, raging seas, thunderous blows… the “pièces à machines” make the most of
the machinery available at the time, to propose spectacular shows to the audience, including
passages set to music. The first model of the genre probably is Andromède (1650) by Pierre
Corneille [1].
But other authors followed him, sometimes collaborating to produce that kind of plays, like
Thomas Corneille (Pierre Corneille’s younger brother) and Jean Donneau de Visé. Amongst
the most impressive shows of the time, the first collaboration between the two authors, Circé
(1675), encouraged them to work together for other works. This play was thus quickly followed
by another collaboration: L’Inconnu (1675). It seems that Donneau de Visé’s implication in
this play could be very significant. It draws its inspiration from Donneau de Visé’s own
tenth short story [18] in Les Nouvelles Galantes, Comiques et Tragiques (1669). In Thomas
Corneille’s obituary notice, Donneau de Visé himself claimed he played an important part in
the writing:
to make progress, I wrote the whole play in prose, and while I was writing the
prose of the second act, he was transforming the prose of the first into verse; and as
prose is easier than verse, I had the time to write those of the entertainments, and
especially the dialogue of Love and Friendship, which did not displease the public
[7].
Yet, despite these claims, the only name cited in the “Privilège du roi” of the first printed
edition is Thomas Corneille. Why omitting to mention Donneau de Visé? We know from
the Registre de La Grange that Donneau de Visé was payed fees for the writing of the play,
and he received the same amount for it as Thomas Corneille. It is thus historically extremely
380
unlikely that Donneau de Visé would not have contributed to the play. But did he overstate
his implication in the writing of the play? Or were there only “commercial reasons” to avoid
mentioning his name - Thomas Corneille being more respected as a playwright than he was?
Thus, using the parameters found optimal in our benchmark phase, and then tested on
Psyché, we will try to work on Circé and L’Inconnu, two pièces à machines also written in
verse and allegedly written collaboratively.
The scarcity of “pièces à machines” however makes it challenging to build a stylometric
approach only on a subgenre-specific approach. We thus compare two methods in this paper:
a genre-specific approach, on a small dataset, and a cross-genre approach, on a considerably
larger dataset (see appendix A).
2. Materials and methods
2.1. Choice of plays and dataset
To investigate these difficulties, we built two different analytic setups, with two very different
philosophies:
a genre-specific approach, in which we built a training corpus including one play from the
same subgenre for each of the three involved authors, as well as two control authors –
one play by Boyer, and two plays for Quinault, as his plays were significantly shorter (see
appendix A.1.1). We then benchmarked different sample lengths, using a leave-one-out
approach. We lacked sufficient data to include Thomas Corneille in this setup, because
the two available plays that could fit the definition, Circé and L’Inconnu are suspected
to be collaborations with Donneau de Visé that we want to analyse later on.
a cross-genre approach in which we took all available single-author verse or mixed plays
containing more than 400 verses from each candidate in the Théâtre classique corpus
[12]. In order to recreate the conditions of the actual analysis on the Pièces à machines,
that is to evaluate the performance in a cross-genre setup on a given specific genre or
sub-genre not represented in the training set, we set apart a subgroup of heroic comedies
as an unseen test set on which to benchmark the models (see appendix A.2.1 and A.2.2).
For the mixed plays, we retained them only if there were more than 400 verses to be
extracted.
For each case studied, the set of candidate authors is known. We thus use an authorship
attribution rather than an authorship verification setup.
In terms of features, after suppressing editorial punctuation and lowercasing the texts, we
extract character 3-grams, a standard choice in authorship attribution [14, 26].
2.2. Calibration
The chosen size of the sample can be seen as a trade-off between accuracy and granularity: the
smaller the samples are, the better the “resolution” of the analysis and the ability to locate
precise stylistic breaks or identify limited shifts in hands [10]; the bigger they are, the more
statistically reliable is our computation, with optima often observed in the 2500-5000 words
range [9]. To find a good compromise, we experimented with a variety of lengths, from 10 to
300 verses, an upper limit close to the size of an act (Table 1). Each time, we normalised the
381
data by using z-scores for variables and applied Euclidean vector-length normalisation (i.e.,
L2 normalisation) to texts [11] and trained a linear Support Vector Classifier (SVC) model,
using a Python sklearn pipeline (see Code section). These results show that a first peak in all
metrics is reached at length 150 verses for the smaller genre-specific corpus, with perfect scores,
as well as for the larger cross-genre corpus when considering the F1 score (with F1= 0.92).
Table 1
Scores resulting from the SVM training benchmark on samples of length from 10 to 500 verses; precision,
recall and F1 score are given, as well as the support (number of test samples). The small same-genre models
are tested using a leave-one-out approach, then averaged; the larger cross-genre models are tested on out-
of-domain plays.
Sample length (verses) genre-specific cross-genre
Prec. Rec. F1 support Prec. Rec. F1 support
10 0.85 0.85 0.85 1069 0.57 0.47 0.46 1120
20 0.92 0.92 0.92 533 0.68 0.59 0.59 558
30 0.95 0.94 0.94 354 0.75 0.67 0.67 371
40 0.96 0.96 0.96 264 0.76 0.71 0.70 277
50 0.98 0.98 0.98 212 0.81 0.75 0.75 222
100 0.99 0.99 0.99 103 0.87 0.81 0.82 110
150 1.00 1.00 1.00 68 0.93 0.91 0.92 72
200 1.00 1.00 1.00 51 0.94 0.92 0.92 53
250 1.00 1.00 1.00 41 0.93 0.90 0.91 42
300 0.97 0.97 0.97 32 0.96 0.94 0.94 35
Detailed scores for the retained sample length on the cross-genre corpus show that Boyer is
the best recognised, while precision for Donneau de Visé and Molière is better (the model is
never wrong in attributing plays to them) than recall (the model misses a few samples), while
the opposite is true for both Corneille brothers (Table 2).
Once the sample size of 150 verses chosen, we train a final SVM model for each setup, on
the complete training set, and then, following rolling stylometry methods, we apply this model
to every successive portion of length n, with a step of 1 (and so, an overlap of n − 1 between
two successive portions, e.g., verses 1-150, 2-151, 3-152…). We then extract the classification,
and plot the decision function for each classifier.
Like with many geometrical methods, this type of analysis rests on the representation of
texts or samples as points in a high dimensional space, based on the frequency of the selected
features. In our case, the frequency of each type of character 3-grams are used as a coordinate
on one axis of an high-dimensional space.
A support vector machine computes a hyperplane in this high-dimensional space, in order
to achieve the best separation between two sets of dots (i.e., text from author 1, text from
all other authors). Intuitively, a good separation is achieved by the hyperplane that has the
largest distance to the nearest training data points of any class (called functional margin),
since in general the larger the margin the lower the generalisation error of the classifier.
The decision function tells us how close each sample is to the hyperplane separating each
class. A negative value means that the sample is outside, a positive, inside. The higher the
score, the deeper inside the class is located a dot, which can be interpreted as a strength of
the authorial markers or an increase in the confidence of the classifier.
By monitoring the decision function of all candidate authors, we can see when a portion of
the text is getting closer to the style of an individual author. When the value of the decision
function for one author gets high, while the values for all others remain low, it is easy to
382
Table 2
Detailed class scores on out-of-domain test for the SVM trained on the cross-genre setup, with sample length
150 verses.
Prec. Rec. F1 support
BOYER 1.00 1.00 1.00 12
CORNEILLEP 0.86 1.00 0.92 12
CORNEILLET 0.82 1.00 0.90 14
DONNEAUDEVISE 1.00 0.91 0.95 11
MOLIERE 1.00 0.75 0.86 12
QUINAULT 0.90 0.82 0.86 11
macro avg 0.93 0.91 0.92 72
weighted avg 0.93 0.92 0.92 72
Figure 1: Value of the decision function for each classifier and each successive rolling sample; each sample
is placed horizontally at it’s median point in verses and vertical grey dashed lines denote scenes. Both setups
achieve largely consistent results.
attribute this portion to this candidate. Yet, in the case where all decision function would
decrease or remain low simultaneously, the status of the portion is hard to assess, and could
be alternatively attributed to a synergy between several candidates, to the intervention of
an author outside the set or could be the results of some kind of noise, in particular generic
discrepancies between the training set and the portion being assessed.
Each sample’s horizontal placement on the graph is determined by its median point, in
verses (fig. 1). For instance, the score attributed to the sample ranging from the 400th to the
550th verse will be placed at the 475th verse. This implies a small but significant distortion
regarding the placement of each score in the timeline of each play. It must be taken into
account when reading the graph.
3. Results
3.1. Psyché by P. Corneille, Molière and Quinault
Both setups provide globally consistent results. The end of the prologue, most of the first act,
and the beginnings of the second and the third acts are attributed to Molière. Molière’s imprint
then seems to gradually fade away, while Pierre Corneille’s contribution appears constantly
growing throughout the play. This result seems solid, and even conservative regarding Molière’s
383
share of the work. When using the cross-genre setting, the precision for Molière is 100 % while
the recall is lower, at 0.75 % (Table 2). On the opposite, the recall for Pierre Corneille is perfect
(100 %), while the precision is lower (86 %). This setting should thus underestimate Molière’s
participation in the play. Yet, even this computation confirms that the statements made in
the notice from the publisher to the reader regarding him are globally accurate. A small but
interesting difference between our analysis and the notice concerns the spike of Corneille’s
decision function at the end of the first act. According to it, it could be possible that Corneille
had a hand in finishing the first act, under a fashion comparable to that of the other acts
(begun by Molière, ended by him) but in much smaller proportions, considering the brevity of
scenes 4 to 6 of the first act (forty verses in total).
Performance on Quinault is also satisfying, even if his interventions can be sometimes quite
short: his alleged part in the prologue barely exceeds 50 verses, his “second intermède” is less
than 20 verses long etc. His intervention in the prologue is a bit masked by the 150 verses
window. Yet, we can see that the decision function for Quinault is high at the very beginning.
The finale (“Cinquième intermède”) and the 60-verse long “troisième intermède” at the end
of the third act are quite visible, and attributed to Quinault, which is consistent with the
notice to the reader. Even the very short “second intermède” is detected by both methods,
and especially clearly in the genre-specific approach.2
3.2. ‘Circé’ by T. Corneille and J. Donneau de Visé
If we consider the results of the cross-genre setup on ‘Circé’ (fig. 2) – the only setup with
training material for Thomas –, the major implication of Thomas Corneille in the writing of
this play is self-evident. Donneau de Visé seems to rise during a small passage of the prologue,
somewhere around the third scene or perhaps the “Prologue de la musique et de la comédie”
(accounting for the blurriness due to window size), a part of sung dialog, which would be
consistent with some of his claims. Yet, no other passage emerges that would make it even
seem plausible that he partly versified them.
The spike of Quinault at act 2, scene 7 matches also a sung dialog (“dialogue de Sylvie et
de Tircis, qui se chante”). Quinault obviously knew how to write passages to be sung and
collaborated with Thomas Corneille in other occasions. Did Donneau de Visé have another
try at plagiarism, after having experimented with such practices earlier in his career? Or did
Donneau or Thomas Corneille ask a small amount of help to a colleague for a specific passage?
Without further analysis, it is difficult to answer, yet it is to be noted that the value of the
decision function only barely crosses 0 (implying positive class membership) on one single
point. This could be an artefact due to generic attractions, because Quinault produced an
important number of musical texts, represented in the training material. It is in any case
deserving of further investigation before drawing any firm conclusion.
3.3. ‘L’Inconnu’ by T. Corneille and J. Donneau de Visé
When applying the same method on L’Inconnu (fig. 3), here again, Donneau de Visé’s implica-
tion is hard to assess, while Thomas Corneille’s style seems easily recognised. Small passages
for which he claims authorship could well be too difficult to detect which such windows - but
appeared in Psyché. Here, the decision function seems barely affected. It is to be noted that
2
It is to be noted that we removed the “Premier intermède”, written in Italian (by Lully), for obvious
reasons.
384
Figure 2: Results for Circé, using the cross-genre setup.
the value of the decision function for Thomas collapses below 0 on a few occasions, especially
during the long dialogues at the end of act 1 and act 2. This could also receive an interpretation
based on generic discrepancies between the training set and this part of the text.
4. Discussion
On Psyché the results of the rolling analysis verifies very closely the self-declaration of the
notice. In comparison, we seem only to identify a non declared limited intervention of Corneille
at the end of the first act.
Our results seem to confirm that Donneau de Visé’s contribution to the final writing of the
plays he co-signed with Thomas Corneille was quite scarce. This of course does not mean
that his contribution to those plays was non-existent. Stylometric analyses such as the ones
performed here mostly detect the style of the person who actually wrote the last version of
the sentences. Thomas Corneille for instance versified Molière’s famous comedy in prose, Dom
Juan, after his death and stylometry attributes the play to Thomas Corneille without a blink
[3], while Molière arguably contributed quite a bit to the final result… Donneau de Visé could
have given a lot of insights about the intrigue, written large passages in prose etc. But for now,
his contribution strictly to the verses seem even scarcer that what he claimed after Thomas
Corneille’s death.
In broader terms, the clues gathered here seem to point mainly towards the first case of
collaborative writing described by Pennebaker and Ireland [21], the “Just-like-another-member-
385
Figure 3: Results for l’Inconnu, using the cross-genre setup.
of-the-team hypothesis”: each portion is mostly attributable to the individual style of a single
author, i.e., the one responsible for the final form of the text, not for its content or for previous
formulations (in particular in the case of –necessarily heavy – transpositions between prose and
verse ). Yet, this would be deserving of further research to improve still our understanding of
authorial collaborations during the Grand Siècle and beyond. For now, it remains impossible
to say if some points of the texts, where the decision function for all candidate authors collapse,
can be attributed to synergies instead of, for instance, generic disturbances (see the case of the
Inconnu).
5. Further research
Further research are still needed to confirm and extend the results on Circé and L’Inconnu,
and more generally on collaborative writing during this age. Increasing the size of the training
set, especially for Donneau de Visé could be a first lead. In particular, we were not able to
secure access to usable digital text of plays such as Les Amours de Vénus et d’Adonis (1670)
or Amours de Bacchus et d’Ariane (1672), two pièces à machine he wrote alone during the
same decade as his collaborations with Thomas Corneille. Running an efficient OCR and post-
correction of those texts, already digitised by the Bibliothèque nationale de France, should thus
be an important next step.
In terms of analysed features, we could extend our work to account for metrical features
[22], accounting for instance for verse length in syllables, and the sequence of such lengths in
386
the “vers libres” parts. We could also try to cross stylistic with thematic features, to further
investigate contribution to the plot by opposition to contributions to the versification.
The Quinault spike in Circé could also make us think that generic attractions are still an
issue in our study: having written numerous opera librettos, Quinault could be a designated
candidate for whatever looks like a sung passage to our SVM model. We should thus check for
possible imbalances in the training corpus. First experiments however seem to show that our
results stay even when downsizing Quinault’s sung passage in the training set.
Finally, we could extend our process to collaborations in prose - which were quite numerous
in the théâtre classique in general, and which also occurred in a pièce à machines such as La
Devineresse by Thomas Corneille and Jean Donneau de Visé.
Code and data availability
Code and datasets are available at 10.5281/zenodo.5517801.
Acknowledgments
We thank Pr. Georges Forestier for his input on Psyche and the subgenre of “pièces à ma-
chines”, Thibault Clérice for fruitful discussions on machine-learning and stylometry, and
anonymous reviewers for their careful reading and insightful suggestions. Errors remain our
own.
References
[1] N. Akiyama. “Corneille et ses pièces à machines”. In: Dix-septième siècle 3 (2010),
pp. 403–417.
[2] R. Bray. “L’Introduction des vers mêlés sur la scène classique”. In: Pmla 66.4 (1951),
pp. 456–484.
[3] F. Cafiero and J.-B. Camps. “Why Molière most likely did write his Plays”. In: Science
advances 5.11 (2019), eaax5489. doi: 10.1126/sciadv.aax5489.
[4] F. Cafiero, J.-B. Camps, S. Gabay, and M. Puren. “La naissance du style: auteur vs
genre aux XVIIe et XIXe siècles”. In: Humanistica 2020. 2020. url: https://hal.archives-
ouvertes.fr/hal-02577853/.
[5] K. van Dalen-Oskam and J. Van Zundert. “Delta for Middle Dutch: Author and Copyist
Distinction in Walewein”. In: Literary and Linguistic Computing 22.3 (2007), pp. 345–
362.
[6] M. De Pure. Idée des spectacles anciens et nouveaux. Minkoff, 1668.
[7] J. Donneau de Visé. “[Notice nécrologique]”. In: Mercure galant (1710), pp. 270–299. url:
https://gallica.bnf.fr/ark:/12148/bpt6k6351123z/f276.
[8] J. Donneau de Visé. Préface à Sganarelle, ou le Cocu imaginaire. Ed. by G. Forestier and
C. Fournial. Paris: Jean Ribou, 1660. url: http://idt.huma-num.fr/notice.php?id=317.
[9] M. Eder. “Does Size matter? Authorship Attribution, Small Samples, Big Problem”. In:
Literary and Linguistic Computing 30.2 (2015), pp. 167–182. doi: 10.1093/llc/fqt066.
url: https://academic.oup.com/dsh/article/30/2/167/390738.
387
[10] M. Eder. “Rolling Stylometry”. In: Digital Scholarship in the Humanities 31.3 (2016),
pp. 457–469.
[11] S. Evert, T. Proisl, F. Jannidis, I. Reger, S. Pielström, C. Schöch, and T. Vitt. “Under-
standing and Explaining Delta Measures for Authorship Attribution”. In: Digital Schol-
arship in the Humanities 32.suppl_2 (2017), pp. ii4–ii16. doi: 10.1093/llc/fqx023. url:
https://academic.oup.com/dsh/article/32/suppl%5C%5F2/ii4/3865676.
[12] P. Fièvre. Théâtre classique. 2007. url: http://www.theatre-classique.fr/.
[13] A. A. Gladwin, M. J. Lavin, and D. M. Look. “Stylometry and Collaborative Authorship:
Eddy, Lovecraft, and ‘The Loved Dead’”. In: Digital Scholarship in the Humanities 32.1
(2017), pp. 123–140.
[14] M. Kestemont. “Function Words in Authorship Attribution: From Black Magic to The-
ory?” In: Proceedings of the 3rd Workshop on Computational Linguistics for Literature
(CLFL). 2014, pp. 59–66.
[15] M. Kestemont, S. Moens, and J. Deploige. “Collaborative Authorship in the Twelfth
Century: A Stylometric Study of Hildegard of Bingen and Guibert of Gembloux”. In:
Digital Scholarship in the Humanities 30.2 (2015), pp. 199–224.
[16] M. Kestemont, J. Stover, M. Koppel, F. Karsdorp, and W. Daelemans. “Authenticating
the Writings of Julius Caesar”. In: Expert Systems with Applications 63 (2016), pp. 86–96.
[17] F.-J. ( Lagrange-Chancel. Œuvres de Monsieur de Lagrange Chancel revues et corrigées
par lui-même. 1758.
[18] P. Mélèse. Un homme de lettres au temps du grand roi, Donneau de Visé: fondateur du
Mercure galant. Librairie Droz, 1936.
[19] Molière. La Princesse d’Elide. Robert Ballard/Thomas Jolly/Guillaume de Luynes/Louis
Billaine, 1665.
[20] Molière. Psiché: tragédie-ballet, par I.B.P. Molière. 1671. url: https://gallica.bnf.fr/ark:
/12148/bpt6k70160j/f7.item.
[21] J. W. Pennebaker and M. E. Ireland. “Using Literature to Understand Authors: The Case
for Computerized Text Analysis”. In: Scientific Study of Literature 1.1 (2011), pp. 34–48.
doi: 10.1075/ssol.1.1.04pen. url: https://www.jbe-platform.com/content/journals/10.
1075/ssol.1.1.04pen.
[22] P. Plecháč. “Relative Contributions of Shakespeare and Fletcher in Henry VIII: An Anal-
ysis Based on Most Frequent Words and Most Frequent Rhythmic Patterns”. In: Digital
Scholarship in the Humanities (2019).
[23] F. G. Quiel. “Comedia palaciega et tragi-comédie française au XVIIe siècle: adresse au
lecteur et transfert littéraire dans la Cassandre de l’abbé Boisrobert”. In: Revista de
Lenguas Modernas (2010).
[24] G. Rouchès. Inventaire des lettres et papiers manuscrits de Gaspare, Carlo et Lodovico
Vigarani conservés aux Archives d’État de Modène, 1634-1684. Champion, 1913.
[25] J. Rybicki, D. Hoover, and M. Kestemont. “Collaborative Authorship: Conrad, Ford and
Rolling Delta”. In: Literary and Linguistic Computing 29.3 (2014), pp. 422–431.
388
[26] U. Sapkota, S. Bethard, M. Montes, and T. Solorio. “Not all Character N-Grams are
Created Equal: A Study in Authorship Attribution”. In: Proceedings of the 2015 con-
ference of the North American chapter of the association for computational linguistics:
Human language technologies. 2015, pp. 93–102.
[27] C. Schöch. “Fine-Tuning our Stylometric Tools: Investigating Authorship, Genre, and
Form in French Classical Theater”. In: Digital Humanities 2013: Conference Abstracts.
2013, pp. 383–86.
[28] C. Schuwey. Un entrepreneur des lettres au XVIIe siècle: Donneau de Visé, de Molière au
Mercure galant. Lire le XVIIe siècle 69. Paris: Classiques Garnier, 2020. doi: 10.15122/
isbn.978-2-406-09572-9.
[29] H. Sun and M. Jin. “Collaborative Writing of ‘Otome no minato’”. In: Structure, Function
and Process in Texts (2018), p. 116.
[30] M. Tschuggnall, E. Stamatatos, B. Verhoeven, W. Daelemans, G. Specht, B. Stein, and
M. Potthast. “Overview of the Author Identification Task at PAN-2017: Style Breach
Detection and Author Clustering”. In: CLEF (Working Notes). 2017.
[31] A. Viala. Naissance de l’écrivain. Minuit, 1985.
[32] H. Visentin. “La tragédie à machines ou l’art d’un théâtre bien ajusté”. In: Littératures
classiques, hors-série, 2002. Mythe et histoire dans le théâtre classique. Hommage à
Christian Delmas (2002). doi: 10.3406/licla.2002.1825. url: https://www.persee.fr/doc/
licla%5C%5F0992-5279%5C%5F2002%5C%5Fhos%5C%5F1%5C%5F1%5C%5F1825.
[33] H. Visentin. “Le théâtre à machines: Succès majeur pour un genre mineur”. In: Littéra-
tures classiques 51.1 (2004), pp. 205–222.
A. Plays used as training
A.1. Genre-specific setup
A.1.1. Train (with leave-one-out test)
author title date n. words
Boyer, Claude LES AMOURS DE JUPITER ET DE SÉMÉLÉ, TRAGÉDIE 1666 20554
Corneille, Pierre LA CONQUÊTE DE LA TOISON D’OR, TRAGÉDIE 1661 22779
Donneau de Visé, Jean LES AMOURS DU SOLEIL, PASTORALE. 1671 20898
Molière Amphitryon, Comédie 1668 17603
Quinault, Philippe THÉSÉE, TRAGÉDIE 1675 9200
Quinault, Philippe CADMUS et HERMIONE, TRAGÉDIE 1673 6840
A.2. Cross-genre setup
A.2.1. Train
author title date n. words
Boyer, Claude AGAMEMNON, TRAGÉDIE. 1680 16638
Boyer, Claude LES AMOURS DE JUPITER ET DE SÉMÉLÉ, TRAGÉDIE 1666 20554
Boyer, Claude ARISTODÈME 1648 13642
Boyer, Claude ARTAXERCE, TRAGÉDIE 1683 16777
Boyer, Claude CLOTILDE, TRAGÉDIE. 1659 20711
Boyer, Claude JUDITH, TRAGÉDIE 1695 14507
Boyer, Claude LISIMÈNE OU LA JEUNE BERGÈRE, PASTORALE 1672 18040
Boyer, Claude LA MORT DES ENFANTS DE BRUTE, TRAGÉDIE. 1648 14634
Boyer, Claude OROPASTE OU LE FAUX TONAXARE 1663 22328
Boyer, Claude LA PORCIE ROMAINE 1646 16131
Boyer, Claude PORUS OU LA GÉNÉROSITÉ D’ALEXANDRE, TRAGÉDIE. 1648 15305
Boyer, Claude TYRIDATE, TRAGÉDIE 1649 16941
Corneille, Pierre AGÉSILAS, TRAGÉDIE 1666 20383
389
author title date n. words
Corneille, Pierre ANDROMÈDE, TRAGÉDIE. 1651 16439
Corneille, Pierre ATTILA, ROI DES HUNS, TRAGÉDIE 1668 18913
Corneille, Pierre LE CID, TRAGI-COMÉDIE 1637 18273
Corneille, Pierre LE CID, TRAGÉDIE 1682 18226
Corneille, Pierre CINNA ou LA CLÉMENCE D’AUGUSTE, TRAGÉDIE 1643 18001
Corneille, Pierre CINNA ou LA CLÉMENCE D’AUGUSTE, TRAGÉDIE 1682 18313
Corneille, Pierre CLITANDRE, COMÉDIE 1682 16383
Corneille, Pierre DON SANCHE D’ARAGON, COMÉDIE HÉROÏQUE 1649 19191
Corneille, Pierre LA GALERIE DU PALAIS ou L’AMIE RIVALE 1637 18411
Corneille, Pierre HÉRACLIUS, EMPEREUR D’ORIENT, TRAGÉDIE 1647 19614
Corneille, Pierre HORACE, TRAGÉDIE 1641 18327
Corneille, Pierre L’ILLUSION COMIQUE, COMÉDIE 1639 17623
Corneille, Pierre MÉDÉE, TRAGÉDIE 1639 15921
Corneille, Pierre MÉDÉE, TRAGÉDIE 1682 15897
Corneille, Pierre MÉLITE OU LES FAUSSES LETTRES, COMÉDIE 1633 20284
Corneille, Pierre MÉLITE, COMÉDIE 1682 18730
Corneille, Pierre LE MENTEUR, COMÉDIE N/A 19189
Corneille, Pierre LA MORT DE POMPÉE, TRAGÉDIE 1644 18583
Corneille, Pierre NICOMÈDE, TRAGÉDIE 1651 19229
Corneille, Pierre OEDIPE, TRAGÉDIE 1659 20532
Corneille, Pierre OTHON, TRAGÉDIE 1665 19119
Corneille, Pierre PERTHARITE ROI DES LOMBARDS, TRAGÉDIE 1653 19334
Corneille, Pierre LA PLACE ROYALE ou L’AMOUREUX EXTRAVAGANT, COMÉDIE 1637 15076
Corneille, Pierre POLYEUCTE MARTYR, TRAGÉDIE 1643 18642
Corneille, Pierre RODOGUNE, TRAGÉDIE 1647 19085
Corneille, Pierre SERTORIUS, TRAGÉDIE 1662 20041
Corneille, Pierre SOPHONISBE, TRAGÉDIE 1663 18885
Corneille, Pierre LA SUITE DU MENTEUR, COMÉDIE 1645 20353
Corneille, Pierre LA SUIVANTE, COMÉDIE 1637 17188
Corneille, Pierre SURENA GÉNERAL DES PARTHES, TRAGÉDIE 1675 18771
Corneille, Pierre THÉODORE, VIERGE ET MARTYRE, TRAGÉDIE CHRÉTIENNE 1646 19451
Corneille, Pierre TITE ET BÉRÉNICE, COMÉDIE HEROÏQUE 1671 18803
Corneille, Pierre LA CONQUÊTE DE LA TOISON D’OR, TRAGÉDIE 1661 22779
Corneille, Pierre LA VEUVE OU LE TRAÎTRE TRAHI, COMÉDIE 1634 19310
Corneille, Pierre LA VEUVE, COMÉDIE 1682 19823
Corneille, Thomas L’AMOUR À LA MODE, COMÉDIE. 1651 20384
Corneille, Thomas ARIANE, TRAGÉDIE 1672 18737
Corneille, Thomas LE BERGER EXTRAVAGANT, PASTORALE BURLESQUE. 1652 19730
Corneille, Thomas BRADAMANTE, TRAGÉDIE 1695 13758
Corneille, Thomas CAMMA, REINE DE GALATIE, TRAGÉDIE 1661 20702
Corneille, Thomas LE CHARME DE LA VOIX, COMÉDIE 1658 19698
Corneille, Thomas LE COMTE d’ESSEX, TRAGÉDIE 1678 17025
Corneille, Thomas LA COMTESSE D’ORGUEIL, COMÉDIE 1690 20697
Corneille, Thomas DARIUS, TRAGÉDIE 1659 20512
Corneille, Thomas DON CÉSAR D’AVALOS, COMÉDIE. 1661 19482
Corneille, Thomas LES ENGAGEMENTS DU HASARD, COMÉDIE. 1662 18651
Corneille, Thomas LE FEINT ASTROLOGUE, COMÉDIE 1651 20223
Corneille, Thomas LE FESTIN DE PIERRE, COMÉDIE 1677 21978
Corneille, Thomas LE GALANT DOUBLÉ, COMÉDIE. 1659 20794
Corneille, Thomas LE GEÔLIER DE SOI-MÊME, COMÉDIE. 1655 19230
Corneille, Thomas MAXIMIAN, TRAGÉDIE 1662 20419
Corneille, Thomas MÉDÉE, TRAGÉDIE EN MUSIQUE 1693 9659
Corneille, Thomas LA MORT D’ANNIBAL, TRAGÉDIE 1669 19903
Corneille, Thomas LA MORT D’ACHILLE, TRAGÉDIE 1673 18121
Corneille, Thomas LA MORT DE L’EMPEREUR COMMODE, TRAGÉDIE 1657 19953
Corneille, Thomas PERSÉE ET DÉMÉTRIUS, TRAGÉDIE. 1662 21509
Corneille, Thomas PYRRHUS, ROI D’ÉPIRE, TRAGÉDIE. 1661 21246
Corneille, Thomas STILICON, TRAGÉDIE 1664 21267
Corneille, Thomas THÉODAT, TRAGÉDIE 1673 18863
Corneille, Thomas TIMOCRATE, TRAGÉDIE 1662 19804
Donneau de visé, Jean LES AMOURS DU SOLEIL, PASTORALE. 1671 20898
Donneau de visé, Jean LA COCUE IMAGINAIRE, COMÉDIE 1660 6278
Donneau de visé, Jean L’EMBARRAS DE GODARD, OU L’ACCOUCHÉE, COMÉDIE 1668 8203
Donneau de visé, Jean LE GENTILHOMME GUESPIN, COMÉDIE 1670 7413
Donneau de visé, Jean LES INTRIGUES DE LA LOTERIE, COMÉDIE 1670 11179
Donneau de visé, Jean LA MÈRE COQUETTE, OU LES AMANTS BROUILLÉS, COMÉDIE 1666 11907
Donneau de visé, Jean LA VEUVE À LA MODE, COMÉDIE 1668 6123
Molière AMPHITRYON, COMÉDIE 1668 17603
Molière LE DÉPIT AMOUREUX 1656 19021
Molière L’ÉCOLE DES FEMMES, COMÉDIE. 1663 19377
Molière L’ÉCOLE DES MARIS, COMÉDIE 1661 12161
Molière L’ÉTOURDI ou LES CONTRE-TEMPS, COMÉDIE 1663 21708
Molière LES FÂCHEUX, COMÉDIE 1662 9607
Molière LES FEMMES SAVANTES, COMÉDIE 1672 19135
Molière MÉLICERTE, COMÉDIE PASTORALE HÉROÏQUE 1682 6328
Molière LE MISANTHROPE ou L’ATRABILAIRE AMOUREUX, COMÉDIE 1667 19590
Molière LA PRINCESSE D’ÉLIDE, COMÉDIE GALANTE 1664 10951
Molière SGANARELLE ou Le COCU IMAGINAIRE, COMÉDIE 1660 6877
Molière LE TARTUFFE ou L’IMPOSTEUR, COMÉDIE 1669 21088
Quinault, Philippe AMADIS, TRAGÉDIE 1684 4733
Quinault, Philippe ARMIDE, TRAGÉDIE. 1686 6811
Quinault, Philippe ATYS, TRAGÉDIE 1676 9181
Quinault, Philippe CADMUS et HERMIONE, TRAGÉDIE 1673 6840
390
author title date n. words
Quinault, Philippe LA COMÉDIE SANS COMÉDIE, COMÉDIE 1667 17813
Quinault, Philippe LES COUPS DE L’AMOUR ET DE LA FORTUNE, TRAGI-COMÉDIE 1655 16101
Quinault, Philippe LE DOCTEUR DE VERRE, COMÉDIE 1689 4324
Quinault, Philippe LE FANTOME AMOUREUX, TRAGI-COMÉDIE 1657 18411
Quinault, Philippe LES FÊTES DE L’AMOUR ET DE BACCHUS, PASTORALE 1672 4180
Quinault, Philippe LA GÉNÉREUSE INGRATITUDE, TRAGI-COMÉDIE PASTORALE 1656 16516
Quinault, Philippe ISIS, TRAGÉDIE en MUSIQUE 1687 7221
Quinault, Philippe LA MÈRE COQUETTE ou LES AMANTS BROUILLÉS, COMÉDIE 1665 19045
Quinault, Philippe PERSÉE, TRAGÉDIE 1682 8040
Quinault, Philippe PROSERPINE, TRAGÉDIE 1680 8342
Quinault, Philippe ROLAND, TRAGÉDIE EN MUSIQUE 1685 8640
Quinault, Philippe STRATONICE, TRAGI-COMÉDIE 1660 18813
Quinault, Philippe LE TEMPLE DE LA PAIX, BALLET 1685 3254
Quinault, Philippe THÉSÉE, TRAGÉDIE 1675 9200
A.2.2. Test
author title date n. words
Boyer, Claude FÉDÉRIC, TRAGI-COMÉDIE 1660 18625
Corneille, Pierre PULCHÉRIE, COMÉDIE HÉROÏQUE 1673 18884
Corneille, Thomas LES ILLUSTRES ENNEMIS, COMÉDIE 1657 20996
Donneau de Visé, Jean DÉLIE, PASTORALE. 1668 17723
Molière DON GARCIE DE NAVARRE, COMÉDIE 1682 19181
Quinault, Philippe AMALASONTE, TRAGI-COMÉDIE 1661 17932
391