=Paper=
{{Paper
|id=Vol-3834/paper97
|storemode=property
|title=Latent structures in french fiction
|pdfUrl=https://ceur-ws.org/Vol-3834/paper97.pdf
|volume=Vol-3834
|authors=Jean Barré
|dblpUrl=https://dblp.org/rec/conf/chr/Barre24
}}
==Latent structures in french fiction==
https://ceur-ws.org/Vol-3834/paper97.pdf
Latent Structures of Intertextuality in French Fiction:
How literary recognition and subgenres are framing
textuality
Jean Barré1,2,∗
1
École normale supérieure - Université PSL, 45 rue d’Ulm, Paris, 75005, France
2
LaTTiCe (Langues, Textes, Traitements informatiques, Cognition), 1 rue Maurice Arnoux, Montrouge, 92049, France
Abstract
Intertextuality is a key concept in literary theory that challenges traditional notions of text, signification
or authorship. It views texts as part of a vast intertextual network that is constantly evolving and being
reconfigured. This paper argues that the field of computational literary studies is the ideal place to
conduct a study of intertextuality since we have now the ability to systematically compare texts with
each others. Specifically, we present a work on a corpus of more than 12.000 French fictions from the
18th, 19th and early 20th century. We focus on evaluating the underlying roles of two literary notions,
sub-genres and the literary canon in the framing of textuality. The article attempts to operationalize
intertextuality using state-of-the-art contextual language models to encode novels and capture features
that go beyond simple lexical or thematic approaches. Our findings suggest that both subgenres and
canonicity play a significant role in shaping textual similarities within French fiction. These discoveries
point to the importance of considering genre and canon as dynamic forces that influence the evolution
and intertextual connections of literary works within specific historical contexts.
Keywords
literary history, intertextuality, computational literary studies, genres, canon, distant reading, cultural
analytics, natural language processing,
1. Introduction
How has textuality been shaped over time? Can we account for the dynamics of influence and
imitation in literary history? What roles have played the underlying structures, such as the
literary canon or the literary genres?
Intertextuality is the concept that centralizes these questions. It was introduced by Kristeva
[21], stating that “every text is a mosaic of quotations; every text is the absorption and trans-
formation of another text.” This perspective suggests that a text is no longer finite and that
its complete interpretation and understanding must involve unraveling the network of textual
relationships. A few years later, Barthes [8] further develops the definition, stating that “every
text is an intertext; other texts are present in it, at varying levels, in more or less recognizable
forms: texts of the previous culture and those of the surrounding culture; every text is a fabric
CHR 2024: Computational Humanities Research Conference, December 4–6, 2024, Aarhus, Denmark
∗
Corresponding author.
£ jean.barre@ens.psl.eu (J. Barré)
ç https://crazyjeannot.github.io/ (J. Barré)
ȉ 0000-0002-1579-0610 (J. Barré)
© 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
21
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
of past quotations.” This metaphor of fabric, recalling that the etymology of the word ”text”
refers to the idea of weaving, illustrates that all literary creation is in reality a web of collective
contributions. The text, conceived as a complex network of interwoven threads, represents a
meeting point where various influences, voices, and cultural traditions interlace.
Intertextuality has been defined differently by successive researchers and can have multiple
approaches1 . The main definition of intertextuality lies in the detection of explicit citation. Lit-
erary studies (computational or not) have naturally taken up tracking citations from one text
to another to understand phenomena of influence or rewriting in literature ([3], [10], [15]). A
second approach to intertextuality, which we will call “weak” to contrast with the first, con-
sists of a simple allusion or thematic and linguistic similarities between a given text and a set
of other texts. Computational methods have enabled the development of this approach in re-
cent years, with examples such as Ganascia [14] who conducted automatic detection of textual
fragments that evoke reuse due to their similarity by n-grams. This approach allows for the
identification of intertextual relationships between texts that may not be immediately apparent
through manual analysis.
In this way, computational literary studies can be extremely useful for analyzing the inter-
textual network in large corpora of text. Researchers can now identify patterns and trends
in the use of intertextuality across different genres, time periods, and cultural contexts. One
example of this is the seminal paper by Manjavacas, Karsdorp, and Kestemont [26], which pro-
poses a thematic and lexical approach to intertextuality. By grouping authors and texts that
share topical or lexical similarities, their aim was to evaluate whether it was possible to detect
intertextual phenomena from a given source (i.e. the Bible) in a particular corpus. In contrast,
our research views potential intertexts as endogenous to the corpus, and we will assess whether
the most significant intertexts are also part of the literary canon or not.
Before expanding on our experiment, we need to define two literary notions that may impact
intertextuality: the first one is the literary canon. The dynamics of prestige, defined as an au-
thor or a book being included in school curricula or reviewed in a prestigious literary journal,
can influence textuality. This is because the literary canon represents what is considered to
be paradigmatic. Previous computational research has uncovered disparities in textual content
between canonical and non-canonical works across various corpora and cultural backgrounds
([1], [34], [9], [6], [7]). These studies have shown that formalist and transcendent definitions
of the literary canon seem to be relevant across different genres and time periods. More inter-
estingly, Underwood found an increase in the probability of a work being canonized over time,
suggesting that canonical literary works constitute what Altieri [4] terms a “cultural grammar.”
In other words, canonical works function as foundational texts shaping the norms, values, and
conventions within a specific cultural tradition.
The second is the concept of literary genre. According to Genette [16], one of the five di-
mensions of intertextuality (or transtextuality, as Genette refers to it) is architextuality2 , which
is dedicated to literary genres. This dimension refers to the interconnectedness and interde-
pendence of various texts within a literary genre, encompassing “all that sets the text in rela-
tionship, whether obvious or hidden, with other texts”. Genette considers genre to be a crucial
1
For an overview of the concept, see Allen [2]’s work.
2
For an attempt to operationalize the concept at the passage level using a computational approach, see Barré [5]
22
component of a work’s overall meaning, arguing that the way in which texts reference and re-
late to one another can reveal important thematic and structural elements. Jauß [18] supports
this argument by introducing the concept of the “horizon of expectations” of the audience,
which may lead authors to adhere to certain expected norms and styles. As a result, intertex-
tuality is stronger between texts from the same genres, and genres constitute a structuring
element for intertextuality.
These two concepts can be seen as structuring elements of intertextuality. Their impact on
writing varies depending on their nature: unconsciously through the authority or influence of
certain works, or consciously through literary subgenres. Our goal is to develop a methodology
capable of automatically detecting the works and authors that have influenced the evolution of
the intertextual network, and to determine which of our two concepts has the greatest impact.
To achieve this, we rely on a massive corpus of over 12,000 French-language fiction.
As a proxy for prestigious literature, we will rely on novels that have been republished over
time. We were assuming that books that have been republished multiple times have either sold
well or are considered important enough to be reread or included in school curricula. The goal
is to evaluate the role of canonical works in shaping the literary tradition. Our hypothesis
is that the influence of canonical works on the intertextual network is more enduring than
the pace of change in the archive, which is heavily represented by texts from popular literary
sub-genres. In other words, we expect that canonical works will continue to shape the literary
tradition over a longer period of time, while the popularity and influence of the others may be
more short-lived.
Outline of the paper The structure of the paper is as follows: We start with a detailed
description of the method we used to model intertextuality (section 2), including the corpus
description (subsection 2.1), the metadata construction (subsection 2.2) and the operationaliza-
tion pipeline (subsection 2.3). Then, we present the results in section 3, including an evaluation
of the method (subsection 3.1), and an analysis of the main results (subsection 3.2), the individ-
ual (subsection 3.3) and collective ones (subsection 3.4). The article concludes with a discussion
and the perspectives opened up by this research (section 4).
2. Modeling Intertextuality
2.1. Corpus
Our corpus relies on a subset of the collection “Fictions littéraires de Gallica” [22] which rep-
resents 19,240 literary fictions drawn from Gallica, the extensive digitization initiative of the
National Library of France. It is made up of works initially categorized as prose literary fiction
spanning the period of 1600-1950. The collection’s massive scale can be attributed to the efÏ-
ciency of the French legal deposit system3 . This legal requirement mandates that publishers
submit copies of all published works to the National Library. As a result, it is estimated that
approximately 40% [22] of all novels published in France during the 19th century are available
in Gallica, based on the BNF catalog. Thus, it offers a representative sample of 19th and early
3
https://en.wikipedia.org/wiki/Legal_deposit#France
23
20th-century literature, encompassing various works such as lesser-known books and subgen-
res. Figure 1 shows the time distribution of the corpus, pointing the large peak of production
in the late nineteenth century, with almost 3.000 novels in the 1890s.
Figure 1: Number of novels over time
In France, texts generally enter the public domain 70 years after the death of their authors.
As the crucial date of 1950 approaches, the corpus of novels significantly decreases in size.
This is because works published after 1950 are still under copyright, and their full text is not
available for use without permission from the rights holders. This limitation affects the overall
size and diversity of the corpus, particularly for more recent literature.
The raw corpus contained several issues, such as the optical character recognition (OCR)
quality, the presence of complete works from an author, or multiple publications of the same
work. To address these issues, we removed all versions of complete works, as our focus was
on individual texts. For novels with multiple editions, we selected the first publication to have
the closest date associated with a text and its writing date (e.g., 6 editions of Hugo’s Toilers of
the Sea from 1866 to 1894). This allowed us to maintain a more accurate representation of the
texts in relation to their original publication dates.
Figure 2 shows the temporal distribution of the works in the corpus based on a calculation of
the proportion of correct words in each work. The goal is to detect works whose OCR quality
is too poor for computational analysis. This evaluation of OCR quality was done by creating
a dictionary of French words from a manually corrected set of literary texts. Each word in a
given text was then compared to this dictionary, allowing us to compute a proxy for the word
error rate metric.
As we can see, the texts before the beginning of the 19th century suffer from rather poor
OCR quality or orthographic standardization. On the other hand, the decrease at the end of
the century can be explained by a decline in the quality of paper used for certain works. This
lower paper quality can negatively impact the OCR process, resulting in a lower overall text
24
Figure 2: OCR quality over time
quality for these specific works. Filtering the corpus with an estimated OCR higher than 95%,
removing multiple publications and complete work from an author, we obtained a list of 12.176
novels.
2.2. Metadata construction
The notion of literary canon or the one of genre are not fixed and stable entities, but rather
dynamic and contested constructions that reflects the values and ideologies of specific cultural
and historic time. Therefore, it is important to approach these concepts with a critical and
historical perspective, taking into account the complex and shifting dynamics of literary pro-
duction and reception.
Setting aside the whole context of publication and reception of the work, which has an im-
pact on textuality but cannot be recovered, the idea is to reintroduce historical context using
proxies for literary reception and events that occurred during the life of the work. The chal-
lenge with this approach lies in the difÏculty of finding large-scale viable metadata. To address
this, we relied on previous research [6] which focused on canonicity in the context of contem-
porary reception in French literature. We then applied it to our corpus, resulting in 11,202 non
canonical elements and 1,083 canonical ones.
For the genre labels, we decided to focus on a single subgenre, as we required clearly defined
and coherent labels. We concentrated on “adventure novels” a dominant subgenre in the late
nineteenth century. For this purpose, we relied on the work of Letourneux [25], who defines
a specific period (1870-1930) for the genre, allowing him to identify its key constants: “The
importance of exotic settings [...] and the central role attributed to violent action, where the
hero faces the risk of death or at least physical peril.” We then generated synthetic metadata
using a binary SVM classifier [31], trained on Letourneux’s labels (with only 102 adventure
labels retrieved), and as a result, we identified 2,114 ”adventure novel” labels in the corpus
25
under study.
2.3. Operationalizing textuality
This article employs intertextuality as its theoretical foundation, drawing upon the work of
structuralists who initially developed the concept. Our aim is to merge their theoretical per-
spectives with empirical experiments on an extensive collection of literary texts. It can be
argued that computational literary studies inherently operationalize intertextuality through
quantitative text comparisons. These comparisons may encompass various textual aspects,
such as lexical, semantic, or thematic dimensions. However, a challenge arises from the ab-
sence of a clear definition of the specific textual components that intertextual theorists refer
to.
Computational literary scholars have been striving to develop methods for comparing dif-
ferent texts. To achieve this, they have utilized various techniques to extract information from
texts, such as topic-based (LDA, neural models), lexicon-based (BoW), or more semantic com-
parisons centered on entities or characters [20]. These methods were selected due to their high
interpretability, as basic machine learning techniques could determine topic or lexical impor-
tance for specific purposes. Nevertheless, certain textual dimensions, like word order or plot
progression, have often been disregarded to obtain more interpretable features. Preventing this,
the use of text embeddings has developed in the past few years, with methods such as Para-
graph Vectors [24] applied to prestige inquiries [13] or genre clustering ones [32]. However,
the ”black box” nature of these methods has deterred some researchers. At the level of passages,
embedding models have shown their strength in retrieving complex textual elements. The pri-
mary reason for using these methods is the uncertainty surrounding the specific information
we want to retrieve from the texts. Some literary analysis tasks may involve different textual
elements, formal or semantic dimensions, and the embedding model is supposedly encoding
these informations in its latent space.
In this study, we employ an encoder model based on a transformer language models, specif-
ically the M3-Embedding dense model [11], an open-source model developed by the Beijing
Academy of Artificial Intelligence. At the time of the experiment, this model led the multilin-
gual MTEB evaluation benchmark4 . We also chose this model because we could fine-tune it
on French literary language, a crucial factor for our research since many encoder models may
lack sufÏcient French language data in their training corpora, and particularly French literary
fiction. This model features an 8192 tokens window, which is significantly larger than that
of a more traditional BERT (512), allowing us to process a wide range of text lengths. This
extended window size enables the model to capture more contextual information, making it
more suitable for analyzing longer parts of literary works.
Consequently, we fine-tuned our model on a passage similarity task. We constructed the
training corpora by selecting 400,000 random passages from our corpus. The ”query” resulted
in a paragraph of ten sentences. For the ”positive” relation with the query, we used the ten
subsequent sentences. As the ”negative” relation, we chose ten random sentences from the
4
The MTEB benchmark stands for the Massive Text Embedding Benchmark. It is a large-scale evaluation framework
designed to assess the performance of text embedding models across a wide variety of natural language processing
(NLP) tasks. See paper [30] and url for more https://huggingface.co/blog/mteb
26
entire corpus. The underlying assumption is that authors maintain a consistent language in
their novels, especially when dealing with consecutive passages. We also replaced all proper
names with a specific token ”[PROPN]” to prevent the encoder model to cluster passages only
based on character names. Proper names of characters have been extracted using BookNLP-fr
[27], a state-of-the-art NER pipeline for literary entities. The model will iteratively enhance
its performance on the author attribution task by leveraging various cues, including formal
elements, thematic content, and stylistic features. As it undergoes training, it is expected to
develop a deeper understanding of these characteristics, allowing it to more accurately identify
the unique voice and style of different texts and authors.
After the fine-tuning, the goal was to infer a vector representation for each novel in our
corpus. Then, a challenge arises, as our encoder can only process a context window of 8096
tokens, while a typical novel contains around 100,000 tokens. We implemented the following
approach to handle this crucial step. For each novel in our corpus, we first randomly draw
100 passages of it, then we run our fine tuned encoder on each passage. Finally we take the
mean embedding of all passages from a novel to represent it as a unique embedding. Thus we
obtained 12.176 embeddings, one for each novel. As a distance metric, we opted for the cosine
similarity, since it is widely used in the NLP and CLS fields. Previous work managed to show
that the cosine similarity between pairs of word embeddings had a robust correlation with hu-
man similarity judgments [36]. However, when applying cosine distance to embeddings drawn
from contextualized language models like ours, previous research found that a small number
of ”rogue dimensions” had a disproportionate impact on the cosine similarity calculation [33],
[29]. To prevent this issue we ran a Standard Scaler normalization before running the cosine
distance between every pair of novels from our corpus.
3. Results
3.1. Sanity check
To validate our approach at each computational step, we implemented a sanity check using a
subset of one thousand novels of our corpus. For each novel, we generated five distinct random
representations by selecting random passages, computing their vector representations, and av-
eraging the resulting vectors. This produced five mean embeddings for each novel, resulting
in a total of 5000 embeddings. The goal of the sanity check is to ensure that, for any given em-
bedding, the four most similar embeddings (based on cosine similarity) correspond to the other
versions of the same novel. If the method fails to identify the four closest versions correctly,
we apply a penalty to the accuracy score. This test verifies whether our process of randomizing
passages, vectorizing them, and averaging the vectors maintains or not the integrity of each
novel’s representation.
Figure 3 compares for each number of draws the performance of the base model BGE-M3 and
the fine-tuned version. It shows that the fine-tuned model consistently outperforms the base
model, and that the number of draws is halved to complete the sanity check (50 vs 100). This
improvement highlights the effectiveness of fine-tuning, enabling an optimization of resource
consumption at inference which is crucial when planning to launch the model on 12,000 novels.
Overall, these results suggest that our computational approach is effective at capturing the
27
Figure 3: Sanity check accuracies between models and number of draws in novels
complexity of what constitutes a novel, while also being robust to the randomization of selected
passages of the works. The evaluation we implemented turned out to be relatively easy for the
model, as a 100% accuracy was achieved quickly. However, this allowed us to determine the
number of random draws needed to capture the complexity and uniqueness of a given novel.
3.2. Similarity over time
After retrieving the embeddings for each of our approximately 12,000 novels, a similarity matrix
is constructed, with as many rows as columns, where we measure the cosine similarity of each
text with all the others. On each row, works by the same author are set to NaN to prevent the
authorial idiolect from skewing the temporal analyses. Next, a similarity matrix is computed for
each text with every year in the corpus. To ensure statistical relevance, we apply downsampling
to each year, meaning that every year must have a minimum of 25 novels and a maximum
of 50 novels. Each text is then aligned with its publication year, centering the analysis at 0
for its date of publication, resulting in the graph presented in Figure 4. This downsampling
process is repeated ten times, and the standard error is displayed in the graph. We chose 30
years before and after because previous research showed it was a good window for measuring
literary change5 . A larger window would have also resulted in a significant loss of data in the
analysis.
The resulting graph measures the similarity between each text and its context of production.
A striking pattern emerges: a clear peak in similarity is observed during the year of the novel’s
publication. As the distance from the publication year increases, the textual similarity dimin-
ishes. This supports clearly previous research [17], where researchers provided quantitative
evidence for the concept of a literary ”style of the time”, highlighting a strong trend toward
more contemporary stylistic influences.
5
See [35] and Moretti [28]’s work on genres, their analysis focused on the cycle of change, considering a timeframe
of 25 to 30 years.
28
Figure 4: Similarity between texts considering dates before and after publication
Surprisingly, before the similarity peak at the time of publication, there is a noticeable down-
ward trend in the cosine similarity between -30 and -15 years before publication. The high
similarity 30 years before publication suggests that earlier works had a strong influence on the
author’s development. This influence decreases as the publication date approaches, reflecting
the gradual divergence from older literary models. The trend aligns with previous research
on the cohort effect, where an author’s writing is shaped by books from their formative years.
However, this goes beyond the analysis of this paper, and we have no evidence to support
this hypothesis here. Underwood, Kiley, Shang, and Vaisey [35] provided strong evidence of
cohorts driving literary change, though their work did not specifically explore their effect on
textual similarity.
3.3. Individual similarity over time
Then, we wanted to grasp the individual trends of specific novels to investigate how they move
in the textual similarity network. For these experiments, we took back the downsampled sim-
ilarity matrix computed for each text with every year in the corpus. Figure 5 represents four
different patterns of textual similarity over time. Figure 5a shows the cosine similarity plot for
Proust’s Le temps retrouvé, the final volume of his monumental À la recherche du temps perdu.
The graph reveals a significant peak around the time of its post-mortem publication in 1927.
The detected pattern is not the author’s signature, as texts by the same author are excluded
from the analysis. Instead, it reflects a kind of ”average language” or a broader intertextual
network at the linguistic level. This peak aligns deeply with the modernist focus on subjectiv-
ity and inner consciousness. Like many modernist writers of the time, Proust was influenced
by emerging psychological theories, especially those of Sigmund Freud, which emphasized the
exploration of the unconscious mind. In Le temps retrouvé, Proust explores the protagonist’s
memories, perceptions, and internal reflections, creating a narrative that is more concerned
with the flow of time as experienced subjectively than with external events or linear plot pro-
29
(a) Cosine similarity over time for Proust’s Le (b) Cosine similarity over time for Leroux’s Le
temps retrouvé mystère de la chambre jaune
(c) Cosine similarity over time for Hugo’s Les (d) Cosine similarity over time for Reybaud’s
Misérables Mézélie
Figure 5: Different patterns of individual textual similarity evolution over time
gression. Proust’s plot exhibits the highest level of noise among the four. This can be explained
by insights from Compagnon [12]’s work which provides a framework for understanding how
Proust’s À la recherche du temps perdu reflects a dialogue between past literary forms and the
new modernist literary movement that was gaining prominence in the early 20th century. This
concept of Proust being “between two centuries” helps explain why Proust’s work resonates
with both 19th-century novelists and modernist writers, which could explain the disparities in
similarity across time.
Figure 5b shows the similarity over time of Leroux’s Le mystère de la chambre jaune, pub-
lished in 1907. A huge increase in cosine similarity appears from the beginning of the twenti-
eth century to the end of the observed time period. This reflects perfectly the novel’s status as
a “proto-detective” story [23] that laid the groundwork for the whodunit detective genre. This
peak can be attributed to the rise of detective fiction following Leroux’s or Doyle’s innovations,
with authors building on their narrative techniques and conventions. As detective stories pro-
liferated in the early 20th century, the post-publication similarity of Le mystère de la chambre
30
jaune became increasingly apparent.
In Figure 5c, Victor Hugo’s Les Misérables offers a more intriguing pattern. If a slight peak
around its publication date in 1862 appears, it is much less pronounced than the decades that
follow between 1880 and 1900. This could be explained by the growing canonical status of
both Hugo and Les Misérables influencing subsequent generations of writers who either drew
inspiration from its themes or echoed its narrative structures. The vivid depictions of social
injustice, poverty, and the struggles of the lower classes in Les Misérables resonate with key
themes of Naturalism, despite the fact that they will take this social critique further, focusing
on deterministic views of human behavior influenced by environment and heredity. The drop
at the end of the graph is likely due to linguistic and thematic changes that are too significant
to maintain high textual similarity.
The last individual trend in Figure 5d is Mézélie, published in 1839 by Madame Charles Rey-
baud. The graph exhibits a gradual decline in cosine similarity over time. Despite her notable
success and popularity during her time, Reybaud’s works, including Mézélie, have largely faded
into obscurity and is now part of the so called “archive”.
3.4. Collective Structuring Frames of Textual Similarity
Our final experiments use the same type of analysis as in subsection 3.2, but we aimed to
explore the extent to which collective factors, such as genres or canonicity, could influence
the different trends in textual observed earlier. To that end we relied on metadata described
in subsection 2.2. To compare our samples (canon vs archive or adventure vs general) with
equal temporal cardinality, we implemented stratified sampling. We randomly selected the
appropriate number of elements from the general and non-canon samples to match those in
the adventure and canonical sets, while also maintaining the temporal distribution per decades
of these specific groups.
(a) Similarity between texts considering dates (b) Similarity between texts considering dates
before and after publication, Canon in blue, before and after publication, Adventure in
Archive in orange blue, in orange
Figure 6: Different patterns of collective textual similarity evolution over time
Figure 6a shows that while both canonical and archival novels follow a similar similarity
trend before publication, the curves diverge notably after the publication date. The pattern is
31
particularly striking, suggesting that canonical works tend to exhibit higher similarity scores
with later publications compared to archival works. This trend persists over time, indicating
that canonical works exert a more enduring influence on the intertextual network, whereas
non-canonical works have a more limited and shorter-lived impact.
Figure 6b shows a less striking pattern but remains interesting and statistically solid. Both
curves follow a similar trend before publication, yet they begin to diverge approximately 10
years before and up to 5 years after the publication date. This pattern suggests that genres
have a precise contextual existence within a specific historical moment. Following the decline
in similarity, we observe that adventure novels maintain a stronger similarity with later texts
(from +10 to +30 years). This is harder to interpret, but one hypothesis could be that many
works within the genre were canonized retrospectively, which, as seen in Figure 6a results in
a higher similarity with subsequent works.
4. Discussion
In this article, we demonstrated that both subgenres and canonicity function as collective struc-
turing frames for textuality. While our operationalization using language model encoders and
cosine similarity is undoubtedly limited in its ability to fully capture the complexity of a novel,
our approach nonetheless uncovered novel patterns of similarity within French fiction.
Firstly, we showed that canonical works tend to be more deeply integrated into the intertex-
tual network after their publication. This investigation sheds light on the collective represen-
tations that shape the cultural frameworks within which we operate. One way to understand
this phenomenon is through the concept of “cultural grammar,” as proposed by Altieri [4]. Ac-
cording to this view, canonical literary works serve as foundational texts that help to shape
the norms, values, and conventions within a particular cultural tradition.
Secondly, we highlighted that texts within the same genre share distinct textual similarities
during the relatively brief period when the genre is being defined. This could be attributed to
the commercial nature of genre, which contributes to increasing the similarity of certain texts
within a specific historical moment.
Ultimately, both canonical and subgenres works establish a set of shared references and
expectations that guide the production and reception of subsequent texts, and contribute to
the formation of a shared cultural imagination.
Narrowing the scope of study appears essential for future research in two ways: reducing the
number of works analyzed and focusing on passages rather than entire novels. This will help
to reestablish a more concrete foundation based on textual evidence. It would be interesting for
example, to examine these dynamics at the scale of a specific sub-genre, taking the example of
detective novels in order to track the dynamics of intertextuality within a highly codified and
well-defined sub-genre. This will also provide an opportunity to test the distant reading obser-
vations on the subgenre’s canonical novels, to determine whether the phenomenon persists at
a smaller scale.
32
Acknowledgments
Jean Barré’s PhD is supported by the EUR (Ecole Universitaire de Recherche) Translitteræ (pro-
gramme “Investissements d’avenir” ANR-10- IDEX-0001-02 PSL and ANR-17-EURE-0025).
References
[1] M. Algee-Hewitt, S. Allison, M. Gemma, R. Heuser, H. Walser, and F. Moretti.
“Canon/Archive. Large-scale Dynamics in the Literary Field”. In: Pamphlets of the Stan-
ford Literary Lab 11 (2016). url: https://litlab.stanford.edu/LiteraryLabPamphlet11.pdf.
[2] G. Allen. Intertextuality. Third edition. The new critical idiom. London New York: Rout-
ledge, Taylor & Francis Group, 2022. 262 pp.
[3] T. Allen, C. Cooney, S. Douard, R. Horton, R. Morrisse, M. Olsen, G. Roe, and R. Voyer.
“Plundering Philosophers: Identifying Sources of the Encyclopédie”. In: Journal of the
Association for History and Computing (2010). url: http://hdl.handle.net/2027/spo.33104
10.0013.107.
[4] C. Altieri. “An Idea and Ideal of a Literary Canon”. In: Critical Inquiry 10.1 (Sept. 1983),
pp. 37–60. doi: 10.1086/448236.
[5] J. Barré. “Détection automatique de l’architextualité dans le roman d’aventures”. In:
Humanistica 2024. Stylométrie. Association francophone des humanités numériques.
Meknès, Morocco, 2024. url: https://hal.science/hal-04559749.
[6] J. Barré, J.-B. Camps, and T. Poibeau. “Operationalizing Canonicity: A Quantitative Study
of French 19th and 20th Century Literature”. In: Journal of Cultural Analytics 8.3 (2023).
doi: 10.22148/001c.88113.
[7] J. Barré and T. Poibeau. “Beyond Canonicity: Modeling Canon/Archive Literary Change
in French Fiction”. In: CEUR Workshop Proceedings CHR2023. 2023, pp. 814–830.
[8] R. Barthes. “Texte (théorie du)”. In: Encyclopædia Universalis (1974). url: https://www.u
niversalis-edu.com/encyclopedie/theorie-du-texte/.
[9] J. Brottrager, A. Stahl, A. Arslan, U. Brandes, and T. Weitin. “Modeling and Predicting
Literary Reception. A Data-Rich Approach to Literary Historical Reception”. In: Journal
of Computational Literary Studies 1.1 (2022). doi: 10.48694/jcls.95.
[10] M. Büchler, G. Crane, M. Moritz, and A. Babeu. “Increasing Recall for Text Re-use in His-
torical Documents to Support Research in the Humanities”. In: Lecture Notes in Computer
Science. Springer Berlin Heidelberg, 2012, pp. 95–100. doi: 10.1007/978-3-642-33290-6
\_11.
[11] J. Chen, S. Xiao, P. Zhang, K. Luo, D. Lian, and Z. Liu. M3-Embedding: Multi-Lingual,
Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distilla-
tion. 2024. arXiv: 2402.03216 [cs.CL].
[12] A. Compagnon. Proust entre deux siècles. Paris: Éd. du Seuil, 2013.
33
[13] A. van Cranenburgh, K. van Dalen-Oskam, and J. van Zundert. “Vector space explo-
rations of literary language”. In: Language Resources and Evaluation 53.4 (Dec. 2019),
pp. 625–650. doi: 10.1007/s10579-018-09442-4.
[14] J.-G. Ganascia. “Détection automatique de phénomènes intertextuels”. In: Genesis 51
(2020), pp. 63–77. doi: 10.4000/genesis.5671.
[15] J.-G. Ganascia, P. Glaudes, and A. Del Lungo. Automatic Detection of Reuses and Citations
in Literary Texts. 2014. doi: 10.48550/arxiv.1404.2997.
[16] G. Genette. “Introduction à l’architexte”. In: Théorie des genres. Ed. by G. Genette and T.
Todorov. Points 181. Paris: Éd. du Seuil, 1986, pp. 110–148.
[17] J. M. Hughes, N. J. Foti, D. C. Krakauer, and D. N. Rockmore. “Quantitative patterns of
stylistic influence in the evolution of literature”. In: Proceedings of the National Academy
of Sciences 109.20 (2012), pp. 7682–7686. doi: 10.1073/pnas.1115407109. url: http://dx.d
oi.org/10.1073/pnas.1115407109.
[18] H. R. Jauß. Toward an aesthetic of reception. In collab. with P. D. Man. Trans. by T. Bahti.
Nachdr. Theory and history of literature 2. Minneapolis, Minn: Univ. of Minnesota Press,
2010. 231 pp.
[19] Jean Barré. fr_literary_bge_base. 2024. doi: 10.57967/hf/3255.
[20] L. Kohlmeyer, T. Repke, and R. Krestel. “Novel Views on Novels: Embedding Multiple
Facets of Long Texts”. In: 2021 Association for Computing Machinery. (2021).
[21] J. Kristeva. Sèméiotikè: recherches pour une sémanalyse. Points. Paris: Éditions Point, 2017.
[22] P.-C. Langlais. Fictions littéraires de Gallica / Literary fictions of Gallica. Version 1. 2021.
doi: 10.5281/zenodo.4751204.
[23] E. d. Lavergne. La naissance du roman policier français: du Second Empire à la Première
Guerre mondiale. Études de littérature des XXe et XXIe siècles 7. Paris: Classiques Garnier,
2009.
[24] Q. V. Le and T. Mikolov. Distributed Representations of Sentences and Documents. 2014.
arXiv: 1405.4053 [cs.CL]. url: https://arxiv.org/abs/1405.4053.
[25] M. Letourneux. Le roman d’aventures: 1870-1930. Limoges: Presses Universitaires de
Limoges et du Limousin, 2010.
[26] E. Manjavacas, F. Karsdorp, and M. Kestemont. “A Statistical Foray into Contextual As-
pects of Intertextuality”. In: Proceedings of the Workshop on Computational Humanities
Research (CHR 2020). Vol. 2723. CEUR Workshop Proceedings, 2020, pp. 77–96.
[27] F. Mélanie-Becquet, J. Barré, O. Seminck, C. Plancq, M. Naguib, M. Pastor, and T. Poibeau.
“BookNLP-fr, the French Versant of BookNLP. A Tailored Pipeline for 19th and 20th
Century French Literature”. In: Journal of computational literary studies (2024). doi: 1
0.26083/tuprints-00027396. url: https://tuprints.ulb.tu-darmstadt.de/id/eprint/27396.
[28] F. Moretti. Graphs, maps, trees: abstract models for literary history. London New York:
Verso, 2007. 119 pp.
34
[29] J. Mu and P. Viswanath. All-but-the-Top: Simple and Effective Postprocessing for Word
Representations. 2018. url: https://openreview.net/forum?id=HkuGJ3kCb.
[30] N. Muennighoff, N. Tazi, L. Magne, and N. Reimers. MTEB: Massive Text Embedding
Benchmark. 2022. url: https://arxiv.org/abs/2210.07316v3.
[31] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P.
Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher,
M. Perrot, and E. Duchesnay. “Scikit-learn: Machine Learning in Python”. In: Journal of
Machine Learning Research 12 (2011), pp. 2825–2830.
[32] O. Sobchuk and A. Šeļa. Computational thematics: Comparing algorithms for clustering
the genres of literary fiction. 2023. arXiv: 2305.11251 [cs.CL].
[33] W. Timkey and M. van Schijndel. “All Bark and No Bite: Rogue Dimensions in Trans-
former Language Models Obscure Representational Quality”. In: Proceedings of the 2021
Conference on Empirical Methods in Natural Language Processing. Ed. by M.-F. Moens, X.
Huang, L. Specia, and S. W.-t. Yih. Online and Punta Cana, Dominican Republic: Associ-
ation for Computational Linguistics, 2021, pp. 4527–4546. doi: 10.18653/v1/2021.emnlp-
main.372.
[34] T. Underwood. Distant horizons: digital evidence and literary change. Chicago: The Uni-
versity of Chicago Press, 2019. 206 pp.
[35] T. Underwood, K. Kiley, W. Shang, and S. Vaisey. “Cohort Succession Explains Most
Change in Literary Culture”. In: Sociological Science 9 (2022), pp. 184–205. doi: 10.15195
/v9.a8.
[36] B. Wang, A. Wang, F. Chen, Y. Wang, and C.-C. J. Kuo. “Evaluating word embedding
models: methods and experimental results”. In: APSIPA Transactions on Signal and Infor-
mation Processing 8.1 (2019). doi: 10.1017/atsip.2019.12.
A. Data, code and model availability
We have made the data and code available on GitHub6 and released the fine-tuned model and
corpus on HuggingFace.7
B. Finetuning supplement
Figure 7 shows a rapid decrease in loss at the beginning of the training, from 1.3 to around
0.5 during the first epoch, indicating that the model is quickly learning to distinguish positive
passages (subsequent passage) from negative ones (random passage). After this initial drop,
the loss decreases more slowly and stabilizes around 0.4 after 2 epochs, suggesting the model
is converging. The small fluctuations are likely due to the stochastic nature of the gradient
updates, as mini-batches vary in content.
6
https://github.com/crazyjeannot/CHR_latent_structures
7
https://huggingface.co/crazyjeannot/fr_literary_bge_base[19]
35
Figure 7: Loss fluctuations during finetuning
36