Jean Barré

@jbarre.bsky.social

486 Followers 534 Following

PhDone ! in Computational Literary Studies at LaTTiCe lab & École normale supérieure in Paris. I study French genre fiction (1860–1945) using quantitative and network-based approaches. https://jeanbarre.eu/

Horizontal bar chart of titles per author for all 569 authors in the final corpus. A handful of headliners top the list — Jules Verne with 60 titles, Paul Féval with 48, Ponson du Terrail with 41 — followed by an extremely long tail: 68.5% of authors appear with a single title, illustrating the mass, largely forgotten production of the genre.
Flowchart of the two-stage pipeline. Stage 1: a seed list of 101 adventure novels cited by Letourneux (2010) plus negatives sampled from Gallica train a supervised classifier, which produces predictions on the full archive. Stage 2: a k-NN graph refinement (k=10, threshold 80%) removes outliers and adds infiltrators, yielding a final corpus of 1,391 adventure novels.
Network visualization of a k-nearest-neighbors graph over 10,998 French novels from Gallica. Each node is a novel, connected to its most similar neighbors by cosine similarity of text embeddings. Orange nodes are predicted adventure novels, green nodes are randomly sampled non-adventure novels.
Six people (Thzierry Poibeau, Thomas Conrad, Marie-Ève Thérenty, Me, Claire Barel-Moisan &Matthieu Letourneux) stand in a row smiling in front of a projection screen for a group photo after a PhD defense. In the centre, a young man with curly dark hair in a dark blazer holds a large bouquet wrapped in kraft paper. The screen behind them reads "Soutenance de thèse" with ENS-PSL and Lattice logos, and the thesis title about computational methods for modelling genre fiction in French-language literature.
Page de couverture de la thèse de doctorat de Jean Barré (PSL / ENS, laboratoire Lattice), intitulée « Modéliser la littérature de genre : Méthodes computationnelles et évolution des genres dans la fiction de langue française ». Soutenance prévue le 5 juin 2026. Jury : Marie-Ève Thérenty (Présidente, Montpellier III), Evelyn Gius (Rapporteure, TU Darmstadt), Matthieu Letourneux (Rapporteur, Paris-Nanterre), Claire Barel-Moisan (Examinatrice, ENS Lyon / CNRS), Ted Underwood (Examinateur, Illinois Urbana-Champaign), Thomas Conrad (Co-encadrant, ENS-PSL), Thierry Poibeau (Directeur, ENS-PSL / CNRS).
Box plot showing accuracy scores for all combinations of 1 to 17 ontological character attribute classes. Three trend lines track maximum (green), mean (red), and minimum (orange) accuracy. The key finding: maximum accuracy peaks at ~0.96 with just 4–5 carefully selected classes, while using all 17 classes simultaneously yields no improvement over the mean (~0.64). A small, well-chosen subset of ontological dimensions outperforms the full representation — more is not better.
Yuri and I standing by a literal cannon while we talk about canonicity in the Luxembourg city casemates
Regression coefficients for subgenres, publication periods & “canonical” status—bars show effect sizes relative to the baseline. Genres explain the lion’s share of the shift!
Smoothed correlations of passé simple with past tenses (imparfait, plus-que-parfait) in warm colors vs. present tenses (présent, futur, passé composé) in cool colors. Notice their uncoupling around WWII—hinting at a shift from narrated to "discussed" modes of storytelling.
The graph shows cosine similarity for "canon" (blue) and "non-canon" (orange) works over time, with both increasing around the publication date (year 0), but "canon" maintains higher similarity values after publication compared to "non-canon," which declines more steeply.
Prompt: Adventure scene with a character in front of steep mountains, ready to fight for her.is life in a both hostile and amazing area