Narratology

Debate over the centrality of character to narrative is as old as poetics. In Poetics I.VI, Aristotle subordinates character to plot: “The Plot, then, is the first principle […] Character holds the second place” (Aristotle 1996, 335 BCE). Yet modern dramaturgy and screenwriting famously invert this hierarchy: “What the reader wants is fascinating, complex characters” (McKee Reference McKee1997, 100). This tension has framed subsequent theorizing.

A durable heuristic is Forster’s (Reference Forster1988) opposition between flat and round characters (constructed around a single idea vs. capable of convincing surprise). But across the 20^th century, narratology diversified the very definition of character. Formalist and structuralist perspectives treat the figure as a textual/functional unit: a “being of paper” (Barthes Reference Barthes1966) or a role in an action grammar, from Propp’s (Reference Propp1928) dramatis personae to Greimas’s actantial schema (Greimas Reference Greimas1966). Barthes’s S/Z radicalizes this by replacing the person with a web of semes tied to a proper name (Barthes Reference Barthes1970). French poetics of description (Hamon) makes this operational: descriptive cues (prosopographie, éthopée, topographie and chronographie) distribute traits across a novel’s personnel, scripting recognizability and difference (Hamon Reference Hamon1998, Reference Hamon1993).

Reception-oriented and cognitive approaches shift emphasis from what a character is to how readers construct one. For Jouve (Reference Jouve1992), character is an effet-personnage: an emergent construct at the intersection of textual cues and readerly activity, addressed to an implicit reader and invested by a real reader. Margolin (Reference Margolin1983) offers a precise account of the inferential process: characterization (ascription of properties from actions, descriptions and setting) builds up to character-building (stabilizing a constellation of traits through classification, hierarchization and temporal scope).

Possible-worlds narratology further insists that characters are processed as entities in structured storyworlds rather than mere bundles of signifiers (Ryan Reference Ryan1992). Syntheses, such as Jannidis (Reference Jannidis and Hühn2013) (and Jannidis Reference Jannidis2008), align these strands by defining characterization as direct/indirect attribution of properties to referential anchors, articulating mimetic (human-likeness), thematic (symbolic roles) and synthetic (constructedness) dimensions. Three knowledge sources drive reader inferences: (a) a basic type for sentient beings (inside/outside and theory of mind), (b) character models/types (e.g., femme fatale and hard-boiled detective), and (c) encyclopedic knowledge (world/genre conventions) (Jannidis Reference Jannidis and Hühn2013). Texts ascribe not only psychological or social attributes but also physiological and spatiotemporal properties; characterization may be explicit or implicit, local or global, and is constrained by the level of narration and the cues available to the reader (Jannidis Reference Jannidis and Hühn2013; Margolin Reference Margolin1983).

Computational narratology

Over the past decade, a significant share of computational narratology research has focused on the question of characters. Most notably Bamman, Underwood, and Smith (Reference Bamman, Underwood and Smith2014), who propose a Bayesian mixed-effects model to infer latent character types from a large corpus of English novels. Their study demonstrates that computational methods can identify underlying character personas by analyzing linguistic patterns associated with character actions, descriptions or possessions. Following this seminal approach, the computational literary studies community has developed specialized tools to automate the extraction of character information. Notable examples include BookNLP for English (Bamman, Lewke, and Mansoor Reference Bamman, Lewke and Mansoor2020; Bamman Reference Bamman2021), as well as its counterparts adapted to other languages, such as French (Mélanie-Becquet et al. Reference Mélanie-Becquet, Barré, Seminck, Plancq, Naguib, Pastor and Poibeau2024), German (Ehrmanntraut, Konle, and Jannidis Reference Ehrmanntraut, Konle and Jannidis2023) and Dutch (van Zundert, van Cranenburgh, and Smeets Reference van Zundert, van Cranenburgh and Smeets2023).

These pipelines have significantly advanced the automatic extraction of character-related information across multiple stages. They enable more accurate identification of character mentions, offer improved coreference resolution tailored to long literary texts (Bourgois and Poibeau Reference Bourgois and Poibeau2025) and enable the extraction of various character-related linguistic features, such as speech acts, verbs and adjectives. In addition, they provide computational representations of characters that can be leveraged for a wide range of applications.

Various methods have been proposed for extracting information and constructing representations of characters. Attribute extraction methods are most commonly syntax-based, as implemented in pipelines, such as BookNLP or FanfictionNLP (Yoder et al. Reference Yoder, Khosla, Shen, Naik, Jin, Muralidharan and Rosé2021). These approaches often rely on hand-crafted rules over part of speech tags and parsing trees to identify features, such as verbs for which a character is the subject/object, possessive dependents and predicatives.

Regarding the aggregation of extracted attributes into computational character representations, different strategies have been explored. A common approach is to merge all extracted features into a single representation for each character, either by constructing a frequency-based vector or by averaging the embeddings of individual attributes (e.g., bitstring (Bamman, Underwood, and Smith Reference Bamman, Underwood and Smith2014), GloVe (Flekova and Gurevych Reference Flekova and Gurevych2015) or BERT-like models (Barré et al. Reference Barré, Seminck, Bourgois and Poibeau2025)).

While such aggregation provides a compact summary of character attributes, it also comes with important limitations. By collapsing heterogeneous information into a single unified representation, much of the nuance of individual attributes is lost, and the resulting vectors can prove difficult to interpret. Not all attribute types are equally informative, and certain syntactic classes may be more relevant depending on the downstream task.

To address this issue, Yang and Anderson (Reference Yang and Anderson2024) propose maintaining separate representations for each class of attributes. By averaging attribute embeddings from a large encoder–decoder transformer Language Model (T5-11B), they represent each character as a set of five high-dimensional vectors. This design encompasses the intuition that some feature types may be more useful than others for specific applications, such as similarity assessment. In their work, they also suggest that concatenating these complementary vectors into a single representation could be a direction for future works. This approach represents a promising step forward in capturing richer character information. However, syntactic classes, while convenient for attribute extraction, still bundle together very heterogeneous types of attributes. For instance, the possessive category may indiscriminately group clothing (hat and coat), vehicles (car and train), human relations (their children) or more abstract notions, such as dreams, power or homeland. Such diversity makes the resulting representations more difficult to interpret and limits their usefulness for fine-grained analyses. From the perspective of cultural studies or computational humanities, where research often targets specific kinds of attributes, this lack of granularity constitutes a significant limitation.

Along similar lines, Mian et al. (Reference Mian, Subbiah, Marcus, Shaalan and McKeown2026) argue that character importance itself should not be treated as a single scalar notion. Drawing on narratological theory, they model character significance as a composition of distinct narrative dimensions (e.g., action, speech, interiority, narration and being talked about by others), and show that different definitions of centrality yield markedly different character rankings. Their work highlights that collapsing heterogeneous narrative evidence into a single score obscures the internal structure of character prominence.

Another line of research leverages generative large language models to construct character representations, either in the form of paragraph-style summaries (Brahman et al. Reference Brahman, Huang, Tafjord, Zhao, Sachan and Chaturvedi2021) or as structured “character sheets.” These methods typically proceed in one of two ways: either by directly prompting the LLM to produce the desired character representation, or, as proposed by Gurung and Lapata (Reference Gurung and Lapata2024), by adopting a question–answer framework that incrementally gathers information about a character across different dimensions, such as dialogue, personality and physical attributes, knowledge, plot role and motivations.

While these representations tend to be rich and readily interpretable by human readers and researchers, they are also computationally expensive to generate and do not easily scale to large-scale computational literary studies involving thousands of full-length books.

Evaluating character representations remains inherently complex. Unlike standard machine-learning evaluations that rely on held-out gold standards, defining what makes a “good” representation of a literary character is difficult because such representations are multidimensional and involve a certain degree of subjectivity. Consequently, evaluation is necessarily approximate, capturing only certain aspects of the representation. Some evaluation protocols are explicit about the dimension they target (e.g., narrative role, social rank and psychological profile), whereas others offer broader assessments against expert judgment without clearly specified criteria.

A first family of approaches relies on supervised evaluation, where representations are tested by training a downstream model to predict predefined classes. The nature of these classes varies across studies. For instance, Flekova and Gurevych (Reference Flekova and Gurevych2015) frame the task as personality prediction, distinguishing extraverted from introverted characters. Srinivasan and Power (Reference Srinivasan and Power2022) propose a role-based ternary system, labeling characters as protagonist, antagonist or supporting figures. Likewise, Chen, Chen, and Chen (Reference Chen, Chen and Chen2019) classify characters into core, secondary and peripheral categories derived from character networks, while Jahan, Mittal, and Finlayson (Reference Jahan, Mittal and Finlayson2021) focus on identifying character roles from Vladimir Propp’s morphology of the folktale (Propp Reference Propp1928).

Another evaluation approach is masked-character identification, which tests whether a human reader (or a generative LLM) can recover the identity of a character from an anonymized description generated by the character representation pipeline (Brahman et al. Reference Brahman, Huang, Tafjord, Zhao, Sachan and Chaturvedi2021; Gurung and Lapata Reference Gurung and Lapata2024).

Some studies adopt unsupervised similarity-based evaluation, which does not require training a classification model. One approach is triplet-based comparisons, in which literary experts annotate triples of characters indicating that character A is more similar to character B than either A or B is to C. Representations are then evaluated on how well they reflect these expert-defined similarity judgments (Bamman, Underwood, and Smith Reference Bamman, Underwood and Smith2014).

Yang and Anderson (Reference Yang and Anderson2024) proposed the AustenAlike benchmark, which evaluates character representations along three notions of similarity: structural, social and expert-defined. Characters are grouped by narrative role, social rank or literary expert judgment, and character representations are compared against these groupings. The conclusions from this work illustrate the current state of character-centered research in computational narratology. While feature-based representations have improved, they still do not fully capture the richness of literary characters: even recent models have difficulty matching expert judgments or representing the multidimensional nature of character similarity.

The gap is conceptual as much as technical. Most feature-based representations remain syntactic or latently lexical: they can say that two figures are close in a vector space, but not along which dimensions of characterization they are. Similarity between characters is not a single dimension; it is a bundle of relations that cut across actions, affect, cognition, social roles, physical description, possessions and more. Without a shared inventory of attribute types, these dimensions get muddled and similarity becomes opaque.

Our response is to move from syntactic proxies to semantic classes. We introduce an explicit ontology based on the work of Jannidis (Reference Jannidis and Hühn2013) using the three theoretic axes that define characters: anthropological, biological and psychological. Each of these three hyper-class axes is further divided into fine-grained subclasses, resulting in a total of 18 interpretable categories for character description. The result is a structured, multi-granular profile for each character: comparable within and across classes, scalable over large corpora and meaningful from a narratological perspective.

Ontology definition

Our ontology is informed by narratology and designed for computational literary studies. Following syntheses, such as Margolin (Reference Margolin1983) and Jannidis (Reference Jannidis and Hühn2013), we assume that characterization proceeds by attributing properties to referential anchors, and that these properties cluster along three macro-domains commonly mobilized in analysis: physical/bodily description and states (Hamon’s prosopographie and descriptive scaffolding), psychological/mental life (affects, cognition and dispositions) and anthropological positioning (roles, relations and identity markers). To make these operational, we articulate this three macro-classes into 18 mutually exclusive and collectively exhaustive fine-grained categories. Two principles guide the design of our ontology: clarity and distinctness. This means that each class has a precise definition to minimize overlap and that adjacent classes are separated by concrete decision rules. For example, there is a difference between physical objects and body parts: the first is part of the character’s body, whether the second is an object that can be carried by it.

Our ontology can capture a large spectrum of attribute types that covers the descriptive surface exploited by readers while remaining legible for large-scale computation.

Table 1 presents the 18 classes and their working definitions used for the manual annotation.

Table 1. Ontology of character attributes grouped into four macro-classes

Operationalizing

Our aim is the automatic, ontology-aware classification of character attributes. The first step toward this goal is to identify characters and extract their attributes. As noted earlier, several pipelines already exist for this purpose. For this work, we build upon the off-the-shelf Propp-fr pipeline (https://github.com/lattice-8094/propp), designed to process full-length novels. It allows us to extract character mentions (pronouns, proper nouns and noun phrases), coreference chains and attributes derived from syntactic dependency rules.

To operationalize our ontology for large-scale computational studies, we must reclassify those syntactic-based attributes into our fine-grained ontology classes. Exhaustive manual annotation is unrealistic given the scale of the task involving thousands of books and millions of attributes.

One possible approach would be to construct an exhaustive dictionary of attributes for each class (e.g., all words referring to body parts). To accelerate this process, one could leverage existing semantic resources, such as WordNet or VerbNet, as done by Flekova and Gurevych (Reference Flekova and Gurevych2015), who assign the most frequent sense to each word. This approach has the limitation of being unable to account for polysemy, as each word is assigned a fixed class, whereas in reality, words often have radically different senses depending on their context, especially in fictional writing.

For example, in French, the verb filer can have many meanings:

• • Il file de la laine.

  (He spins wool.)

• • En voyant la police, elle a filé.

  (When she saw the police, she ran away.)

• • Elle m’a filé un coup de main pour tondre la pelouse.

  (She gave me a hand mowing the lawn.)

• • Ce garçon file un mauvais coton.

  (That boy is headed for trouble.)

To overcome this issue, models that incorporate contextual information are required. Neural classifiers based on contextual embeddings provide a particularly suitable solution (Huang et al. Reference Huang, Sun, Qiu and Huang2019). These models generate dynamic representations of words conditioned on their surrounding context, thereby capturing polysemy and subtle usage patterns that static embeddings or hand-curated lexicons cannot.

We therefore adopt a contextual neural classification approach. To enable ontology-based labeling of extracted attributes, the classifier must first be trained. This necessitates a gold-standard dataset: a manually annotated collection of attribute instances, each assigned to the appropriate class within our ontology. Such a resource provides the supervision required for the model to learn robust, context-sensitive mappings between character attributes and fine-grained ontological categories.

Gold dataset

To ensure that the classifier will be able to generalise effectively, and be useful for downstream computational literary studies, the gold dataset must contain a representative pool of character attributes drawn from diverse texts.

For this purpose, we apply the Propp-fr pipeline to the full Chapitres corpus (Leblond Reference Leblond2022), which contains 2,960 full-length French novels published between 1811 and 2020. This corpus provides diversity across authors, periods, literary subgenres, narrative perspectives and representation of the literary canon.

The corpus contains 234M tokens, from which 25M attributes were extracted (10.83%). (This ratio of approximately 1:10 between character attributes and total tokens further underscores the character-centric nature of narratives, providing additional justification for the approach adopted in this work). The distribution of attribute types is as follows: agents 49.5 percent, patients 12.0 percent, possessives 27.3 percent and modifiers 11.2 percent.

Neural classifier training and evaluation

The following section presents the training setup and evaluation of a neural classifier for automatic attribute categorization.

Character representation vectors

Following Margolin’s (Reference Margolin1983) approach to character conceptualization, we aggregate classified attributes into unified, vectorial representations of characters. For each ontological category (17 fine-grained and 4 coarse-grained), in-class attribute embeddings are aggregated and averaged to produce a single 1,024-dimensional vector per category.

The average vectors are then processed through the following steps:

1. 1. Normalization: Each mean vector is L2-normalized to unit length. This ensures that subsequent operations, such as PCA, are insensitive to the overall magnitude of the embeddings, while preserving directional information on a unit hypersphere.

2. 2. Principal component removal: To reduce dominant, non-informative directions shared across characters, the first principal component (PC) of all non-zero normalized vectors for a given attribute class is computed. Each vector is then projected orthogonally to this component, highlighting unique, discriminative characteristics for each character.

3. 3. PCA-based dimensionality reduction: For each attribute class, PCA is applied to all non-zero vectors after PC removal. The top K PCs (e.g., $K=15$ ) are retained, and each vector is projected onto this reduced subspace.

The resulting centered and reduced vectors provide a compact, discriminative representation of a character for each attribute class, suitable for downstream tasks, such as clustering, similarity analysis or classification.

To enable comparison, we also consider two simpler baselines. The first, following Yang and Anderson (Reference Yang and Anderson2024), computes separate average embeddings for each syntactic category extracted by the Propp-fr pipeline (agent, patient, modifier and possessive), yielding four additional vectors per character. The second baseline collapses all attributes into a single undifferentiated vector by averaging embeddings across all attribute instances, regardless of category. For both baselines, the resulting vectors are centered and reduced following the same post-processing steps described above, ensuring comparability across methods.

Bamman, Underwood, and Smith (Reference Bamman, Underwood and Smith2014) French benchmark

The Bamman, Underwood, and Smith (Reference Bamman, Underwood and Smith2014) baseline performance steadily decreases from Class A to Class D, reflecting the increasing difficulty of the triplets. Despite this decline, overall performance remains strong, with an overall accuracy of 64 percent (where random choice would yield $\approx $ 0.33). This provides a meaningful reference for evaluating our character representations (see Table 2).

Table 2. Accuracy results on the French adaptation of the Bamman, Underwood, and Smith (Reference Bamman, Underwood and Smith2014) similarity benchmark

Note: We display the original baseline results alongside our evaluations. The reported baseline is computed as the weighted average of the best-performing model for each class, selected from the three models presented in Bamman, Underwood, and Smith (Reference Bamman, Underwood and Smith2014). For each attribute group, we also report the average ratio of all character attributes used in constructing the corresponding vector. Per-class performance (A–D) and overall accuracy are shown.

Turning to the syntactic attribute classes, performance is mixed. The modifiers class performs particularly well (77% accuracy), while patient lags behind (42%). Interestingly, when all attributes from Propp-fr are aggregated into a single vector (all syntactic), performance rises to 65 percent, surpassing most individual syntactic classes and slightly above the baseline.

For the coarse-grained ontological classes, overall accuracy is relatively consistent (50%–58%) but remains lower than for some of the syntactic-based categories, suggesting that broad semantic categories do not capture character similarity as effectively. Notably, the non-informative class behaves as intended: aggregating attributes labeled as non-informative yields very low accuracy (27%), even below random chance ( $\approx $ 0.33). This confirms that these attributes contribute little to character discrimination, and removing them helps focus on more meaningful semantic signals.

In contrast, the fine-grained ontological classes exhibit a highly diverse pattern. Some classes, such as possession (12%) and communication (23%), perform poorly, while others, like motion, personality traits and affect emotion, achieve high overall accuracy (69%–73%), in some cases surpassing the baseline, but not quite reaching the best performing syntactic-based vector from modifiers attributes.

Combining fine-grained ontological vectors

Individual fine-grained ontological vectors capture specific dimensions of character similarity, but human judgments often integrate multiple, distinct aspects simultaneously. A human annotator might consider personality, actions, social relations and other traits together when assessing which characters are most alike.

To model this multi-dimensional perspective, we exhaustively explored all combinations of the 17 fine-grained ontological classes (from pairs to the full set). The total number of combinations is

$$\begin{align*}\sum_{k=2}^{17} \binom{17}{k} = 131{,}054. \end{align*}$$

For each combination, we concatenate all vectors and similarity scores were recomputed on our benchmark triplets. We then grouped the results by the number of classes included and computed the distribution of accuracy scores, as well as average, minimum and maximum accuracy for each number of classes (1–17).

The full results of this experiment are displayed in Appendix B. Mean accuracy steadily increases as more classes are combined, rising from 0.49 for single-class vectors to a peak of approximately 0.646 around 9–10 classes. Beyond this point, adding additional classes yields diminishing returns and a slight decrease in mean accuracy, reflecting the inclusion of lower-performing or noisy attributes.

Maximum accuracy is achieved with carefully selected small combinations of 3–5 classes, reaching up to 0.96 on the French adaptation of the Bamman, Underwood, and Smith (Reference Bamman, Underwood and Smith2014) similarity benchmark.

As shown in Table 2, the optimal combinations of fine-grained ontological classes confirm the value of selective mixing. For example, pairing affect emotion + personality traits already yields 85 percent accuracy, surpassing any individual class. Adding a third class (motion) increases accuracy to 88 percent, while carefully chosen four- and five-class mixes (affect emotion + personality traits + action + physical objects and affect emotion + personality traits + action + motion + physical objects) reach 96 percent.

This result highlights that a judicious mixture of complementary ontological dimensions can strongly enhance character similarity predictions, approaching multi-dimensional human-like judgment. In contrast, including all possible classes does not further improve performance, as lower-quality attributes can dilute the signal.

Finally, the correlation between attribute usage ratio (frequency of appearance in character vectors) and accuracy is weak. Some infrequent attributes, such as personality traits (4.52%), are highly informative, achieving 73 percent accuracy, whereas more common attributes, like agent (54.9%), perform relatively poorly; the same pattern is observed for coarse-grained ontological classes. This indicates that informativeness is not proportional to frequency and underscores the importance of careful attribute selection when constructing effective character representations.

CharaSim-fr benchmark

Evaluation and results

The evaluation on CharaSim-fr broadly confirms the tendencies observed in the first benchmark, while also revealing notable differences.

Syntactic-based vectors again provide a strong signal, with accuracies ranging from 53 percent to 71 percent. As shown in Table 3, the aggregated syntactic vector achieves 74 percent, outperforming all individual categories. Among the fine-grained ontological classes, results are more heterogeneous: possession (29%) and capability/obligation (35%) perform poorly, while action, affect emotion, communication and body parts reach 62 percent to 68 percent. Interestingly, the strongest ontological classes here are not the same as in the previous benchmark. Moreover, the vector built from attributes labeled as non-informative unexpectedly reaches 68 percent, suggesting that some of these features may still capture relevant signals in this context.

Table 3. CharaSim-fr benchmark results summary

Note: Full results in Appendix C.

Combining ontological classes again improves performance, though the optimal mixes differ from those observed previously. In CharaSim-fr, the best combination (action + personality traits + physical traits + human relations) yields 82 percent accuracy, whereas in Bamman, the most effective mixes were centered on emotional attributes. This divergence may reflect cultural or corpus-specific differences, or differences in how literary experts conceive the notion of similarity between characters.

Qualitative analysis

Unpacking the model behavior, we look closely at per-dimension similarities for each triplet. For each ontological class $k,$ we retrieve pairwise similarities $s_{AB}^k, s_{AC}^k, s_{BC}^k$ (cosine on k-restricted vectors). We call the dimension-wise outsider the character whose average similarity to the other two is minimal in class k.

Easy case

Triplet: A = Julien Sorel (Stendhal), B = Lucien Leuwen (Stendhal), C = Emma Bovary (Flaubert)

Predicted outsider: C = 16, B = 1

Table 4 shows that $s_{AB}^k$ dominates across 16/17 classes, while both $s_{AC}^k$ and $s_{BC}^k$ are markedly lower or negative-leaning, yielding C as the outsider in almost every dimension. The strongest margins occur in sociological, action, perception, motion and physical traits, which coheres with narratological expectations: Julien and Lucien are Stendhalian ambitieux (male aspirants in Restoration/July Monarchy France) with comparable social roles and activity patterns, whereas Emma’s profile diverges along gendered social constraints and affective-aesthetic desire.

Table 4. Cosine similarities over fine-grained ontological classes

Note: A: Julien Sorel from Stendhal — B: Lucien Leuwen from Stendhal — C: Emma Bovary from Gustave Flaubert.

The lone dimension favoring B (environment) is driven almost entirely by the feature salon: Lucien is the only one of the three who regularly attends “the salon,” which situates him sociologically within the Parisian grand monde (upper bourgeoisie). By contrast, Emma Bovary is excluded from that world (and suffers for it), and Julien Sorel remains a provincial outsider.

Harder case

Triplet: A = Rosanette (Flaubert), B = Mme Arnoux (Flaubert), C = Mme Dambreuse (Flaubert)

Predicted outsider: A = 7, B = 5, C = 3

Table 5 shows that the harder triplet exhibits a fractured signal across classes, mirroring the interpretive subtlety of L’Éducation sentimentale. Dimension-wise outsiders alternate: some classes push A (Rosanette) out (cognition, sociological, existence state and body parts), others isolate B (Mme Arnoux) (affect emotion, communication, perception and physical traits) and others C (Mme Dambreuse) (action, motion and physical objects). This non-convergence is informative: Flaubert distributes Frédéric’s field of desire across three poles, sensuality (Rosanette), ideal (Mme Arnoux) and ambition/social power (Dambreuse), so that similarity depends on which facet is foregrounded. Our ontology’s facets capture these shifting alignments, producing a pattern of partial agreements rather than a single majority outsider.

Table 5. Cosine similarities over fine-grained ontological classes

Note: A: Rosanette from Gustave Flaubert $|$ B: Mme Arnoux from Gustave Flaubert $|$ C: Mme Dambreuse from Gustave Flaubert.

The dispersion of votes reflects genuinely competing interpretive frames in L’Éducation sentimentale: each woman anchors a distinct pole of Frédéric’s sentimental and social field, so the “outsider” depends on which facet of characterization the reader foregrounds. Expert feedback articulates three defensible choices for the outsider:

• • Rosanette is less refined, more eroticized and far more present in dialogue. In human relations/sociological, her lexicon skews to demi-monde ties (friend, lover and miss), vs. Mme Arnoux’s domestic roles and Mme Dambreuse’s status/kinship network. Weighting communication+affect/body cleanly isolates A.

• • On affect emotion/perception/traits, Mme Arnoux concentrates idealized feeling and inwardness (peaks on to love, fear, with evaluatives like good, heart and poor). Rosanette reads more performative/interactional; Mme Dambreuse cooler and milieu-coded. If one foregrounds affect/ideality, B emerges as the outsider.

• • Mme Dambreuse is more effaced, enters late and functions mainly as access to a milieu. Empirically, she is less mobile (0 occurrence on to go/to come/to do) and clusters on postural/possessional verbs and salon props, while the others show higher locomotion/transactions. Emphasizing action+motion+physical objects makes C the outsider.

Read this way, the dispersion of expert judgments is structured: distinct facet mixes, sociological+body (Rosanette), affect+ personality (Mme Arnoux) and action+motion+objects (Mme Dambreuse), each license a defensible outsider. This also explains our results: selective bundles beat using all classes (weak facets dilute signal), while targeted mixes approximate scholars’ multi-dimensional reasoning and argue for expert-aligned facet weighting.