Narrative Flattening in Computational Analysis

• Narrative flattening is the measurable compression of a story’s thematic, affective, and stylistic dimensions into a lower-dimensional representation.
• It employs methods such as dynamic character networks, causal DAGs, and role-sensitive embeddings to transform narratives into structured data for analysis.
• This approach balances analytical clarity with the loss of narrative nuance, influencing both computational narrative analysis and creative text generation.

Searching arXiv for recent and directly relevant papers on “narrative flattening” and adjacent narrative-representation work. Narrative flattening denotes the compression of a narrative trajectory into a lower-dimensional, lower-variance, or otherwise simplified representation. In the most explicit recent formulation, it is defined as “the measurable compression of this variation—regularized movement, muted affect, and reduced sensitivity to the source domain” in LLM fiction (Li et al., 27 May 2026). Across adjacent computational research, the same general operation appears in more constructive forms: narratives are converted into dynamic character networks, causal DAGs, actant-structured embeddings, or narrative-anchored geometric projections in order to make them analyzable, comparable, and visualizable (Min et al., 2016, Huang et al., 2 Mar 2026, Elfes, 2024, Keith-Norambuena et al., 29 Jun 2026). The concept therefore names both a methodological reduction of narrative complexity and, in generative modeling, a degradation in narrative variation.

1. Definition and semantic range

The most direct definition appears in the study of post-trained LLMs for fiction, where a story is treated as a trajectory through thematic, affective, and linguistic space, and narrative flattening is the measurable compression of that trajectory into more regular thematic motion, more neutral affect, and less stylistic spread across stories (Li et al., 27 May 2026). In that usage, flattening is a loss phenomenon: dynamic variation that characterizes human fiction is compressed by post-training.

Taken together, the broader literature suggests a second usage in which flattening is a deliberate representational strategy. Network analysis turns an unfolding text into a character graph plus textual attributes; graph annotation reduces narrative discourse to event nodes and causal edges; embedding-based narrative mapping projects high-dimensional semantic space into a navigable surface; and role-sensitive embeddings attempt to resist the flattening induced by ordinary semantic vectors (Min et al., 2016, Huang et al., 2 Mar 2026, Keith-Norambuena et al., 29 Jun 2026, Elfes, 2024). This suggests that “narrative flattening” is not a single method but a family of reductions that differ in what they preserve and what they discard.

A central distinction in this literature is between semantic content and narrative configuration. Standard embeddings and topic models largely preserve “what is being talked about,” but they can blur “who is doing what to whom,” role structure, narrative perspective, and causal depth. The result is a form of flattening in which different story logics occupy the same semantic region unless additional structure is imposed (Elfes, 2024).

2. Structural reductions: from text to networks and causal graphs

One influential structural approach models a narrative as a dynamically unfolding character network. In the Les Misérables case study, nodes are characters and edges are defined by co-appearance in the same chapter, yielding an undirected adjacency matrix

$$A_{\alpha\beta} = \begin{cases} 1,& \text{if characters } \alpha,\beta \text{ co-appear in at least one chapter}\ 0,& \text{otherwise.} \end{cases}$$

Narrative time is encoded by chapters, books, volumes, and data-driven Sequences, and cumulative snapshots $G(t)=(V(t),E(t))$ track how the network grows over the 365 chapters (Min et al., 2016).

This reduction produces a tractable representation of progression. The framework tracks cumulative node and edge growth through $n(t)=|V(t)|$ and $m(t)=|E(t)|$, along with node-level appearance counts and degrees. In the final pruned network there are 63 characters and 504 edges, with density approximately $25.8\%$, mean shortest path length $1.85$, diameter $4$, and clustering coefficient $0.77$ (Min et al., 2016). The resulting representation supports narratological interpretations: early spikes in character introduction correspond to exposition, flatter regions to digression or isolated character focus, and late bursts in edges among previously separate groups to climactic convergence.

The same paper extends flattening beyond pure co-occurrence by attaching sentiment and topic information to the network. Chapters receive a Sentiment Polarity Index from LIWC; average chapter sentiment is then propagated to characters and co-appearing pairs, and pairwise “cosentiment” is defined relative to the global average pair sentiment. Lexical content is compressed by Non-Negative Matrix Factorization of a TF–IDF matrix into 50 topics, yielding a character–topic association matrix $T=[t_{\alpha k}]$ (Min et al., 2016). What is preserved is the cast, interaction pattern, temporal order via chapter index and Sequences, and abstracted emotional and topical state; what is lost includes prose style, syntax, narrative voice, internal monologue, precise causality, and distinctions among interaction types.

A different structural reduction appears in graph annotation for economic news narratives, where narratives are represented as DAGs of event categories and causal links:

$$g=\{(e_i,r_j,e_k)\}, \qquad r_j \in \{\text{Increases},\text{Decreases}\}.$$

The annotation framework defines six representational levels: All Events, Adjacent Events, Relations, Full Story, Adjacent Story, and Extended Story. Moving from Full Story to Adjacent Story truncates causal depth; moving from graphs to set-based representations removes edges entirely (Huang et al., 2 Mar 2026).

This graph-theoretic flattening is explicitly tied to reliability under human label variation. The paper reports that lenient metrics overestimate reliability: for All Events, Krippendorff’s $G(t)=(V(t),E(t))$0 falls from $G(t)=(V(t),E(t))$1 under the lenient metric to $G(t)=(V(t),E(t))$2 under the strict metric; for Full Story it falls from $G(t)=(V(t),E(t))$3 to $G(t)=(V(t),E(t))$4 (Huang et al., 2 Mar 2026). By contrast, locally constrained representations are more invariant: Relations scores $G(t)=(V(t),E(t))$5, $G(t)=(V(t),E(t))$6, and $G(t)=(V(t),E(t))$7 under lenient, moderate, and strict metrics respectively, and the authors explicitly conclude that “Adjacent Story provides the most reliable and consistent graph representation” (Huang et al., 2 Mar 2026). In this setting, flattening is a way to reduce annotation variability by restricting the representational space to direct causes and local structure.

3. Embedding, projection, and the attempt to counter semantic flattening

A major criticism of standard NLP representations is that they flatten narrative into a semantic cloud. In news analysis, ordinary document embeddings can place articles close together because they share vocabulary even when they differ sharply in actantial configuration, such as whether a particular actor is Subject, Opponent, Helper, or Receiver (Elfes, 2024). To address this, a narrative-structured embedding framework uses Greimas’ six actants—Subject, Object, Sender, Receiver, Helper, Opponent—as explicit slots.

The pipeline extracts actants with Llama-3-8B-Instruct under deterministic decoding, embeds the selected actor phrase for each slot with E5-large into $G(t)=(V(t),E(t))$8, applies SVD separately to each actant position with $G(t)=(V(t),E(t))$9, and concatenates the six reduced vectors into a 204-dimensional narrative-structured embedding:

$n(t)=|V(t)|$0

UMAP and Ward-linkage agglomerative clustering on 5,342 Israel–Palestine articles then yield 18 clusters interpreted as cultural narratives (Elfes, 2024). The key point is structural sensitivity: swapping the same actor between Subject and Opponent changes different coordinate blocks, so similarity becomes role-sensitive rather than merely topical.

Another embedding-based flattening strategy appears in Information Terra, which projects a corpus of normalized document embeddings from the unit hypersphere onto an Earth-like globe whose poles are user-chosen endpoint documents. If $n(t)=|V(t)|$1 and $n(t)=|V(t)|$2 are source and target embeddings and $n(t)=|V(t)|$3, then narrative progress is the clipped parameter of the closest point on the SLERP geodesic:

$n(t)=|V(t)|$4

with latitude defined as

$n(t)=|V(t)|$5

Longitude is computed from the residual of $n(t)=|V(t)|$6 orthogonal to the SRC–TGT plane, reduced by PCA and mapped by $n(t)=|V(t)|$7 into a circular deviation coordinate (Keith-Norambuena et al., 29 Jun 2026).

Here flattening is explicit and constructive: latitude encodes narrative progress between endpoints, longitude encodes thematic deviation from the geodesic, and a geodesic-monotone Maximum Capacity Path extracts a northward-moving storyline. On a 540-article Cuban Protests corpus, KDE thresholding on the sphere yields 27 landmasses—8 continents and 19 islands, plus 20 archipelagos and 43 singleton islets—and the monotonicity constraint eliminates backtracking in 814 out of 1000 random endpoint pairs with mean relative coherence loss of approximately $n(t)=|V(t)|$8, a 95th percentile of approximately $n(t)=|V(t)|$9, and only 13 out of 1000 cases exceeding $m(t)=|E(t)|$0 (Keith-Norambuena et al., 29 Jun 2026). The method deliberately compresses high-dimensional story space into a navigable 2D manifold whose axes are semantically anchored by user-chosen endpoints.

4. Narrative flattening in post-trained language-model fiction

The explicit coinage of the term in generative modeling arises from a matched story-continuation study comparing four OLMo 32B checkpoints—Base, SFT, DPO, and RLVR—against matched human continuations from three domains: The New Yorker, TMAS, and StoryStar (Li et al., 27 May 2026). Because these checkpoints share architecture, scale, tokenizer, and pretraining corpus, the design isolates the post-training effect.

The study measures three sentence-level dimensions. Thematic motion is defined from normalized sentence embeddings, with jump sizes

$m(t)=|E(t)|$1

and variability summarized by

$m(t)=|E(t)|$2

Affective prevalence is computed by mapping top-1 GoEmotions labels into four families and defining

$m(t)=|E(t)|$3

with affective charge

$m(t)=|E(t)|$4

Stylistic diversity is assessed with StyleDistance embeddings via MMD$m(t)=|E(t)|$5 between human and model distributions and by across-story variance

$m(t)=|E(t)|$6

Across all three facets, post-training compresses variation. For The New Yorker continuations, human thematic CV is approximately $m(t)=|E(t)|$7, compared with $m(t)=|E(t)|$8 for Base, $m(t)=|E(t)|$9 for SFT, and $25.8\%$0 for DPO and RLVR, corresponding to a reported loss versus human of $25.8\%$1 at DPO/RLVR (Li et al., 27 May 2026). The human 5–95% range of thematic CV is $25.8\%$2, while RLVR compresses it to $25.8\%$3. In mixed-effects modeling, RLVR versus Human reduces thematic CV with $25.8\%$4, 95% CI $25.8\%$5, $25.8\%$6.

Affective dynamics shift from high-intensity states toward neutrality. In The New Yorker continuations, human conflict prevalence is approximately $25.8\%$7, surprise–curiosity approximately $25.8\%$8, and neutral approximately $25.8\%$9. The Base model is over-marked, with conflict approximately $1.85$0, surprise–curiosity approximately $1.85$1, and neutral approximately $1.85$2. By RLVR, conflict falls to approximately $1.85$3, surprise–curiosity to approximately $1.85$4, and neutral rises to approximately $1.85$5 (Li et al., 27 May 2026). The paper therefore characterizes post-training not as a simple move toward human affect but as an overshoot into muted, neutral territory. This is reinforced by mixed-effects estimates: for The New Yorker, RLVR-Human conflict prevalence is $1.85$6 and RLVR-Human neutral prevalence is $1.85$7, both with $1.85$8.

Stylistic variation collapses into a narrow attractor. Relative to The New Yorker human distribution, Base has MMD$1.85$9 and very high cross-story variance, with Var/human $4$0; SFT moves further from human in mean style with MMD$4$1 while dropping Var/human to $4$2; DPO and RLVR push MMD$4$3 to approximately $4$4 and $4$5 respectively and reduce Var/human to approximately $4$6 and $4$7 (Li et al., 27 May 2026). The interpretation offered is that Base exhibits stylistic sprawl, whereas post-training collapses that sprawl onto a narrow but offset stylistic regime.

The flattening is directionally stable across domains, but gap size depends on the human baseline. Cross-domain range in thematic CV shrinks from $4$8 in human continuations to $4$9 under RLVR, a reduction of approximately $0.77$0; cross-domain range in affective charge shrinks from 17.8 percentage points to 3.3 points, a reduction of approximately $0.77$1; and the largest human cross-domain stylistic MMD$0.77$2 of $0.77$3 falls to approximately $0.77$4 under RLVR (Li et al., 27 May 2026). Professional literary fiction is compressed most because its baseline distribution lies farthest from the post-trained model’s “default rhythm.”

5. Preservation, loss, and controversy

The major epistemic trade-off in narrative flattening is between analytical clarity and narrative richness. Character-network models preserve who appears, with whom, and when, plus abstracted sentiment and topics, but they lose prose style, metaphor, syntax, internal monologue, precise causality, focalization, and distinctions among interaction types (Min et al., 2016). Causal-DAG annotation preserves event identity and edge direction at selected levels of abstraction, but it removes tone, framing, and discourse-level organization, and stricter comparison reveals substantial human label variation in rich graphs (Huang et al., 2 Mar 2026).

Role-sensitive embeddings counter one kind of flattening—the collapse of role structure into generic semantic similarity—but still reduce each article to a single static actantial configuration, usually keeping only the first actor per actant. This leaves temporal evolution, multi-threaded narratives, predicate semantics, and intra-textual heterogeneity underrepresented (Elfes, 2024). Information Terra makes the flattening itself explicit: it compresses all off-axis thematic directions into a single circular longitude coordinate and depends entirely on the analyst’s chosen source and target documents. Poor endpoint choices can therefore induce misleading or forced narratives, and residual PCA may under-represent rare but important thematic deviations (Keith-Norambuena et al., 29 Jun 2026).

In annotation research, a common misconception is that higher agreement necessarily reflects better narrative representation. The factorial analysis of inflation narratives shows that lenient overlap-based metrics can mask disagreement: any non-empty overlap yields distance $0.77$5, which can make divergent annotations appear fully aligned. The authors therefore recommend reporting multiple $0.77$6 scores under different metrics and treating high lenient agreement cautiously, especially when representations are already flattened (Huang et al., 2 Mar 2026).

In fiction generation, a corresponding misconception is that flattening is simply professionalization or editing. The post-training study explicitly rejects that reading. The endpoint regime does not converge on human literary fiction; rather, it overshoots toward muted affect and compressed stylistic variation, and the domain most strongly affected is The New Yorker, the corpus with the richest baseline literary modulation (Li et al., 27 May 2026). A plausible implication is that alignment objectives optimized for assistant behavior can be well matched to consistency and safety while being poorly matched to literary variation.

6. Applications and future directions

Deliberate flattening remains useful because it converts narrative into structured data. Dynamic character networks support comparative narrative analysis, recommendation by character-topology or sentiment dynamics, literary studies of centrality and thematic focus, social interaction analysis, and tools for automatic narrative generation or draft diagnostics (Min et al., 2016). Narrative-anchored spherical projections support readable storyline extraction, globe-based exploration, antipodal reading of opposed themes, and semantic terrain maps with continents and islands (Keith-Norambuena et al., 29 Jun 2026). Role-sensitive embeddings support the study of competing media narratives across outlets by distinguishing articles that share topics but assign different actantial roles (Elfes, 2024).

The next research problem is not whether to flatten, but how much, along which dimensions, and for what downstream purpose. In annotation, the recommended direction is principled flattening: use QCA-style iterative refinement to reduce annotation error, then choose the least reductive representation that still yields acceptable invariance, possibly with semantically aware distance metrics and more diverse annotator pools (Huang et al., 2 Mar 2026). In narrative representation, future work includes richer dynamic modeling, multiple actants per role, multiple narrative threads per document, dialog and finer interaction extraction, and explicit integration of time into geometric narrative maps (Min et al., 2016, Elfes, 2024, Keith-Norambuena et al., 29 Jun 2026).

For creative LLMs, the problem is inverted: the goal is to prevent unwanted flattening. Suggested directions include domain-conditional reward models, distributional matching objectives that preserve cross-story variance, preference data sampled from the literary tail, and the use of thematic CV, affective charge, and stylistic variance as training-time diagnostics (Li et al., 27 May 2026). More broadly, the literature suggests that narrative flattening should be treated neither as an intrinsic defect nor as a neutral preprocessing step. It is a controlled compression whose value depends on whether the task is explanation, annotation, visualization, or creative generation—and on which dimensions of narrative one can afford to lose.