Semantic Novelty at Scale: Narrative Shape Taxonomy and Readership Prediction in 28,606 Books

W. Frederick Zimmerman

cs.CLFeb 24, 2026

Semantic Novelty at Scale: Narrative Shape Taxonomy and Readership Prediction in 28,606 Books

Authors: W. Frederick Zimmerman

Organizations: Nimble Books LLC

Abstract

I introduce semantic novelty--cosine distance between each paragraph's sentence embedding and the running centroid of all preceding paragraphs--as an information-theoretic measure of narrative structure at corpus scale. Applying it to 28,606 books in PG19 (pre-1920 English literature), I compute paragraph-level novelty curves using 768-dimensional SBERT embeddings, then reduce each to a 16-segment Piecewise Aggregate Approximation (PAA). Ward-linkage clustering on PAA vectors reveals eight canonical narrative shape archetypes, from Steep Descent (rapid convergence) to Steep Ascent (escalating unpredictability). Volume--variance of the novelty trajectory--is the strongest length-independent predictor of readership (partial rho = 0.32), followed by speed (rho = 0.19) and Terminal/Initial ratio (rho = 0.19). Circuitousness shows strong raw correlation (rho = 0.41) but is 93 percent correlated with length; after control, partial rho drops to 0.11--demonstrating that naive correlations in corpus studies can be dominated by length confounds. Genre strongly constrains narrative shape (chi squared = 2121.6, p < 10 to the power negative 242), with fiction maintaining plateau profiles while nonfiction front-loads information. Historical analysis shows books became progressively more predictable between 1840 and 1910 (T/I ratio trend r = negative 0.74, p = 0.037). SAX analysis reveals 85 percent signature uniqueness, suggesting each book traces a nearly unique path through semantic space. These findings demonstrate that information-density dynamics, distinct from sentiment or topic, constitute a fundamental dimension of narrative structure with measurable consequences for reader engagement. Dataset: https://huggingface.co/datasets/wfzimmerman/pg19-semantic-novelty

Figures & tables

Figure 1: Eight narrative shape archetypes discovered by Ward-linkage hierarchical clustering on 16-segment PAA vectors from 28,606 books. Each panel shows the centroid curve (bold) with the 25th–75th percentile envelope (shaded). These eight shapes—from Steep Descent to Steep Ascent—constitute the core taxonomy of this paper.

Figure 2: The three canonical novelty trajectory types: convergent (green, 22.5%), plateau (blue, 59.0%), and divergent (red, 18.6%). These legacy categories are refined into eight archetypes by Ward-linkage clustering (Figure 1 ).

Figure 3: Distribution of

\log_{10}

(downloads) across the eight narrative clusters. Download distributions vary systematically by cluster, reflecting the interplay between narrative shape, book length, and readership patterns explored in Section 4 .

Genre	$N$	T/I	Circ	Med DL
Fiction	4,974	1.022	264.9	315
Children’s	3,470	1.072	189.4	291
History	3,242	1.088	210.6	424
Social Science	2,042	1.059	199.2	353
Travel/Geogr.	1,598	1.118	157.7	474
Poetry	1,449	1.031	69.4	319

Table 1: Semantic novelty metrics by genre. T/I = Terminal/Initial ratio; Circ = median circuitousness; DL = median downloads.

Figure 4: Genre composition within each of the eight narrative clusters. Fiction dominates plateau clusters (Flat, Late Plateau); History and Other dominate descent and ascent clusters. Genre and cluster are strongly associated (

\chi^{2}=543.2

p=2.41\times 10^{-74}

Metric	$\rho_{\text{raw}}$	$\rho_{\text{partial}}$	$\rho_{\text{length}}$
Circuitousness †	0.406	0.113	0.928
Compr. progress †	0.388	0.040	0.944
Reversal count †	0.381	$-0.180$	0.998
Volume	0.049	0.317	$-0.495$
Std novelty	0.046	0.316	$-0.500$
T/I ratio	0.207	0.191	0.088

Table 2: Spearman correlations with

\log_{10}

(downloads): raw vs. length-controlled,

n=28{,}433

. Metrics marked

\dagger

are strongly correlated with book length (

|\rho_{\text{length}}|>0.90

Quintile	Med Circ	$N$	Med DL	%Red
Q1 (straight)	37.0	5,687	267	44.8%
Q2	97.4	5,686	329	52.1%
Q3	172.1	5,687	368	52.7%
Q4	275.4	5,686	402	47.6%
Q5 (winding)	506.9	5,687	539	42.4%

Table 3: Downloads by circuitousness quintile. The gradient parallels book length; see Section 4.3 for length-controlled analysis.

Figure 5: Circuitousness vs.

\log_{10}

(downloads), colored by curve type. The raw Spearman

\rho=0.41

is substantially confounded by book length (

\rho_{\text{circ,length}}=0.93

); after length control, the partial

\rho=0.11

. See Table 2 for length-controlled rankings.

Figure 6: SAX symbol frequency by position. The opening and closing positions show the greatest variance; middle positions converge toward $c$ (mid-level novelty). The rightward shift from $b$ / $c$ dominance to $c$ / $d$ dominance reflects the overall tendency toward slightly increasing novelty.

Decade	$N$	T/I	%Green	%Blue	%Red
1840s	29	1.078	24.1	24.1	51.7
1850s	50	1.113	12.0	24.0	64.0
1860s	57	1.101	12.3	29.8	57.9
1870s	49	1.111	8.2	46.9	44.9
1880s	70	1.087	14.3	34.3	51.4
1890s	157	1.042	11.5	49.0	39.5

Table 4: Narrative shape metrics by publication decade, 1840–1910.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Cluster	$s_{1}$	$s_{2}$	$s_{3}$	$s_{4}$	$s_{5}$	$s_{6}$	$s_{7}$	$s_{8}$	$s_{9}$	$s_{10}$	$s_{11}$	$s_{12}$	$s_{13}$	$s_{14}$	$s_{15}$	$s_{16}$
Steep Desc.	.55	.69	.50	.26	.11	$-.01$	$-.12$	$-.16$	$-.23$	$-.24$	$-.27$	$-.27$	$-.27$	$-.26$	$-.23$	$-.02$
Grad. Desc.	$-2.25$	1.36	.78	.25	.21	.04	$-.02$	$-.13$	$-.16$	$-.17$	$-.18$	$-.19$	$-.12$	$-.17$	$-.16$	.14
Early Plat.	.55	.09	$-.06$	$-.10$	$-.11$	$-.11$	$-.10$	$-.10$	$-.08$	$-.04$	.01	$-.01$	.00	.02	$-.03$	.09
Late Plat.	$-.16$	$-.28$	$-.24$	$-.18$	$-.10$	$-.03$	.03	.09	.12	.11	.09	.05	.07	.08	.10	.24
U-Shape	.32	.10	$-.04$	$-.09$	$-.13$	$-.16$	$-.15$	$-.16$	$-.16$	$-.17$	$-.16$	$-.18$	$-.16$	$-.14$	.06	1.21
Flat	$-.00$	.02	.08	.10	.08	.06	.04	.02	$-.01$	$-.03$	$-.05$	$-.05$	$-.06$	$-.09$	$-.09$	$-.03$

Appendix

Table 5: Ward-linkage cluster centroids (16-segment PAA, z-scored). Each row represents the centroid of one cluster.

$k$	WCSS	Silhouette	Marginal %
2	112,325	0.091	—
3	105,243	0.059	6.3%
4	101,064	0.058	4.0%
5	97,663	0.029	3.4%
6	95,643	0.029	2.1%
7	93,789	0.022	1.9%

Appendix

Table 6: Cluster selection diagnostics (

n=8{,}000

subsample).

Computational Literary Studies

Computational Literary Studies uses data analysis and language technologies to examine literary style, themes, narratives, influence, and change.

See all 49 papers

Semantic Novelty Trajectories in 80,000 Books: A Cross-Corpus Embedding Analysis
Mar 2, 2026Fred ZimmermanComputational Literary Studies
Nimble Books LLC, Ann Arbor, MI
Narrative Fingerprints: Multi-Scale Author Identification via Novelty Curve Dynamics
Apr 1, 2026Fred Zimmerman, Hilmar AIAuthorship Attribution Stylometry
Nimble Books LLC / Big Five Killer. · Contributing Editor, RKHS Multiverses Research.
Novels generated by language models show compressed formal variation
Aug 12, 2026Mehdy Sedaghat Payam, Justin QuinnLLM Evaluation Long-Form Document Generation
University of Maryland · University of West Bohemia · Charles University

Mar 2, 2026J/K move · Enter open · S save

Semantic Novelty Trajectories in 80,000 Books: A Cross-Corpus Embedding Analysis

Fred Zimmerman

Nimble Books LLC, Ann Arbor, MI

Computational Literary Studies

Open paper AskPDF arXiv

I apply Schmidhuber's compression progress theory of interestingness at corpus scale, analyzing semantic novelty trajectories in more than 80,000 books spanning two centuries of English-language publishing. Using sentence-transformer paragraph embeddings and a running-centroid novelty measure, I compare 28,730 pre-1920 Project Gutenberg books (PG19) against 52,796 modern English books (Books3, approximately 1990-2010). The principal findings are fourfold. First, mean paragraph-level novelty is roughly 10% higher in modern books (0.503 vs. 0.459). Second, trajectory circuitousness -- the ratio of cumulative path length to net displacement in embedding space -- nearly doubles in the modern corpus (+67%). Third, convergent narrative curves, in which novelty declines toward a settled semantic register, are 2.3x more common in pre-1920 literature. Fourth, novelty is orthogonal to reader quality ratings (r = -0.002), suggesting that interestingness in Schmidhuber's sense is structurally independent of perceived literary merit. Clustering paragraph-level trajectories via PAA-16 representations reveals eight distinct narrative-shape archetypes whose distribution shifts substantially between eras. All analysis code and an interactive exploration toolkit are publicly available at https://bigfivekiller.online/novelty_hub.

Abstract

Computational Literary Studies

Explore similar work

Semantic Novelty Trajectories in 80,000 Books: A Cross-Corpus Embedding Analysis

Narrative Fingerprints: Multi-Scale Author Identification via Novelty Curve Dynamics

Novels generated by language models show compressed formal variation