Source-linked AI summary
The Plot Thins: Uniformity and Linearity in Literary Summaries
Rebecca M. M. Hicke, Sil Hamilton, David Mimno, Ross Deans Kristensen-McLachlan
TL;DR
Literary summaries privilege plot, but little prior work has examined how they operationalize plot and privilege some story content over the rest. This paper aligns 150 Wikipedia novel summaries to source chapters and finds that summaries often break linearity and uniformity, with deviations reflecting emphasis and clarification choices.
Problem
Little prior work has examined how summaries operationalize plot and privilege some story content over the rest.
Method
The paper aligns sentences from 150 Wikipedia novel summaries with source chapters and analyzes the alignments using quantitative metrics and close readings.
Results
Summaries often break linearity and uniformity, while the alignment pipeline achieves ∼84% F1 on four human-aligned summaries and texts.
Takeaways & Limitations
Summary deviations from source order and coverage provide insight into how summaries emphasize significant events and clarify events and themes.
Takeaways & Limitations
The study is restricted to Wikipedia summaries, and mapping summary sentences to source texts remains challenging for human annotators and frontier LLMs.
Abstract
from arXiv · showhide
Works of literature are complicated; they balance plot, suspense, surprise, and artistic expression. Summaries of literature prioritize plot, and therefore may deviate from their sources. Using a combination of manual and LLM-based annotation, we construct a dataset mapping sentences from 150 novel summaries to their respective source chapters. We find the task unexpectedly difficult for both human and model annotators. Using the sentence-to-chapter mappings, we then measure summary linearity, the degree to which it maintains the source's order of events, and uniformity, the degree to which a summary spreads attention equally across a source. By examining when and how summaries break linearity and uniformity, we identify differences in how literary works and summaries express plot, particularly with regard to the clarity and prominence with which narrative details are described.
1 Introduction
This work studies how literary summaries compress plot by aligning 150 human-written summaries with source chapters and analyzing their departures from linearity and uniformity. Although frontier LLMs struggle with this reverse-summarization task, the authors develop a model-in-the-loop pipeline achieving ∼84% F1 on four human-aligned summaries and texts.
- Motivation: Literary summaries separate plot from artistic expression by compressing narratives to events deemed most impactful or necessary, making their omissions and modifications analytically informative.Examining these changes can characterize event significance while revealing how authors’ artistic expression shapes plot execution.
- Motivation: Summaries may deliberately violate linearity and uniformity because plot summaries need not recap every scene or moment, and lengthy middle sections may be condensed before climactic encounters.Such departures are permitted and sometimes encouraged by Wikipedia’s plot-summary guidelines.
- Contribution and findings: The study combines 150 human-written summaries with the literary works they describe, then uses quantitative metrics and close readings to characterize summary writing.Novels are split into chapters before summaries and chapters are aligned, as illustrated in Figure 1.
- Contribution and findings: ∼84% F1: A model-in-the-loop summary sentence-to-chapter mapping pipeline achieves this score on four human-aligned summaries and texts, despite frontier LLMs struggling with reverse summarization.The pipeline aligns summary sentences with chapters or other delineated narrative segments in the original works.
2 Related Work
Prior work distinguishes literary story events from their often non-chronological, ambiguous textual presentation, while computational summarization initially focused mainly on journalistic headlines and scientific abstracts. This paper addresses the limited study of how summaries operationalize plot and privilege some story content over other material.
- Stories, discourse, and plot: Literary texts may intentionally impede narrative comprehension through non-chronological retellings, unreliable narrators, and other stylistic devices, so textual order may not align cleanly with story order.This distinction motivates examining how summaries represent plot rather than assuming that literary texts present events clearly or chronologically.
- Stories, discourse, and plot: Little prior work has examined how summaries operationalize plot or why they privilege some story content over the rest, which this paper investigates.The paper frames this gap as a central motivation for its analysis of literary summaries.
- Summaries and computational narratology: Early computational summarization research primarily developed systems for generating journalistic headlines and scientific abstracts, using datasets commonly drawn from news sources such as CNN and NYT [21; 29; 10; 11; 24].Researchers viewed summarization as a potential route toward more general semantic parsing of natural language.
3 Data
The dataset pairs 150 public-domain novels with their Wikipedia plot summaries to study how summaries represent and reorganize narratives. The authors assemble and preprocess full texts and summaries, splitting novels into validated chapters and summaries into sentences for sentence-to-chapter mapping.
- Dataset construction: The dataset contains 150 public-domain novels and their Wikipedia plot summaries, chosen to study narrative representation rather than individual readers’ summarization.Wikipedia summaries were selected because they are typically co-written by multiple authors and are not necessarily biased toward one writer’s style.
- Dataset construction: The authors identified English-language novels through Project Gutenberg and matched them to English-language Wikipedia pages containing plot-oriented subheadings, yielding 80 initial candidate texts.They removed common single-word titles before searching for matching Wikipedia pages.
- Preprocessing: Novels were automatically chapterized and manually validated, while summaries were manually extracted, stripped of publication metadata, and split into sentences with spaCy.Punctuation was standardized and illustration markers removed so footnotes remained in the correct chapter files.
- Dataset characteristics: Novels range from 3 to 86 chapters and 9,780 to 488,101 tokens, averaging 32 chapters and 147,173 tokens per book.Summaries average 992 tokens and range from 243 to 3,534 tokens.
4 Generating Sentence-to-Chapter Mappings
The authors construct sentence-to-chapter mappings with an iteratively developed, three-prompt pipeline. On four annotated novels, the pipeline achieves 84.3% F1 overall, although accuracy varies dramatically by book.
- 4 Generating Sentence-to-Chapter Mappings: The mapping task represents bipartite edges between summary sentences and chapters describing their dramatized events, allowing one-to-many, many-to-one, or unmatched cases.Each sentence can map to any number of chapters, including zero, and vice versa.
- 4 Generating Sentence-to-Chapter Mappings: An evaluation set of four novels produced high expert agreement, with mean Cohen’s κ of 0.74; disagreements were resolved through discussion.The novels were Rookwood, Kazan, Kim, and Carmilla, paired with their Wikipedia summaries.
- 4 Generating Sentence-to-Chapter Mappings: Prompt development used iterative evaluation on the four-book dataset, with Claude Opus 4.8 grouping errors and proposing targeted edits.The resulting pipeline used preliminary chapter-wise mapping, correction of off-by-one errors, and a further prompt applied to selected matches.
- 4 Generating Sentence-to-Chapter Mappings: 84.3% F1: The complete mapping pipeline achieves this score across four evaluation novels, with performance varying dramatically by book.The pipeline ran Qwen 3.6 27B with temperature 0 and top_p 1.0, then mapped all 150 Wikipedia summaries.
5 Summary Linearity
Literary summaries are usually highly, though not perfectly, linear: 62.7% have l ≥0.9, while deviations reflect choices that can make narratives easier to parse. Linearity also increases moderately with summary length relative to book length.
- Method: Linearity measures Kendall’s Tau between chapter matches in summary order and their perfectly ordered version.This captures how much a summary’s event ordering differs from the source narrative.
- Results: 62.7% of summaries have l ≥0.9, but only 6.7% achieve perfect linearity (l = 1.0), with a median of 0.92.The distribution is left-tailed, with clear low-linearity outliers.
- Results: Out-of-order matches arise when summaries foreground later revelations, combine recurring events, separate concurrent plotlines, or reorder information for clarity.These deviations may therefore reflect deliberate summary-writing choices intended to make narratives easier to parse.
- Results: The least-linear outlier is The Expedition of Humphry Clinker by Tobias Smollett, with l ≈0.10, because broad plot statements match chapters throughout the narrative.Other outliers include L’Assommoir by Émile Zola (l ≈0.46) and The Marrow of Tradition by Charles W. Chesnutt (l ≈0.51), which intensify similar ordering choices.
- Results: Linearity correlates moderately positively with the log-ratio of summary length to book length (r ≈0.32, p < 10−3).The Ulysses summary is both unusually long and unusually linear (l ≈0.93), though editorial attention may contribute to longer summaries.
6 Summary Uniformity
Summary attention is generally nonuniform: most summaries skip chapters, while others dwell disproportionately on selected chapters or novel regions. The three metrics show that these patterns reflect both summary length and substantive choices about which events to foreground.
- Chapter coverage: Only 79% of novel chapters are matched on average, and just 12% of summaries cover every chapter, with coverage ranging from 0.31 to 1.0.Summary-to-book length correlates positively with chapter coverage (r ≈0.55, p < 10^-11), but explains only R2 ≈0.31 of its variation.
- Chapter coverage: Chapter coverage gaps can reflect streamlining or abstraction rather than irrelevance, as summaries sometimes introduce information at a broader level or condense narrative structure.The Murder of Roger Ackroyd maps to only 56% of chapters, while Flatland abstracts to the intentions of chapters; both illustrate alternative ways of presenting information.
- Unequal chapter focus: Among chapters included in a summary, sentence attention is usually uneven but moderate, with median Gc = 0.28 and every novel below 0.5.The most uniform cases have Gc < 0.1, while the least uniform summaries reach approximately 0.42–0.44 by concentrating sentences on subsets of chapters.
- Novel-position attention: Across novels, attention is slightly shifted toward second halves, while low or high µm values identify summaries that skip endings or beginnings; longer summaries also have higher Gc (r ≈0.42, p < 10^-6).µm ranges from 0.32 to 0.71 with median 0.52; some summaries omit final chapters, whereas others omit early chapters or describe later sections in greater detail.
7 Combining Linearity and Uniformity
The section combines linearity and uniformity into mean off-diagonal distance (µODD), measuring summaries against a perfectly linear and evenly distributed baseline. Human summaries generally have low µODD, but compression is associated with less linear and less even narrative delivery.
- Metric: µODD averages each sentence-to-chapter match’s distance from the perfectly linear, evenly distributed matching represented by the diagonal.The combined metric enables direct comparison between summaries while flattening some nuance captured by the separate metrics.
- Results: Human summaries have median µODD ≈ 0.11, spanning 0.02–0.35, with lower values resembling mostly linear summaries that distribute attention relatively evenly.Examples include Mike (µODD ≈ 0.02), The Cave Girl (µODD ≈ 0.03), and Sister Carrie: A Novel (µODD ≈ 0.03).
- Results: Summaries can combine nonlinearity and uneven attention: The Expedition of Humphry Clinker is extremely nonlinear, while Beauvallet and three others are predominantly linear but heavily non-uniform.The latter summaries are Beauvallet, The Triumph of the Scarlet Pimpernel, Captain Blood, and Main Street.
- Results: µODD is moderately negatively correlated with the log-ratio of summary-to-book length (r ≈ −0.36, p < 10−4), indicating that greater condensation changes narrative delivery toward less linearity and evenness.The finding suggests summaries with less space to describe a book are less linear and less evenly distributed.
8 Differences in Summary Sentences
This section evaluates whether summary sentences form comparable information units using sentence coverage and chapter-distribution metrics. Most summaries map nearly all sentences to chapters, while low-Gs summaries typically map each sentence to only 1–2 chapters and are often strongly linear.
- Metrics: The analysis uses ps, the proportion of summary sentences matched to at least one chapter, and Gs, the Gini coefficient of chapters matched to each sentence.These metrics assess sentence coverage and the distribution of source chapters across summary sentences.
- Sentence coverage: All sentences are matched in 52% of summaries, while at least 90% are matched in 91%; unmatched sentences are typically alignment errors.Examples include The Sound and the Fury by William Faulkner [ps ≈0.60], where stylistic complexity appears to confound the mapping pipeline, and several other summaries with ps ≈0.79–0.81.
- Chapter distribution: Gs ranges from 0.00 to 0.58 and has a lower median than Gc, 0.22 versus 0.28, despite being more widely distributed.The comparison is based on the distributions shown in Fig. 13b and Fig. 5a.
- Chapter distribution: Summaries with very low Gs, including Metamorphosis by Franz Kafka, Heart of Darkness by Joseph Conrad, and Ulysses by James Joyce, usually match each sentence with 1–2 chapters and are often strongly linear.Their reported Gs values are approximately 0.00, 0.00, and 0.01, respectively; Fig. 14 illustrates low-Gs sentence-to-chapter mappings.
9 Conclusion
The study treats literary summaries as interpretations and cultural artifacts, finding that their form reflects both stylistic choices and source structure. It also identifies challenging mapping work and several directions for expanding the dataset, authorship scope, and predictive analysis.
- Literary summaries vary in form and structure, shaped by summarization choices driven by stylistic desires and by the internal structure of the works they summarize.
- Summaries often depart from the expected order of events in their source works, breaking linearity.
- Next steps: Mapping summary sentences to source-text origins was surprisingly challenging for both human annotators and frontier LLMs, motivating larger datasets and improved mapping pipelines.
- Next steps: Future work could compare solo human-authored and AI-authored summaries and predict how summarizers respond to different source works.
A Novel Dataset · B Alignment Prompts
The paper assembles a literary-summary alignment dataset spanning novels from authors including Abbott, Alcott, Austen, Brontë, Dickens, and Stoker, then defines prompts for matching summary sentences to chapters. The prompts require decisions based only on chapter text, distinguishing dramatized or actively advancing events from recalled, foreshadowed, or merely mentioned material.
- A Novel Dataset: The dataset covers a broad range of novels, including Flatland: A Romance of Many Dimensions, Little Women, Emma, Jane Eyre, Wuthering Heights, and Dracula.The listed works span authors and publication periods from Abbott and Alcott through Stoker.
- B.1 Prompt 1: Prompt 1 instructs annotators to judge every summary sentence using only the supplied chapter text, without relying on prior knowledge of the novel, plot, or characters.The output must be a JSON mapping from every sentence id to YES or NO.
- B.1 Prompt 1: A sentence matches when the chapter dramatizes one of its events or first establishes or directly shows the described state, relationship, or characterization.The criterion is based on the underlying event or state rather than identical wording.
- B.1 Prompt 1: The prompts reject matches based only on recalled, summarized, discussed, foreshadowed, alluded-to, or merely recurring material, as well as setting or character presence alone.A recurring character counts only when the sentence’s own event occurs in the chapter.
- B.1 Prompt 1: Prompt 1 allows multiple chapter matches when a sentence bundles events occurring in different chapters or describes a continuing event actively unfolding across consecutive chapters.Already-matched sentences can be matched again when the same event continues or a different event from the sentence occurs in the current chapter.
- B.2 Prompt 2: Prompt 2 resolves tentative matches among adjacent chapters by selecting every chapter where the event actually happens or actively advances, including both chapters when the event spans their boundary.It excludes chapters that merely lead up to, recall, foreshadow, or follow the event.
- B.3 Prompt 3: Prompt 3 re-checks one match from the chapter text alone and requires the specific described occurrence, not a similar event involving different characters or a merely ongoing situation.Sentence order is explicitly non-determinative because flashbacks, recurring events, and later-fulfilled foreshadowing may produce out-of-order matches.
D Summary Texts
This section presents the Wikipedia summaries cited in the paper in order of appearance and marks unmatched summary sentences in red.
- D Summary Texts: The section compiles the Wikipedia summaries referenced in the paper.
- D Summary Texts: The summaries appear in the order in which they are referenced in the paper.
- D Summary Texts: Summary sentences unmatched to any source-text chapter after alignment are shown in red.
D.1 Ivanhoe: A Romance by Walter Scott
Ivanhoe follows the disinherited Wilfred through tournaments, captivity, rescue, and political upheaval after his return from the Crusades. The plot culminates in King Richard’s return, Rebecca’s witchcraft trial, and Cedric’s reconciliation with Ivanhoe and acceptance of his marriage to Rowena.
- D.1 Ivanhoe: A Romance by Walter Scott: Wilfred of Ivanhoe is disinherited for supporting Richard and loving Rowena, then secretly returns from the Crusades as the masked knight Desdichado.He wins the first tournament day, chooses Rowena as its queen, and is revealed after the second day’s melee.
- D.1 Ivanhoe: A Romance by Walter Scott: After Ivanhoe is wounded, Rebecca treats him, but he, Rebecca, and Isaac are captured with Cedric, Athelstane, and Rowena at Torquilstone.Gurth and Wamba escape and join Locksley and the Black Knight in planning the rescue.
- D.1 Ivanhoe: A Romance by Walter Scott: The besiegers storm Torquilstone after Wamba enables Cedric’s escape, while Ulrica sets the castle ablaze and Front-de-BÏuf dies.King Richard reveals himself as the Black Knight, releases de Bracy, and rescues Ivanhoe; Bois-Guilbert escapes with Rebecca.
D.2 Oil! by Upton Sinclair
The summary introduces Dad Ross and Bunny as they encounter an oil discovery amid family conflict in southern California, then follows their separation and Dad’s eventual death overseas.
- Dad Ross and Bunny drive through southern California to meet the Watkins family, whose dispute concerns how their oil properties should be managed and divided.
- While quail hunting on the Watkins ranch, Dad and Bunny discover oil, and Bunny urges Dad to intervene when the elder Watkins beats his daughter Ruth.
- Bunny later accompanies Dad to Vernon Roscoe’s mansion; Dad flees the country over the Teapot Dome scandal, marries Mrs. Olivier overseas, and dies of pneumonia.
- Before Dad leaves, Bunny proposes becoming independent, and Dad responds with confusion and hurt but remains supportive.
D.3 Carmilla by Joseph Sheridan Le Fanu
Laura’s childhood vision of Carmilla unfolds into a vampire attack as nearby girls die, her health declines, and Carmilla’s identity is linked to Mircalla Karnstein. Spielsdorf’s testimony and Baron Vordenburg’s investigation expose Carmilla’s history before Mircalla’s body is exhumed and destroyed.
- D.3 Carmilla by Joseph Sheridan Le Fanu: Carmilla’s resemblance to a 1698 portrait of Countess Mircalla Karnstein, together with her aversion to prayers and nocturnal behavior, marks her as supernatural.Laura also experiences nightmares of a cat-like beast attacking her and sees Carmilla standing bloodied at her bedside.
- D.3 Carmilla by Joseph Sheridan Le Fanu: Laura’s health deteriorates after nightmares of being bitten, prompting a doctor to find a blue spot and advise that she never be left unattended.Her father then travels with her toward the ruined village of Karnstein, leaving instructions for Carmilla to follow after she wakes.
- D.3 Carmilla by Joseph Sheridan Le Fanu: At Karnstein, Spielsdorf attacks Carmilla, who disarms him and disappears, after which he explains that her names are anagrams of Mircalla’s.The party then meets Baron Vordenburg, whose ancestor loved Mircalla and whose notes help locate her hidden tomb.
- D.3 Carmilla by Joseph Sheridan Le Fanu: Carmilla is revealed as Millarca and Mircalla, the vampire responsible for the deaths of Bertha and other young women.Spielsdorf recounts how Millarca visited Bertha, whose illness matched Laura’s, before a vampire attacked her and Bertha died.
- D.3 Carmilla by Joseph Sheridan Le Fanu: An imperial commission exhumes Mircalla’s blood-filled body, drives a stake through its heart, and decapitates it after the vampire’s tomb is found.The body appears to breathe and its heart to beat before the killing is completed.
D.4 L’Assommoir by Émile Zola … D.27 Ulysses by James Joyce
The supplied summaries span novels whose plots range from tragic social decline and dimensional allegory to forbidden love, individual rebellion, and settlement on the prairie. Across these works, characters confront abandonment, coercive societies, imprisonment, social shame, and isolation.
- D.4 L’Assommoir by Émile Zola: L’Assommoir follows Gervaise and Coupeau from marriage into debt, hunger, alcoholism, and death, ending with Gervaise’s corpse left unnoticed for two days.The summary begins with Lantier abandoning Gervaise and her sons, then traces her marriage to Coupeau and their eventual ruin.
- D.5 Flatland: A Romance of Many Dimensions by Edwin Abbott Abbott: In Flatland, the Square encounters Lineland and Spaceland, learns that dimensional understanding is constrained by perspective, and faces authorities who suppress such knowledge.The Sphere demonstrates the third dimension through analogies and physical acts, while Flatland’s leaders respond to witnesses and dissenters with imprisonment or massacre.
- D.6 Adam Bede by George Eliot: Adam Bede centers on Hetty’s pregnancy, abandonment of her infant, conviction for child murder, and Dinah’s compassionate support before execution.Hetty becomes pregnant after agreeing to marry Adam, searches unsuccessfully for Arthur, and is sentenced to hang after the infant dies of exposure.
- D.8 The Phantom of the Opera by Gaston Leroux: The Phantom of the Opera culminates in Erik’s coercion of Christine, his release of her and Raoul after receiving compassion, and his reported death before Christine and Raoul elope.Erik threatens the Opera House with explosives, but Christine’s sympathy transforms his behavior and enables the couple’s escape.
- D.9 O Pioneers! by Willa Cather: O Pioneers! contrasts Alexandra’s persistence with her brothers’ desire to leave, tracing the family’s investment in land, later prosperity, renewed conflict, and Alexandra’s isolation.After crop failures, Alexandra persuades her brothers to mortgage the farm and buy land; sixteen years later, prosperity is followed by departures and estrangement.
- D.10 The Triumph of the Scarlet Pimpernel by Baroness Orczy: The Triumph of the Scarlet Pimpernel portrays revolutionary danger, a failed denunciation of Robespierre, and the rescue of the de Servals before Theresia pursues further intrigue in England.The Scarlet Pimpernel intervenes while the mob turns against Moncrif and the de Servals; Theresia later arrives disguised and seeks to manipulate Sir Percy’s household.
- D.11 Beauvallet by Georgette Heyer; D.12 The Golovlyov Family by Mikhail Saltykov-Shchedrin; D.13 Anne of Green Gables by L. M. Montgomery: Beauvallet, The Golovlyov Family, and Anne of Green Gables introduce protagonists whose family, social, or romantic circumstances drive subsequent conflict and adaptation.Beauvallet assumes a dead man’s identity to reach Dominica; Arina’s family is weakened by irresponsibility and financial control; Anne remains at Green Gables after arriving through a misunderstanding.
D.28 Micromegas by Voltaire
Voltaire’s seven-chapter tale follows Micromegas and a Saturnian companion from cosmic travels to their encounter with tiny human philosophers. The narrative uses scale, measurement, and cross-species dialogue to relativize humanity while exposing both human intelligence and vanity.
- Scale and relativization: The travelers’ immense bodies and astronomical journeys repeatedly relativize Earth’s size, human stature, and the brevity of earthly life.The text compares planetary dimensions, lifespans, and human height with cosmic scales, including calculations about human conflict.
- Plot: Micromegas, banished after his scientific book is deemed heretical, travels from Sirius through Saturn and other planets with a Saturnian companion in pursuit of knowledge.Their journey includes visits to Saturn, Jupiter, Mars, and their satellites before reaching Earth.
- Encounter with humanity: After arriving on Earth, the aliens initially mistake the planet for lifeless, then discover a whale and a boat of philosophers through increasingly powerful magnification and improvised instruments.They use a diamond magnifying glass, a fingernail hearing tube, and toothpicks to communicate without deafening the humans.
- Human philosophy: The encounter reveals human intellectual capacity while also exposing human vanity and philosophical pretension, which the travelers mock.The Saturnian first concludes that the humans lack intelligence, but Micromegas reasons with him; the humans can measure their visitors and describe other scales of animal life.
E Per-Book Metric Values
This section presents metric values for *The Red Badge of Courage: An Episode of the American Civil War*.
- E Per-Book Metric Values: The section covers *The Red Badge of Courage: An Episode of the American Civil War*.