小説何以好讀？文學研究、編輯工藝與手稿數據

Why does a novel read well—and why do editors and readers notice that a chapter feels sluggish long before they can point to a grammatical error?

Over the past decade, computational literary studies and empirical editing research have examined thousands of novels to find out which measurable textual patterns correlate with literary craft, reading fluency, and commercial reader engagement. Two landmark studies stand out:

1. Jodie Archer & Matthew Jockers (The Bestseller Code, 2016) analyzed 20,000+ novels and found that strong reader engagement relies on symmetrical pacing beats, conversational immediacy (a higher share of active verbs, contractions, and short exchanges over static noun-plus-adjective stacks), and distinct character voices.
2. Ted Underwood (Distant Horizons, 2019) and Mark Algee-Hewitt et al. (Stanford Literary Lab) traced two centuries of fiction and showed a steady shift toward concrete scenic immediacy (Show vs. Tell: sensory perception and physical action in scene) over abstract summary.
3. Susan Greenberg (A Poetics of Editing, 2018) and Peter Ginna (What Editors Do, 2017) describe how developmental and line editors work in practice: they look for scene economy, voice contrast between dialogue and narration, and paragraph-local cohesion (how sentences hand over focus without monotonous sentence openings).

Most writing tools only inspect the surface: spelling, grammar rules, or single-formula readability scores like Flesch-Kincaid (which penalizes polysyllabic words and treats a choppy sequence of 11-word sentences identically to a varied, rhythmic paragraph).

With Lixity—an open-source (AGPL-3.0-or-later), local, stdlib-only manuscript analysis engine and reading workbench—we can measure these structural and stylistic signals across an entire book offline, without sending a word to the cloud and without turning prose into an opaque "quality score."

Let's look at what the data actually shows when we run Lixity on two public-domain classics bundled with the repository: Theodor Fontane's *Effi Briest* (German, 36 chapters, 95,449 words) and Jane Austen's *Pride and Prejudice* (English, 61 chapters, 117,916 words).

1. Whole-Novel Comparison: Effi Briest vs. Pride and Prejudice

Running lixity dialogue and lixity pacing on both novels produces the following whole-book baselines:

2. Finding 1: The -10 Word Voice Contrast Between Dialogue and Narration

Look at the Voice contrast (ASL Δ) row in the table above. Despite writing in different languages, decades apart, and in different narrative traditions:

• In Fontane's Effi Briest, sentences outside quotation marks average 21.1 words, while sentences inside quotation marks average 11.2 words—a voice contrast of −9.9 words per sentence (49.9% of all quoted spans are 1–12 words long).
• In Austen's Pride and Prejudice, sentences outside quotation marks average 23.4 words, while sentences inside quotation marks average 13.7 words—a voice contrast of −9.7 words per sentence (45.2% of all quoted spans are 1–12 words long).

Why this matters in manuscript editing

One of the most frequent developmental and line-editing observations in fiction drafts is that "the characters sound just like the narrator." When an author writes exposition inside quotation marks, dialogue_asl creeps up toward narration_asl, and the voice contrast (ASL Δ) approaches 0 or turns positive.

Watch what happens when we look at individual chapters in both novels where ASL Δ flips:

• Austen, Chapter 1 (ASL Δ = −3.5, 78.8% quoted speech, 55.6% short turns): The famous opening exchange between Mr. and Mrs. Bennet ("My dear Mr. Bennet, have you heard that Netherfield Park is let at last?"). Fast, clipped banter sets the comic pace immediately.
• Austen, Chapter 35 (ASL Δ = +0.8, 87.6% quoted text, 0.0% short turns, 5 long turns >= 80 words, dialogue_asl = 27.4): Why is the "dialogue" in Chapter 35 suddenly longer than the narration, with zero short turns? Because Chapter 35 is Mr. Darcy's letter to Elizabeth. It is enclosed in quotation marks, yet written in formal epistolary register (27.4 words per sentence).
• Austen, Chapter 49 (ASL Δ = −14.2, short_turn_pct = 69.5%, dialogue_asl = 9.2 vs. narration_asl = 23.4): When the Bennet family learns the terms of Lydia and Wickham's marriage settlement, nearly 70% of all spoken spans are 12 words or shorter, while the surrounding narrative sentences stay past 23 words.
• Fontane, Chapter 10 (ASL Δ = +0.4, dialogue_asl = 19.6 vs. narration_asl = 19.2): Effi explores her new house in Kessin while Innstetten explains the building's history and the upstairs curtains. Because Innstetten is lecturing rather than conversing, his spoken sentences match the length of the narrator's prose (+0.4 words)—a measurable signal of his pedagogical distance.

In Lixity (lixity dialogue in the CLI or the Quoted passages panel in the browser overview), you can inspect short_turn_pct, long_turns, dialogue_asl, narration_asl, and asl_contrast per chapter. If a tense confrontation scene shows short_turn_pct < 20% and ASL Δ ≈ 0, characters may be delivering essays to each other instead of reacting in scene.

3. Finding 2: Paragraph-Local Cohesion and Sentence-Start Monotony

Cognitive reading research (McNamara, Graesser, McCarthy & Cai, Automated Evaluation of Text and Discourse with Coh-Metrix, 2014) shows that readers process complex ideas smoothly when adjacent sentences share a lexical bridge—a repeated content noun or verb stem that hands focus from sentence $i$ to sentence $i+1$. Conversely, when consecutive sentences in a paragraph neither share content words nor vary their grammatical opening, the paragraph feels like a disconnected list (He walked to the window. He looked at the street. He turned back to the table.).

Lixity's Pacing & scenes report (lixity pacing) measures two complementary paragraph-local signals:

1. Adjacent sentence overlap (sentence_overlap_pct): The percentage of adjacent sentence pairs within the same paragraph that share at least one content word (excluding language-specific function words).
2. Repeated sentence-start runs (repeated_start_runs): Runs of >= 3 consecutive sentences inside the same paragraph that start with the exact same word token.

Look at how remarkably consistent the whole-book baseline is across both classics:

• Effi Briest: 23.7% of within-paragraph adjacent sentence pairs share a content word (1,105 of 4,661 pairs), with only 5 repeated sentence-start runs across 95,449 words.
• Pride and Prejudice: 23.6% of within-paragraph adjacent sentence pairs share a content word (869 of 3,689 pairs), with 21 repeated sentence-start runs across 117,916 words.

And when a chapter deviates sharply from that ~24% baseline, it points directly to a distinct rhetorical mode:

• In Effi Briest, Chapter 26, adjacent sentence overlap jumps to 46.6% (27 of 58 pairs) while narration_asl rises to 29.2 words. That is the pivotal chapter in which Innstetten discovers Major von Crampas's letters in Effi's sewing table and debates with Wüllersdorf whether honor demands a duel. The dense lexical chaining reflects Innstetten's obsessive, step-by-step legalistic reasoning.
• In Pride and Prejudice, Chapter 15, adjacent sentence overlap reaches 47.1% while narration_asl climbs to 36.6 words—the chapter introducing Mr. Collins's pompous, self-referential courtship calculations.

4. How to Use These Signals on Your Own Manuscript in Lixity

Lixity orders its Analysis overview along the sequence an author or editor uses during a revision pass:

1. Flagged passages & Work markers: Start with concrete paragraph anomalies (tense shifts, long sentence chains, high function-word density, spoken-delivery clustering) and your own open #TODO / #CHECK manuscript notes.
2. Pacing & scenes & Scenes & style registers: Check chapter length, sentence-length rhythm (staccato vs. long), chapter-ending hooks, adjacent sentence overlap (sentence_overlap_pct), and repeated sentence-start runs (repeated_start_runs). Assign intentional chapter registers (such as epistolary for a letter chapter like Darcy's in Chapter 35) so deliberate form shifts are documented rather than misread as errors.
3. Quoted passages & Sentence lengths: Compare short turns (1–12 words), long turns (>= 80 words), and voice contrast (ASL Δ = Dialogue ASL − Narration ASL) across chapters.
4. Signal groups, Characters & Motifs & echoes: Check perceptual/filter-word groups, character presence across the book, and nearby phrase echoes within a sliding word window.
5. Style & Chapter matrix: Review lexical diversity (MTLD, HD-D), corpus z*-scores, 3D chapter style space, and the full chapter-by-chapter comparison table.

Try it in the browser or via CLI

You can explore both full classic analyses directly in your browser—no installation required:

• Interactive Effi Briest analysis (de)
• Interactive Pride and Prejudice analysis (en)
• Full article with chapter tables (EN) · Deutsche Fassung (DE)

Or run the commands locally on your own Markdown, plain-text, DOCX, ODT, EPUB, or PDF draft:

pipx install lixity
lixity dialogue manuscript.md --language en
lixity pacing manuscript.md --language en
lixity dashboard manuscript.md

Every metric in Lixity is an invitation to read a specific passage with fresh eyes—never an automated grade.

Note: This article was drafted with AI assistance and verified against the Lixity software and documentation.