MSc Thesis — Larisa Ivanova · Saarland University / DFKI · 2026
Supervisor: Dr. rer. nat. Gerrit Großmann · Reviewers: Prof. Dr. Stefania Degaetano-Ortlieb, MSc. Sofía Aguilar
TL;DR — This dataset comprises 873 temporal signed character networks extracted from literary works, spanning 8 genres and 6 historical epochs. Each network tracks character relationships (positive / negative / neutral) at chapter-level resolution. Fictional social worlds obey real-world network laws (small-world structure, superlinear densification); structural balance is persistently high, genre-stratified, and temporally stable — with a non-monotonic U-shaped historical arc that peaks in Pre-Modern and 21st-Century fiction.
- What Are Temporal Signed Character Networks?
- Pipeline Overview
- Dataset
- Key Findings
- Case Studies: Per-Book Profiles
- Metrics and Parameters
- Repository Structure
- Getting Started
- Citation
- License
Literary narratives are bounded, intentionally constructed social systems. Characters form alliances, rivalries emerge and resolve, and relationships evolve across chapters in ways that stylize real-world social dynamics.
This project operationalizes these dynamics as temporal signed networks: one network per literary work, updated chapter by chapter.
- Nodes — named characters
- Edges — character relationships labeled
positive,negative, orneutral - Time — chapter index, normalized to τ ∈ (0, 1]
- Polarity convention — most-recent-wins: if two characters are allies in chapter 3 but enemies in chapter 7, the edge is coded as negative from chapter 7 onward
Relationships were extracted from publicly available chapter-level book summaries using Meta Llama-3.3-70B-Instruct via an automated LLM pipeline.
The pipeline proceeds in three stages:
Stage 1 — Data collection. Chapter-level narrative summaries and curated character lists are collected from online literary study guides. Each work is stored as two structured plain-text files: a summary file (chapter partitions delimited by === PARTITION N: [Title] ===) and a character file (MAJOR / MINOR sections with canonical names).
Stage 2 — LLM extraction (scripts/network_extraction.ipynb). Each chapter partition is processed by Meta Llama-3.3-70B-Instruct (via AcademicCloud and DeepInfra APIs), yielding signed edge tuples (A, B, σ, t) with σ ∈ {positive, negative, neutral}. The pipeline is crash-resilient (SQLite progress tracking), supports deterministic resume, and rotates across multiple API keys.
Stage 3 — Network construction and analysis (scripts/analysis/network_analysis.ipynb). The graph builder assembles cumulative signed graphs G₁, …, G_T under a most-recent-wins convention. Per-book 142-dimensional feature vectors are extracted. Analysis blocks cover: static network metrics (A), temporal dynamics (B), structural balance (C), temporal motifs (D), genre classification (E), epoch stratification (F), and per-book case studies (G).
Processing time: approximately 132 hours total on consumer hardware (Apple MacBook Pro, M-series chip). No GPU required.
→ Download Dataset v1 (networks + metadata, .zip)
Or use the files directly from the data/ folder in this repository.
109,855 rows × 6 columns. One row = one character-pair observation in one chapter.
| Column | Description |
|---|---|
Book |
"Author: Title" — unique book identifier |
Chapter |
Chapter number (1-indexed) |
Character A |
First character (canonical name from curated cast list) |
Character B |
Second character |
Relationship |
positive, negative, or neutral |
source |
Source corpus (sparknotes or litcharts) |
900 rows × 9 columns. One row per literary work.
| Column | Description |
|---|---|
Author |
Author name |
Book |
Title |
Year |
Year of first publication |
Epoch |
Historical epoch (6 categories) |
Genre_Primary |
Primary genre label (8 categories) |
Genre_Secondary |
Secondary genre label (where applicable) |
Form |
Novel, Play, or Other |
Remarks |
Curation notes (edge cases, non-fiction flags) |
source |
Source corpus |
| Statistic | Value |
|---|---|
| Works (after quality filtering) | 873 |
| Total character-relationship observations | 109,855 |
| Authors | 544 |
| Genres | 8 |
| Historical epochs | 6 |
| Median cast size | 16 characters |
| Median chapter count | 16 |
| Median edges per work | 29 |
Genre distribution:
| Genre | n | Dominant form |
|---|---|---|
| Literary / Realist Fiction | 286 | Novel |
| Coming-of-Age / Bildungsroman | 107 | Novel |
| Tragedy / Drama | 85 | Play (91.8%) |
| Fantasy / Adventure | 65 | Novel |
| Science Fiction / Dystopia | 52 | Novel |
| Comedy / Satire | 48 | Play (66.7%) |
| Mystery / Thriller | 47 | Novel |
| Romance / Domestic Fiction | 36 | Novel |
⚠️ Form–genre confound: Tragedy/Drama and Comedy/Satire are play-dominated; all other genres are predominantly novels. Plays have smaller, more densely interconnected casts. Cross-genre structural comparisons involving these two genres must be interpreted with caution. See the thesis (§VIII.3) for full discussion.
Of 696 works with sufficient density, 99.3% satisfy σ_sw > 1 (corpus median σ_sw = 2.83), confirming universal small-world structure. The corpus-wide median densification exponent is α = 1.48 (97.1% of books yield α > 1): each newly introduced character forms, on average, more than one relationship with the existing cast — mirroring laws documented in empirical social networks.
The strong-balance index B(τ) lies in [0.65, 0.84] across all genres and narrative positions, far above the B = 0.50 expectation for a randomly signed network. No genre shows a systematic monotone trend. Social-structural complexity is established early and maintained throughout.
Genre differences are statistically reliable (Kruskal–Wallis H = 38.8, p = 2.1 × 10⁻⁶, η² = 0.062) but partly attributable to the play/novel form distinction. The global median per-book argmin of B(τ) falls at τ* = 0.667, within the Freytag climax window [0.60, 0.75] (KS non-uniformity p < 10⁻⁶).
An exploratory U-shaped epoch effect (ΔB ≈ 0.10–0.12) reveals that Pre-Modern and 21st-Century works occupy the highest balance band, while 19th- and Early 20th-Century works (realism and modernism) occupy the lowest — a magnitude comparable to the genre effect.
Stable-alliance (+++) and stable-enmity (−−−) motifs together account for 76–87% of all length-3 polarity sequences. Comedy/Satire shows the highest reconciliation rate (−−+ = 0.047); Tragedy/Drama is the only genre where antagonism rivals alliance in base rate (−−− = 0.404, +++ = 0.389).
Random Forest macro-F1 = 0.139 on the full corpus (stratified baseline: 0.127; L1 Logistic Regression: 0.196). Entropy trajectory features dominate the top predictors; no B(τ) features appear in the top 10. Weak classification reflects genuine structural overlap among genres, not pipeline failure.
Benchmarked against majority-vote human annotations from Massey et al. (2015): macro-F1 = 0.58 on 966 matched character pairs across 91 works — within the inter-annotator agreement range (κ ≈ 0.65).
Three works were selected to illustrate how corpus-level findings manifest at the level of individual books — one per genre. Each profile shows the temporal evolution of the signed character network, the structural balance trajectory B(τ), and the full per-chapter relationship polarity heatmap. Detailed written analyses are in Appendix C of the thesis.
N = 15 characters · E = 32 edges · T = 20 chapters · σ_sw = 1.86
B(τ) stays at 1.0 from τ = 0.25 to τ = 0.75 — the court of Elsinore locks into a stable two-faction configuration. Five simultaneous sign flips in the final chapter collapse balance to B = 0.45, the most dramatic late-narrative structural breakdown in the corpus. Per-book argmin τ* = 1.0 — an outlier relative to the corpus median of 0.667, reflecting the Shakespearean convention that social dissolution concentrates in the denouement.
 |
 |
Temporal evolution of the signed character network across 20 partitions
N = 14 characters · E = 47 edges · T = 32 chapters · σ_sw = 1.56
The highest positive-edge fraction of the three case studies (f⁺ = 0.51). B(τ) oscillates in [0.50, 0.85] throughout — high and flat in aggregate, but driven by continuous dyadic reorganization at the pair level. The Elizabeth Bennet–Fitzwilliam Darcy edge cycles through four sign states (negative → neutral → negative → positive), the textbook reconciliation arc of the Romance genre. Per-book argmin τ* = 0.156, consistent with the Romance genre median of 0.654: peak tension arrives early and the narrative resolves it.
 |
 |
Temporal evolution of the signed character network across 32 partitions
N = 19 characters · E = 50 edges · T = 59 chapters · σ_sw = 1.90
The corpus's clearest U-shaped B(τ) arc. Balance begins at 0.833, drops to a minimum of 0.400 at τ* = 0.22 (Pip's period of maximum social dislocation), then recovers monotonically to 0.833 by τ = 1.0. The Estella–Pip edge records five sign changes across 59 chapters — the highest dyadic volatility of the three case studies. Degree entropy grows from H = 2.0 to 4.0, reflecting Pip's progressive social integration as he moves from the marshes to London.
 |
 |
Temporal evolution of the signed character network across 59 partitions
| Metric | Symbol | Description |
|---|---|---|
| Cast size | N | Number of distinct characters |
| Edge count | E | Number of unique character pairs |
| Density | δ = 2E / N(N−1) | Fraction of possible edges realized |
| Avg. clustering | C_avg | Mean local clustering coefficient |
| Avg. path length | L | Mean shortest path over the largest connected component |
| Small-world coeff. | σ_sw = (C/C_rand) / (L/L_rand) | σ_sw > 1 indicates small-world structure |
| Positive fraction | f⁺ | Fraction of edges labeled positive |
| Negative fraction | f⁻ | Fraction of edges labeled negative |
| Metric | Symbol | Description |
|---|---|---|
| Network density | D(τ) | δ evaluated at narrative position τ |
| Positive density | D⁺(τ) | n_pos(τ) / C(N,2) |
| Negative density | D⁻(τ) | n_neg(τ) / C(N,2) |
| Alliance ratio | R⁺(τ) = D⁺/(D⁺+D⁻) | Fraction of signed edges that are positive |
| Degree entropy | H(τ) | Shannon entropy of the degree distribution |
| Avg. clustering | C(τ) | Mean local clustering at snapshot τ |
| Densification exponent | α (from E ∝ N^α) | α > 1 = superlinear densification |
| Edge volatility | — | Fraction of edges that change polarity between consecutive chapters |
| Burstiness | κ = (σ_Δ − μ_Δ)/(σ_Δ + μ_Δ) | κ < 0 = more regular than Poisson; κ > 0 = bursty |
| Metric | Description |
|---|---|
| Balance index | B(τ) = fraction of closed signed triads that are balanced (+++ or +−−) |
| Weak balance | B_w(τ) = fraction of triads satisfying the weaker Harary (1953) criterion (+−− only) |
| Minimum balance position | τ* = argmin B(τ) — narrative position of peak social tension |
| End-minus-start balance | ΔB = B(τ=1) − B(τ=0) |
| Triad types | f_+++ , f_++− , f_+−− , f_−−− (fractions of the four signed triad types) |
Length-3 polarity sequences on recurring character pairs (neutral edges excluded):
| Motif | Label | Interpretation |
|---|---|---|
+++ |
Stable alliance | Relationship maintains positive polarity |
--- |
Stable enmity | Relationship maintains negative polarity |
+-- |
Alliance collapse | Friendship turns to lasting enmity |
--+ |
Reconciliation | Conflict resolves to positive |
++- |
Drift to conflict | Friendship gradually turns antagonistic |
-++ |
Drift to alliance | Conflict gradually resolves |
+-+ |
Betrayal–reconciliation | Temporary rupture of an alliance |
-+- |
Truce–relapse | Temporary resolution of a conflict |
| Parameter | Value | Description |
|---|---|---|
| Narrative time bins | 10 equispaced bins at τ ∈ {0.05, 0.15, …, 0.95} | Used for trajectory interpolation |
| Balance threshold | ≥ 5 signed triads per snapshot | Minimum for B(τ) to be computed |
| Small-world threshold | σ_sw > 1 | Condition for small-world classification |
| Motif length | 3 | Length of polarity sequences extracted |
| LLM temperature | 0.1 | Near-deterministic extraction |
| CV folds | 10-fold stratified | Genre classification cross-validation |
| RF trees | 500 | Random Forest ensemble size |
| Feature vector | 142 dimensions | 7 metrics × 19 features + 9 scalar features |
litnet-balance/
├── README.md # This file
├── LICENSE # MIT License (code)
├── LICENSE-DATA.md # CC BY 4.0 (dataset)
│
├── data/
│ ├── README.md # Data documentation
│ ├── combined_networks_all.csv # 109,855 character-relationship edges
│ └── combined_metadata_all.csv # 900-row per-work metadata
│
├── scripts/
│ ├── README.md # Script documentation and setup guide
│ ├── network_extraction.ipynb # LLM extraction pipeline (key rotation, crash resilience)
│ ├── requirements.txt # Python dependencies
│ └── analysis/
│ └── network_analysis.ipynb # Full analysis notebook (873 books)
│
├── figures/
│ ├── README.md
│ ├── pipeline_overview.png
│ ├── balance_trajectory_by_genre.png
│ ├── motif_analysis.png
│ ├── confusion_matrix.png
│ ├── feature_importance.png
│ ├── tsne_genre.png
│ ├── balance_by_epoch.png
│ ├── smallworld_sigma_distribution.png
│ └── case_studies/
│ ├── hamlet_profile.png
│ ├── hamlet_heatmap.png
│ ├── hamlet_network.gif
│ ├── pride_prejudice_profile.png
│ ├── pride_prejudice_heatmap.png
│ ├── pride_prejudice_network.gif
│ ├── great_expectations_profile.png
│ ├── great_expectations_heatmap.png
│ └── great_expectations_network.gif
│
└── thesis/
└── Ivanova_2026_Temporal_Dynamics_Social_Balance.pdf
Python 3.12. Install dependencies:
pip install -r scripts/requirements.txtOr individually:
pip install pandas numpy networkx scikit-learn matplotlib seaborn tqdm openaiimport pandas as pd
# Character-relationship edges
df = pd.read_csv("data/combined_networks_all.csv")
# Per-work metadata
meta = pd.read_csv("data/combined_metadata_all.csv")
# Merge for analysis
meta["BookKey"] = meta["Author"] + ": " + meta["Book"]
df_merged = df.merge(meta, left_on="Book", right_on="BookKey", how="left")
print(f"Corpus: {df['Book'].nunique()} works, {len(df):,} edges")Open scripts/analysis/network_analysis.ipynb in Jupyter. Run blocks in order:
Block 0 → 1 → A → B → C → D → E → F → H → G
Each block is self-contained with section headers and inline comments.
Before running, edit the two path variables at the top of Block 0 to point to your local copies of data/combined_networks_all.csv and data/combined_metadata_all.csv.
Open scripts/network_extraction.ipynb. Set API keys as environment variables before running — do not hardcode them:
export ACADEMICCLOUD_KEY_1="your_key_here"
export DEEPINFRA_KEY_1="your_key_here"See scripts/README.md for full setup instructions and supported endpoints.
If you use this dataset or code, please cite:
@mastersthesis{ivanova2026temporal,
author = {Larisa Ivanova},
title = {Temporal Dynamics of Social Balance in Fictional Character Networks},
school = {Saarland University},
year = {2026},
type = {MSc Thesis},
url = {https://github.com/L-Earthling/litnet-balance},
note = {Department of Language Science and Technology / DFKI}
}Code (scripts/): MIT License
Dataset (data/): Creative Commons Attribution 4.0 International (CC BY 4.0)
Thesis PDF (thesis/): © Larisa Ivanova, 2026. All rights reserved. Shared here for academic reference only.