---
license: cc-by-4.0
task_categories:
  - text-classification
  - feature-extraction
language:
  - en
tags:
  - computational-literary-studies
  - narrative-shape
  - semantic-novelty
  - time-series
  - sentence-transformers
  - digital-humanities
  - PG19
size_categories:
  - 10K<n<100K
pretty_name: PG19 Semantic Novelty
dataset_info:
  features:
    - name: gutenberg_id
      dtype: int64
    - name: title
      dtype: string
    - name: authors
      sequence: string
    - name: pub_year
      dtype: int64
    - name: subjects
      sequence: string
    - name: bookshelves
      sequence: string
    - name: download_count
      dtype: int64
    - name: primary_genre
      dtype: string
    - name: paragraph_count
      dtype: int64
    - name: mean_novelty
      dtype: float64
    - name: std_novelty
      dtype: float64
    - name: ti_ratio
      dtype: float64
    - name: trend_slope
      dtype: float64
    - name: mean_compression_progress
      dtype: float64
    - name: curve_type_3
      dtype: string
    - name: cluster_8
      dtype: int64
    - name: cluster_name
      dtype: string
    - name: speed
      dtype: float64
    - name: volume
      dtype: float64
    - name: circuitousness
      dtype: float64
    - name: reversal_count
      dtype: int64
    - name: sax_16_5
      dtype: string
    - name: novelty_curve
      sequence: float64
    - name: paa_16
      sequence: float64
  splits:
    - name: train
      num_examples: 28535
---

# PG19 Semantic Novelty Dataset

Paragraph-by-paragraph **semantic novelty curves** for **28,535 books** from the [PG19 corpus](https://huggingface.co/datasets/deepmind/pg19) (Project Gutenberg, pre-1920 English literature).

## What is Semantic Novelty?

For each paragraph in a book, we compute:

```
novelty(p) = 1 - cosine_similarity(embedding(p), running_centroid)
```

where `embedding()` uses **SBERT all-mpnet-base-v2** (768-dimensional) and `running_centroid` is the mean of all preceding paragraph embeddings. This measures how much *new information* each paragraph introduces relative to everything before it.

The resulting **novelty curve** captures the information-delivery shape of a narrative — whether a book front-loads its ideas (convergent/green), sustains steady novelty (parallel/blue), or builds to increasingly novel content (divergent/red).

## Key Features

| Feature | Description |
|---------|-------------|
| `novelty_curve` | Full paragraph-level novelty series (variable length, typically 100-5000 values) |
| `paa_16` | 16-segment Piecewise Aggregate Approximation (fixed-length summary) |
| `sax_16_5` | SAX string representation (16 chars, 5-letter alphabet) |
| `cluster_8` | 8-cluster Ward-linkage taxonomy (1-8) |
| `cluster_name` | Human-readable archetype label |
| `curve_type_3` | Legacy 3-type classification (green=convergent, blue=parallel, red=divergent) |
| `ti_ratio` | Tail/initial novelty ratio (>1 = divergent, <1 = convergent) |
| `speed`, `volume`, `circuitousness` | Toubia et al. (2021) narrative shape metrics |
| `mean_compression_progress` | Schmidhuber (2009) compression progress proxy |

## Dataset Statistics

- **Books**: 28,535
- **Publication years**: 1531\u20132014
- **Mean paragraphs per book**: 1128
- **Mean semantic novelty**: 0.4590
- **Books with novelty curves**: 28,535
- **Books with derived metrics**: 28,433

### 3-Type Distribution
| Type | Count | Description |
|------|-------|-------------|
| green (convergent) | 2,573 | Novelty decreases - ideas consolidate |
| blue (parallel) | 12,338 | Novelty stays steady |
| red (divergent) | 13,624 | Novelty increases - ideas expand |

### 8-Cluster Taxonomy
| Cluster | Count | % |
|---------|-------|---|
| Flat | 7442 | 26.1% |
| Late Plateau | 6540 | 22.9% |
| Early Plateau | 4486 | 15.7% |
| U-Shape | 2791 | 9.8% |
| Gradual Ascent | 2638 | 9.2% |
| Steep Ascent | 2637 | 9.2% |
| Steep Descent | 1677 | 5.9% |
| Gradual Descent | 222 | 0.8% |

## Usage

```python
from datasets import load_dataset
import numpy as np

ds = load_dataset("wfzimmerman/pg19-semantic-novelty", split="train")

# Get a book's novelty curve
book = ds[0]
print(f"{book['title']} by {book['authors']}")
print(f"Cluster: {book['cluster_name']} ({book['curve_type_3']})")
print(f"Paragraphs: {book['paragraph_count']}, Mean novelty: {book['mean_novelty']:.4f}")

# Plot a novelty curve
import matplotlib.pyplot as plt
curve = book["novelty_curve"]
plt.plot(curve)
plt.xlabel("Paragraph")
plt.ylabel("Semantic Novelty")
plt.title(book["title"])
plt.show()

# Filter by cluster
steep_descent = ds.filter(lambda x: x["cluster_name"] == "Steep Descent")
print(f"Steep Descent books: {len(steep_descent)}")

# Use PAA for fixed-length comparison
paa_matrix = np.array([x["paa_16"] for x in ds if x["paa_16"] is not None])
print(f"PAA matrix shape: {paa_matrix.shape}")  # (N, 16)
```

## Methodology

1. **Corpus**: PG19 (Rae et al., 2020) - 28,535 English-language books from Project Gutenberg published before 1920
2. **Embeddings**: Sentence-BERT `all-mpnet-base-v2` (Reimers & Gurevych, 2019), 768 dimensions
3. **Novelty computation**: `1 - cosine_similarity(paragraph_embedding, running_centroid)` where centroid accumulates all prior paragraphs
4. **Clustering**: Ward-linkage hierarchical clustering on 16-segment PAA vectors, k=8
5. **Toubia metrics**: Speed, volume, circuitousness following Toubia et al. (2021)
6. **SAX encoding**: Lin et al. (2003) Symbolic Aggregate Approximation, 16 segments, 5-letter alphabet

## Citation

If you use this dataset, please cite:

```bibtex
@dataset{zimmerman2026pg19novelty,
  title={PG19 Semantic Novelty: Paragraph-Level Information Curves for 28,000+ Books},
  author={Zimmerman, W. Frederick},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/wfzimmerman/pg19-semantic-novelty}
}
```

## Related Work

- Toubia, O., et al. (2021). How quantifying the shape of stories predicts their success. *PNAS*.
- Reagan, A. J., et al. (2016). The emotional arcs of stories. *EPJ Data Science*.
- Schmidhuber, J. (2009). Simple algorithmic theory of subjective beauty. *arXiv:0812.4360*.
- Reimers, N. & Gurevych, I. (2019). Sentence-BERT. *EMNLP*.
- Rae, J. W., et al. (2020). Compressive Transformers for Long-Range Sequence Modelling.

## License

CC-BY-4.0. The underlying texts are public domain (Project Gutenberg). The novelty analysis, clustering, and derived metrics are original contributions.