---
task_categories:
- summarization
- text-generation
language:
- en
pretty_name: BookSum Summarization Dataset Clean
size_categories:
- 1K<n<10K
configs:
- config_name: books
  data_files: 
  - split: train
    path: "books/train.jsonl"
  - split: test
    path: "books/test.jsonl"
  - split: validation
    path: "books/val.jsonl"
- config_name: chapters
  data_files: 
  - split: train
    path: "chapters/train.jsonl"
  - split: test
    path: "chapters/test.jsonl"
  - split: validation
    path: "chapters/val.jsonl"
---

# Table of Contents

1. [Description](#description)
2. [Usage](#usage)
3. [Distribution](#distribution)
   - [Chapters Dataset](#chapters-dataset)
   - [Books Dataset](#books-dataset)
4. [Structure](#structure)
5. [Results and Comparison with kmfoda/booksum](#results-and-comparison-with-kmfodabooksum)

# Description:
This repository contains the Booksum dataset introduced in the paper [BookSum: A Collection of Datasets for Long-form Narrative Summarization
](https://arxiv.org/abs/2105.08209).

This dataset includes both book and chapter summaries from the BookSum dataset (unlike the kmfoda/booksum one which only contains the chapter dataset). Some mismatched summaries have been corrected. Uneccessary columns has been discarded. Contains minimal text-to-summary rows. As there are multiple summaries for a given text, each row contains an array of summaries.

# Usage
Note: Make sure you have [>2.14.0 version of "datasets" library](https://github.com/huggingface/datasets/releases/tag/2.14.0) installed to load the dataset successfully.
```
from datasets import load_dataset

book_data = load_dataset("ubaada/booksum-complete-cleaned", "books")
chapter_data = load_dataset("ubaada/booksum-complete-cleaned", "chapters")


# Print the 1st book
print(book_data["train"][0]['text'])

# Print the summary of the 1st book 
print(book_data["train"][0]['summary'][0]['text'])
```

# Distribution

<div style="display: inline-block; vertical-align: top; width: 45%;">
  
## Chapters Dataset

| Split   | Total Sum. | Missing Sum. | Successfully Processed | Chapters |
|---------|------------|--------------|------------------------|------|
| Train   | 9712       | 178          | 9534 (98.17%)          | 5653 |
| Test    | 1432       | 0            | 1432 (100.0%)          | 950  |
| Val     | 1485       | 0            | 1485 (100.0%)          | 854  |

</div>

<div style="display: inline-block; vertical-align: top; width: 45%; margin-left: 5%;">

## Books Dataset

| Split   | Total Sum. | Missing Sum. | Successfully Processed | Books |
|---------|------------|--------------|------------------------|------|
| Train   | 314        | 0            | 314 (100.0%)           | 151  |
| Test    | 46         | 0            | 46 (100.0%)            | 17   |
| Val     | 45         | 0            | 45 (100.0%)            | 19   |

</div>


# Structure:
```
Chapters Dataset
  0 - bid (book id) 
  1 - book_title
  2 - chapter_id
  3 - text (raw chapter text)
  4 - summary (list of summaries from different sources)
      - {source, text (summary), analysis}
      ...
  5 - is_aggregate (bool) (if true, then the text contains more than one chapter)

Books Dataset:
  0 - bid (book id)
  1 - title
  2 - text (raw text)
  4 - summary (list of summaries from different sources)
      - {source, text (summary), analysis}
      ...
```

# Reults and Comparison with kmfoda/booksum


Tested on the 'test' split of chapter sub-dataset. There are slight improvement on R1/R2 scores compared to another BookSum repo likely due to the work done on cleaning the misalignments in the alignment file. In the plot for this dataset, first summary \[0\] is chosen for each chapter. If best reference summary is chosen from the list for each chapter, theere are further improvements but are not shown here for fairness.
![image/png](https://cdn-uploads.huggingface.co/production/uploads/62a7d1e152aa8695f9209345/lUNes4SFXVMdtebGMEJK0.png)