File size: 2,859 Bytes
2b14e78 5c25213 2b14e78 ffbbe24 2b14e78 ffbbe24 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 |
---
dataset_info:
features:
- name: src
sequence: string
- name: tgt
sequence: string
- name: labels
sequence: int64
splits:
- name: test
num_bytes: 53831114
num_examples: 11490
- name: train
num_bytes: 1376640992
num_examples: 287113
- name: validation
num_bytes: 62200550
num_examples: 13368
download_size: 857262516
dataset_size: 1492672656
license: mit
task_categories:
- summarization
language:
- en
size_categories:
- 100K<n<1M
---
## Data Card for Extractive CNN/DailyMail Dataset
### Overview
This is an extractive version of the [CNN/Dailymail](https://huggingface.co/datasets/cnn_dailymail) dataset. The structure of this dataset is identical to the original except for a minor modification in the data representation and the introduction of labels to denote the extractive summary.
The labels are generated following a greedy algorithm, as proposed by [Liu (2019)](https://arxiv.org/abs/1903.10318). The curation process can be found in the [bertsum-hf](https://github.com/eReverter/bertsum-hf) repository. I am uploading it in case someone does not want to go through the preprocessing, although Liu has a version ready for training in its [bertsum](https://github.com/nlpyang/BertSum) repository!
In this dataset:
- 'src' corresponds to 'article',
- 'tgt' equates to 'abstract',
- 'labels' represents a mapping of sentences forming the extractive summary.
### Data Architecture
Each entry in the dataset contains the following fields:
- `id`: a unique `string` identifier for each example.
- `src`: a `list[string]` field representing the original news article. Each string in the list is a separate sentence from the article.
- `tgt`: a `list[string]` field representing the professionally edited highlights or abstract of the article.
- `labels`: a `list[bool]` field with binary values. Each boolean value corresponds to a sentence in 'article', indicating whether that sentence is part of the extractive summary (1 for True, 0 for False).
### Sample Data Entry
Here is an illustrative example from the dataset:
```json
{
"id": "1",
"src": ["This is the first sentence",
"This is the second"],
"tgt": ["This is one of the highlights"],
"labels": [1, 0]
}
```
In this example, the first sentence of the article is selected as part of the extractive summary (as indicated by '1' in the 'labels'), while the second sentence is not ('0' in the 'labels').
### Usage
The extractive CNN/DailyMail dataset can be used to train and evaluate models for extractive text summarization tasks. It allows models to learn to predict which sentences from an original text contribute to a summary, providing a binary mapping as a reference. The 'tgt' or 'abstract' field can serve as a basis for comparison, helping to assess how well the selected sentences cover the key points in the abstract. |