|
--- |
|
language: |
|
- en |
|
multilinguality: |
|
- monolingual |
|
size_categories: |
|
- 100K<n<1M |
|
task_categories: |
|
- summarization |
|
- text-generation |
|
task_ids: [] |
|
tags: |
|
- conditional-text-generation |
|
dataset_info: |
|
- config_name: document |
|
features: |
|
- name: article |
|
dtype: string |
|
- name: abstract |
|
dtype: string |
|
splits: |
|
- name: train |
|
num_bytes: 7118099012 |
|
num_examples: 203037 |
|
- name: validation |
|
num_bytes: 216177092 |
|
num_examples: 6436 |
|
- name: test |
|
num_bytes: 216560120 |
|
num_examples: 6440 |
|
download_size: 3614462676 |
|
dataset_size: 7550836224 |
|
- config_name: section |
|
features: |
|
- name: article |
|
dtype: string |
|
- name: abstract |
|
dtype: string |
|
splits: |
|
- name: train |
|
num_bytes: 7201498856 |
|
num_examples: 203037 |
|
- name: validation |
|
num_bytes: 218793173 |
|
num_examples: 6436 |
|
- name: test |
|
num_bytes: 219196348 |
|
num_examples: 6440 |
|
download_size: 3649452807 |
|
dataset_size: 7639488377 |
|
configs: |
|
- config_name: document |
|
data_files: |
|
- split: train |
|
path: document/train-* |
|
- split: validation |
|
path: document/validation-* |
|
- split: test |
|
path: document/test-* |
|
- config_name: section |
|
data_files: |
|
- split: train |
|
path: section/train-* |
|
- split: validation |
|
path: section/validation-* |
|
- split: test |
|
path: section/test-* |
|
default: true |
|
train-eval-index: |
|
- config: document |
|
task: summarization |
|
task_id: summarization |
|
splits: |
|
eval_split: test |
|
col_mapping: |
|
article: text |
|
abstract: target |
|
--- |
|
|
|
# Arxiv dataset for summarization |
|
|
|
Dataset for summarization of long documents.\ |
|
Adapted from this [repo](https://github.com/armancohan/long-summarization).\ |
|
Note that original data are pre-tokenized so this dataset returns " ".join(text) and add "\n" for paragraphs. \ |
|
This dataset is compatible with the [`run_summarization.py`](https://github.com/huggingface/transformers/tree/master/examples/pytorch/summarization) script from Transformers if you add this line to the `summarization_name_mapping` variable: |
|
```python |
|
"ccdv/arxiv-summarization": ("article", "abstract") |
|
``` |
|
|
|
### Data Fields |
|
|
|
- `id`: paper id |
|
- `article`: a string containing the body of the paper |
|
- `abstract`: a string containing the abstract of the paper |
|
|
|
### Data Splits |
|
|
|
This dataset has 3 splits: _train_, _validation_, and _test_. \ |
|
Token counts are white space based. |
|
|
|
| Dataset Split | Number of Instances | Avg. tokens | |
|
| ------------- | --------------------|:----------------------| |
|
| Train | 203,037 | 6038 / 299 | |
|
| Validation | 6,436 | 5894 / 172 | |
|
| Test | 6,440 | 5905 / 174 | |
|
|
|
|
|
# Cite original article |
|
``` |
|
@inproceedings{cohan-etal-2018-discourse, |
|
title = "A Discourse-Aware Attention Model for Abstractive Summarization of Long Documents", |
|
author = "Cohan, Arman and |
|
Dernoncourt, Franck and |
|
Kim, Doo Soon and |
|
Bui, Trung and |
|
Kim, Seokhwan and |
|
Chang, Walter and |
|
Goharian, Nazli", |
|
booktitle = "Proceedings of the 2018 Conference of the North {A}merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers)", |
|
month = jun, |
|
year = "2018", |
|
address = "New Orleans, Louisiana", |
|
publisher = "Association for Computational Linguistics", |
|
url = "https://aclanthology.org/N18-2097", |
|
doi = "10.18653/v1/N18-2097", |
|
pages = "615--621", |
|
abstract = "Neural abstractive summarization models have led to promising results in summarizing relatively short documents. We propose the first model for abstractive summarization of single, longer-form documents (e.g., research papers). Our approach consists of a new hierarchical encoder that models the discourse structure of a document, and an attentive discourse-aware decoder to generate the summary. Empirical results on two large-scale datasets of scientific papers show that our model significantly outperforms state-of-the-art models.", |
|
} |
|
``` |
|
|
|
|