Dataset Card for Reddit Webis-TLDR-17

Dataset Summary

This corpus contains preprocessed posts from the Reddit dataset (Webis-TLDR-17). The dataset consists of 3,848,330 posts with an average length of 270 words for content, and 28 words for the summary.

Features includes strings: author, body, normalizedBody, content, summary, subreddit, subreddit_id. Content is used as document and summary is used as summary.

Supported Tasks and Leaderboards

Summarization (abstractive)

Known ROUGE scores achieved for the Webis-TLDR-17:

Model	ROUGE-1	ROUGE-2	ROUGE-L	Paper/Source
Transformer + Copy (Gehrmann et al., 2019)	22	6	17	Generating Summaries with Finetuned Language Models
Unified VAE + PGN (Choi et al., 2019)	19	4	15	VAE-PGN based Abstractive Model in Multi-stage Architecture for Text Summarization

(Source: https://github.com/sebastianruder/NLP-progress/blob/master/english/summarization.md)

Languages

English

Dataset Structure

Data Instances

default

• Size of downloaded dataset files: 3.14 GB
• Size of the generated dataset: 18.94 GB
• Total amount of disk used: 22.08 GB

An example of 'train' looks as follows.

{
    "author": "me",
    "body": "<>",
    "content": "input document.",
    "id": "1",
    "normalizedBody": "",
    "subreddit": "machinelearning",
    "subreddit_id": "2",
    "summary": "output summary."
}

Data Fields

The data fields are the same among all splits.

default

• author: a string feature.
• body: a string feature.
• normalizedBody: a string feature.
• subreddit: a string feature.
• subreddit_id: a string feature.
• id: a string feature.
• content: a string feature.
• summary: a string feature.

Data Splits

name	train
default	3848330

This corpus does not contain a separate test set. Thus it is up to the users to divide the corpus into appropriate training, validation and test sets.

Dataset Creation

Curation Rationale

In the scope of the task of absractive summarization the creators of the Webis-TLDR-17 propose mining social media for author-provided summaries and taking advantage of the common practice of appending a "TL;DR" to long posts. A large Reddit crawl was used to yield the Webis-TLDR-17 corpus. This dataset intends to complement the existing summarization corpora primarily from the news genre.

Source Data

Reddit subreddits posts (submissions & comments) containing "TL;DR" from 2006 to 2016. Multiple subreddits are included.

Initial Data Collection and Normalization

Initial data: a set of 286 million submissions and 1.6 billion comments posted to Reddit between 2006 and 2016. Then a five-step pipeline of consecutive filtering steps was applied.

Who are the source language producers?

The contents of the dataset are produced by human authors, bot-generated content was eliminated by filtering out all bot accounts with the help of an extensive list provided by the Reddit community, as well as manual inspection of cases where the user name contained the substring "bot."

Annotations

Annotation process

More Information Needed

Who are the annotators?

More Information Needed

Personal and Sensitive Information

More Information Needed

Considerations for Using the Data

Social Impact of Dataset

This dataset has been created to serve as a source of large-scale summarization training data. It is primarily geared towards the automatic abstractive summarization task, that can be considered one of the most challenging variants of automatic summarization. It also aims to tackle the lack of genre diversity in the summarization datasets (most are news-related).

Discussion of Biases

More Information Needed

Other Known Limitations

Reddit users write TL;DRs with various intentions, such as providing a “true” summary, asking questions or for help, or forming judgments and conclusions. As noted in the paper introducing the dataset, although the first kind of TL;DR posts are most important for training summarization models, yet, the latter allow for various alternative summarization-related tasks.

Although filtering was performed abusive language maybe still be present.

Additional Information

Dataset Curators

Michael Völske, Martin Potthast, Shahbaz Syed, Benno Stein

Licensing Information

More Information Needed

Citation Information

@inproceedings{volske-etal-2017-tl,
    title = "{TL};{DR}: Mining {R}eddit to Learn Automatic Summarization",
    author = {V{"o}lske, Michael  and
      Potthast, Martin  and
      Syed, Shahbaz  and
      Stein, Benno},
    booktitle = "Proceedings of the Workshop on New Frontiers in Summarization",
    month = sep,
    year = "2017",
    address = "Copenhagen, Denmark",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/W17-4508",
    doi = "10.18653/v1/W17-4508",
    pages = "59--63",
    abstract = "Recent advances in automatic text summarization have used deep neural networks to generate high-quality abstractive summaries, but the performance of these models strongly depends on large amounts of suitable training data. We propose a new method for mining social media for author-provided summaries, taking advantage of the common practice of appending a {``}TL;DR{''} to long posts. A case study using a large Reddit crawl yields the Webis-TLDR-17 dataset, complementing existing corpora primarily from the news genre. Our technique is likely applicable to other social media sites and general web crawls.",
}

Contributions

Thanks to @mariamabarham, @patrickvonplaten, @thomwolf for adding this dataset.