kelm / README.md
albertvillanova's picture
Reorder split names (#1)
cee05ec
metadata
annotations_creators:
  - found
language_creators:
  - found
language:
  - en
license:
  - cc-by-sa-3.0
multilinguality:
  - monolingual
size_categories:
  - 1M<n<10M
source_datasets:
  - original
task_categories:
  - other
task_ids: []
paperswithcode_id: kelm
pretty_name: Corpus for Knowledge-Enhanced Language Model Pre-training (KELM)
tags:
  - data-to-text-generation
dataset_info:
  features:
    - name: triple
      dtype: string
    - name: sentence
      dtype: string
  splits:
    - name: train
      num_bytes: 1343187306
      num_examples: 6371131
    - name: validation
      num_bytes: 167790917
      num_examples: 796471
    - name: test
      num_bytes: 167921750
      num_examples: 796493
  download_size: 1631259869
  dataset_size: 1678899973

Dataset Card for Corpus for Knowledge-Enhanced Language Model Pre-training (KELM)

Table of Contents

Dataset Description

Dataset Summary

Data-To-Text Generation involves converting knowledge graph (KG) triples of the form (subject, relation, object) into a natural language sentence(s). This dataset consists of English KG data converted into paired natural language text. The generated corpus consists of ∼18M sentences spanning ∼45M triples with ∼1500 distinct relations.

Supported Tasks and Leaderboards

The intended task is data-to-text generation, taking in a knowledge graph tuple and generating a natural language representation from it. Specifically, the data is in the format the authors used to train a seq2seq language model with the tuples concatenated into a single sequence.

Languages

The dataset is in English.

Dataset Structure

Data Instances

Each instance consists of one KG triple paired with corresponding natural language.

Data Fields

  • triple: Wikipedia triples of the form <subject> <relation> <object> where some subjects have multiple relations, e.g. <subject> <relation1> <object1> <relation2> <object2> <relation3> <object3>. For more details on how these relations are grouped, please refer to the paper.
  • sentence: The corresponding Wikipedia sentence.

Data Splits

The dataset includes a pre-determined train, validation, and test split.

Dataset Creation

Curation Rationale

The goal of the dataset's curation and the associated modeling work discussed in the paper is to be able to generate natural text from a knowledge graph.

Source Data

Initial Data Collection and Normalization

[More Information Needed]

Who are the source language producers?

The data is sourced from English Wikipedia and it's associated knowledge graph.

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

[More Information Needed]

Discussion of Biases

From the paper:

Wikipedia has documented ideological, gender6, and racial biases in its text. While the KELM corpus may still contain some of these biases, certain types of biases may be reduced.

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

This dataset has been released under the CC BY-SA 2.0 license.

Citation Information

@misc{agarwal2020large,
      title={Large Scale Knowledge Graph Based Synthetic Corpus Generation for Knowledge-Enhanced Language Model Pre-training}, 
      author={Oshin Agarwal and Heming Ge and Siamak Shakeri and Rami Al-Rfou},
      year={2020},
      eprint={2010.12688},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Contributions

Thanks to @joeddav for adding this dataset.