The viewer is disabled because this dataset repo requires arbitrary Python code execution. Please consider
removing the
loading script
and relying on
automated data support
(you can use
convert_to_parquet
from the datasets
library). If this is not possible, please
open a discussion
for direct help.
Dataset Card for XSum Hallucination Annotations
Dataset Summary
Neural abstractive summarization models are highly prone to hallucinate content that is unfaithful to the input document. The popular metric such as ROUGE fails to show the severity of the problem. This dataset contains a large scale human evaluation of several neural abstractive summarization systems to better understand the types of hallucinations they produce. The dataset consists of faithfulness and factuality annotations of abstractive summaries for the XSum dataset. The dataset has crowdsourced 3 judgements for each of 500 x 5 document-system pairs. This will be a valuable resource to the abstractive summarization community.
Supported Tasks and Leaderboards
summarization
: : The dataset can be used to train a model for Summarization,, which consists in summarizing a given document. Success on this task is typically measured by achieving a high/low ROUGE Score.
Languages
The text in the dataset is in English which are abstractive summaries for the XSum dataset. The associated BCP-47 code is en
.
Dataset Structure
Data Instances
Faithfulness annotations dataset
A typical data point consists of an ID referring to the news article(complete document), summary, and the hallucination span information.
An example from the XSum Faithfulness dataset looks as follows:
{
'bbcid': 34687720,
'hallucinated_span_end': 114,
'hallucinated_span_start': 1,
'hallucination_type': 1,
'summary': 'rory mcilroy will take a one-shot lead into the final round of the wgc-hsbc champions after carding a three-under',
'system': 'BERTS2S',
'worker_id': 'wid_0'
}
Factuality annotations dataset
A typical data point consists of an ID referring to the news article(complete document), summary, and whether the summary is factual or not.
An example from the XSum Factuality dataset looks as follows:
{
'bbcid': 29911712,
'is_factual': 0,
'summary': 'more than 50 pupils at a bristol academy have been sent home from school because of a lack of uniform.',
'system': 'BERTS2S',
'worker_id': 'wid_0'
}
Data Fields
Faithfulness annotations dataset
Raters are shown the news article and the system summary, and are tasked with identifying and annotating the spans that aren't supported by the input article. The file contains the following columns:
bbcid
: Document id in the XSum corpus.system
: Name of neural summarizer.summary
: Summary generated by ‘system’.hallucination_type
: Type of hallucination: intrinsic (0) or extrinsic (1)hallucinated_span
: Hallucinated span in the ‘summary’.hallucinated_span_start
: Index of the start of the hallucinated span.hallucinated_span_end
: Index of the end of the hallucinated span.worker_id
: Worker ID (one of 'wid_0', 'wid_1', 'wid_2')
The hallucination_type
column has NULL value for some entries which have been replaced iwth -1
.
Factuality annotations dataset
Raters are shown the news article and the hallucinated system summary, and are tasked with assessing the summary whether it is factual or not. The file contains the following columns:
- `bbcid1: Document id in the XSum corpus.
system
: Name of neural summarizer.summary
: Summary generated by ‘system’.is_factual
: Yes (1) or No (0)worker_id
: Worker ID (one of 'wid_0', 'wid_1', 'wid_2')
The is_factual
column has NULL value for some entries which have been replaced iwth -1
.
Data Splits
There is only a single split for both the Faithfulness annotations dataset and Factuality annotations dataset.
train | |
---|---|
Faithfulness annotations | 11185 |
Factuality annotations | 5597 |
Dataset Creation
Curation Rationale
[More Information Needed]
Source Data
Initial Data Collection and Normalization
[More Information Needed]
Who are the source language producers?
[More Information Needed]
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
[More Information Needed]
Other Known Limitations
[More Information Needed]
Additional Information
Dataset Curators
[More Information Needed]
Licensing Information
Creative Commons Attribution 4.0 International
Citation Information
@InProceedings{maynez_acl20,
author = "Joshua Maynez and Shashi Narayan and Bernd Bohnet and Ryan Thomas Mcdonald",
title = "On Faithfulness and Factuality in Abstractive Summarization",
booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
year = "2020",
pages = "1906--1919",
address = "Online",
}
Contributions
Thanks to @vineeths96 for adding this dataset.
- Downloads last month
- 121