20_Newsgroups_Fixed / README.md
albertvillanova's picture
Fix `license` metadata (#1)
147dd30
metadata
annotations_creators:
  - crowdsourced
language_creators:
  - crowdsourced
language:
  - en
license:
  - unknown
multilinguality:
  - monolingual
pretty_name: 20_Newsgroups_Fixed
size_categories:
  - 10K<n<100K
source_datasets:
  - original
task_categories:
  - text-classification
task_ids:
  - multi-class-classification
  - topic-classification

Dataset Card for 20_Newsgroups_Fixed

Table of Contents

Dataset Description

Dataset Summary

This dataset is a version of the 20 Newsgroups dataset fixed with the help of the Galileo ML Data Intelligence Platform. In a matter of minutes, Galileo enabled us to uncover and fix a multitude of errors within the original dataset. In the end, we present this improved dataset as a new standard for natural language experimentation and benchmarking using the Newsgroups dataset.

Curation Rationale

This dataset was created to showcase the power of Galileo as a Data Intelligence Platform. Through Galileo, we identify critical error patterns within the original Newsgroups training dataset - garbage data that do not properly fit any newsgroup label category. Moreover, we observe that these errors permeate throughout the test dataset.

As a result of our analysis, we propose the addition of a new class to properly categorize and fix the labeling of garbage data samples: a "None" class. Galileo further enables us to quickly make these data sample changes within the training set (changing garbage data labels to None) and helps guide human re-annotation of the test set.

Total Dataset Errors Fixed: 1163 (6.5% of the dataset)

Errors / Split. Overall Train Test
Garbage samples fixed 718 396 322
Empty samples fixed 445 254 254
Total samples fixed 1163 650 650

To learn more about the process of fixing this dataset, please refer to our Blog.

Dataset Structure

Data Instances

For each data sample, there is the text of the newsgroup post, the corresponding newsgroup forum where the message was posted (label), and a data sample id.

An example from the dataset looks as follows:

{'id': 1,
'text': 'I have win 3.0 and downloaded several icons and BMP\'s but I can\'t figure out\nhow to change the "wallpaper" or use the icons.  Any help would be appreciated.\n\n\nThanx,\n\n-Brando'
'label': comp.os.ms-windows.misc}

Data Fields

  • id: the unique numerical id associated with a data sample
  • text: a string containing the text of the newsgroups message
  • label: a string indicating the newsgroup forum where the sample was posted

Data Splits

The data is split into a training and test split. To reduce bias and test generalizability across time, data samples are split between train and test depending upon whether their message was posted before or after a specific date, respectively.

Data Classes

The fixed data is organized into 20 newsgroup topics + a catch all "None" class. Some of the newsgroups are very closely related to each other (e.g. comp.sys.ibm.pc.hardware / comp.sys.mac.hardware), while others are highly unrelated (e.g misc.forsale / soc.religion.christian). Here is a list of the 21 classes, partitioned according to subject matter:

comp.graphics
comp.os.ms-windows.misc
comp.sys.ibm.pc.hardware
comp.sys.mac.hardware
comp.windows.x
rec.autos
rec.motorcycles
rec.sport.baseball
rec.sport.hockey
sci.crypt
<sci.electronics
sci.med
sci.space
misc.forsale talk.politics.misc
talk.politics.guns
talk.politics.mideast
talk.religion.misc
alt.atheism
soc.religion.christian
None