stefan-it's picture
readme: mention dataset creation notebook
159b0ab verified
metadata
license: cc-by-4.0
task_categories:
  - token-classification
language:
  - de

Filtered GermEval 2014 NER Dataset

This repository hosts a filtered version of the great GermEval 2014 NER Dataset.

After some analysis of the annotated examples in this dataset, it can be seen that the dataset is highly biased by Wikipedia articles.

Dataset Stats

We present an overview of the top 10 top-level domains where annotations were retrieved from for training, development and test splits:

Training Split

TLD Number of examples (Percentage)
wikipedia.org 12,007 (50.03%)
welt.de 662 (2.76%)
spiegel.de 512 (2.13%)
tagesspiegel.de 424 (1.77%)
handelsblatt.com 369 (1.54%)
fr-aktuell.de 344 (1.43%)
sueddeutsche.de 308 (1.28%)
abendblatt.de 283 (1.18%)
berlinonline.de 255 (1.06%)
szon.de 249 (1.04%)

Development Split

TLD Number of examples (Percentage)
wikipedia.org 1,119 (50.86%)
welt.de 46 (2.09%)
spiegel.de 43 (1.95%)
fr-aktuell.de 38 (1.73%)
tagesspiegel.de 37 (1.68%)
handelsblatt.com 35 (1.59%)
sueddeutsche.de 28 (1.27%)
szon.de 25 (1.14%)
feedsportal.com 24 (1.09%)
berlinonline.de 22 (1.0%)

Test Split

TLD Number of examples (Percentage)
wikipedia.org 2,547 (49.94%)
welt.de 139 (2.73%)
spiegel.de 88 (1.73%)
tagesspiegel.de 86 (1.69%)
handelsblatt.com 84 (1.65%)
sueddeutsche.de 78 (1.53%)
abendblatt.de 72 (1.41%)
fr-aktuell.de 62 (1.22%)
berlinonline.de 59 (1.16%)
szon.de 57 (1.12%)

Summary

For each dataset split it can be seen, that the portion of annotated examples from Wikipedia are around 50%!

Filtered Version & Motivation

We now create a Wikipedia-filtered-out version of the GermEval 2014 dataset. Here's one scenario for the main motivation:

Imagine you are pretraining a nice language model and you want to measure performance on GermEval 2014 for named entity recognition. Additionally, you want of course to compare performance to other existing language models.

What would be the easiest way to get high performance on GermEval 2014 dataset? Yes, you can literally pretrain a language model on Wikipedia only (just as I did)! It will outperform models that are even pretrained on 100+ GB! See the great ScandEval leaderboard and have a look at the gwlms models. However, the model performance for this pretrained model on Wikipedia-only will be worse on other downstream tasks such as Question Answering.

So this Wikipedia-filtered-out version could help to achieve better comparisons between LMs.

Stats for Filtered Version

Additionally, we now present the stats for the filtered version of GermEval 2014 dataset:

Training Split

TLD Number of examples (Percentage)
welt.de 662 (5.52%)
spiegel.de 512 (4.27%)
tagesspiegel.de 424 (3.54%)
handelsblatt.com 369 (3.08%)
fr-aktuell.de 344 (2.87%)
sueddeutsche.de 308 (2.57%)
abendblatt.de 283 (2.36%)
berlinonline.de 255 (2.13%)
szon.de 249 (2.08%)
n-tv.de 195 (1.63%)

Development Split

TLD Number of examples (Percentage)
welt.de 46 (4.26%)
spiegel.de 43 (3.98%)
fr-aktuell.de 38 (3.52%)
tagesspiegel.de 37 (3.42%)
handelsblatt.com 35 (3.24%)
sueddeutsche.de 28 (2.59%)
szon.de 25 (2.31%)
feedsportal.com 24 (2.22%)
berlinonline.de 22 (2.04%)
rp-online.de 21 (1.94%)

Test Split

TLD Number of examples (Percentage)
welt.de 139 (5.44%)
spiegel.de 88 (3.45%)
tagesspiegel.de 86 (3.37%)
handelsblatt.com 84 (3.29%)
sueddeutsche.de 78 (3.06%)
abendblatt.de 72 (2.82%)
fr-aktuell.de 62 (2.43%)
berlinonline.de 59 (2.31%)
szon.de 57 (2.23%)
feedsportal.com 52 (2.04%)

Dataset Creation

We provide a notebook that shows how to recreate this filtered version of GermEval 2014. It can be found here.

Additionally, we provide a dataset loader for the awesome Flair library!

Licence

We keep the original license of GermEval 2014 dataset ( CC-BY-4.0).