Datasets:
wmt
/

Modalities:
Text
Formats:
parquet
Libraries:
Datasets
Dask
License:

Warn about issues with the Common Crawl corpus data

#2
by albertvillanova HF staff - opened
WMT: Workshop on Statistical Machine Translation org
edited Dec 2, 2022

After investigating this issue, we have discovered that there are issues with the source data file of the Common Crawl corpus: https://www.statmt.org/wmt13/training-parallel-commoncrawl.tgz

  • Non-English files contain many English sentences.
  • Their "parallel" sentences in English are not aligned: they are uncorrelated with their counterpart

This PR adds a warning to the dataset card explaining this.

albertvillanova changed pull request status to open
albertvillanova changed pull request status to merged

Hello, I was using the wmt14 and saw a similar trend, several german sentences are completely unrelated to the English counterpart:

image.png

Does the issue proliferate the wmt14 as well? If yes, is there a solution to this like a better dataset or a tool that can help filter out the noisy sentence pairs? Thanks!

WMT: Workshop on Statistical Machine Translation org

Hi

WMT14 is now 10 years old. I would recommend you use data from the latest WMT (http://www2.statmt.org/wmt24/).

All parallel data contains some noise however, and there are many approaches to filtering. For instance, have a look at the WMT20 filtering task for ideas. https://www.statmt.org/wmt20/parallel-corpus-filtering.html

best
Barry

Sign up or log in to comment