Question regarding the german translation of SQuAD v1.1 mentioned for GELECTRA fine tuning

[TiloMichel]

Hi,

I'm currently trying to fine tune a model for the task of end-to-end question generation in German.
I used your published GermanQuAD and transformed it for the task, though it generated questions they weren't mostly suitable for the context (I haven't evaluated with automatic metrics yet). It may be that I need a larger pretrained LM (I currently use a german version of GPT2), a multilingual LM or more training data.

Since I don't have exactly a clue how much training data for the fine-tuning is necessary and most approaches are using the English SQuAD dataset, I searched for an automatic translated SQuAD dataset and found something here https://huggingface.co/deepset/gelectra-base-germanquad#performance .

My question is now if you could share that automatic translation of SQuAD v1.1 and if not maybe give me an insight with which tools you approached this translation.

Thanks in advance and best regards

[bogdankostic]

Hey @TiloMichel,

The German part of the MLQA dataset was used for the performance comparisons. Hope this helps :)

[TiloMichel]

Thanks for the answer.
When I read on huggingface I assumed there would exist a german translated version of the SQuAD v1.1 dataset (with all 100.000+ qa pairs). The MLQA split for german isn't that big either (about 4.500 qa pairs) so I think the impact may not be that big, but it can of course be used for the task.

[TiloMichel]

I think there was a misunderstanding from my side. The MLQA dataset itself is not that large, but in a subsection of the readme file in their repository there is a machine translated version of the squad v1.1 dataset for german https://github.com/facebookresearch/MLQA#translate-train-and-translate-test-data

[julian-risch]

As a coincidence, we just released a blogpost on NLP Resources Beyond English today 😁 https://www.deepset.ai/blog/nlp-resources-beyond-english

[TiloMichel]

Great timing, thanks 😅