metadata
language: de
license: mit
tags:
- german
We released the German Question Answering model fine-tuned with our own German Question Answering dataset (deQuAD) containing 130k training and 11k test QA pairs.
Overview
- Language model: electra-base-german-uncased
- Language: German
- Training data: deQuAD2.0 training set (~42MB)
- Evaluation data: deQuAD2.0 test set (~4MB)
- Infrastructure: 8xV100 GPU
Evaluation
We benchmarked the question answering performance on our deQuAD test data with some German language models. The fine-tuned electra-base-german-uncased model gives the best performance (Exact Match/F1).
| Model | All | HasAns | NoAns |
|---|---|---|---|
| electra-base-german-uncased | 70.97/76.18 | 67.73/78.02 | 74.29/74.29 |
| bert-base-german-cased | 58.98/64.77 | 49.19/60.63 | 69.03/69.03 |
| bert-base-german-dbmdz-uncased | 63.70/68.00 | 57.03/65.52 | 70.51/70.51 |
| dbmdz/bert-base-german-europeana-uncased | 58.79/63.38 | 52.14/61.22 | 65.59/65.59 |