Commit
•
b84ed2c
1
Parent(s):
28d5429
Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: Multilingual
|
| 3 |
+
datasets:
|
| 4 |
+
- deepset/germanquad
|
| 5 |
+
license: mit
|
| 6 |
+
thumbnail: https://thumb.tildacdn.com/tild3433-3637-4830-a533-353833613061/-/resize/720x/-/format/webp/germanquad.jpg
|
| 7 |
+
tags:
|
| 8 |
+
- exbert
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
## Overview
|
| 14 |
+
**Language model:** deepset/xlm-roberta-base-squad2-distilled
|
| 15 |
+
**Language:** German
|
| 16 |
+
**Training data:** GermanQuAD train set (~ 12MB)
|
| 17 |
+
**Eval data:** GermanQuAD test set (~ 5MB)
|
| 18 |
+
**Infrastructure**: 1x V100 GPU
|
| 19 |
+
**Published**: Apr 21st, 2021
|
| 20 |
+
|
| 21 |
+
## Details
|
| 22 |
+
- We trained a German question answering model with a gelectra-base model as its basis.
|
| 23 |
+
- The dataset is GermanQuAD, a new, German language dataset, which we hand-annotated and published [online](https://deepset.ai/germanquad).
|
| 24 |
+
- The training dataset is one-way annotated and contains 11518 questions and 11518 answers, while the test dataset is three-way annotated so that there are 2204 questions and with 2204·3−76 = 6536answers, because we removed 76 wrong answers.
|
| 25 |
+
- In addition to the annotations in GermanQuAD, haystack's distillation feature was used for training. deepset/xlm-roberta-large-squad2-distilled was used as the teacher model.
|
| 26 |
+
|
| 27 |
+
See https://deepset.ai/germanquad for more details and dataset download in SQuAD format.
|
| 28 |
+
|
| 29 |
+
## Hyperparameters
|
| 30 |
+
```
|
| 31 |
+
batch_size = 24
|
| 32 |
+
n_epochs = 6
|
| 33 |
+
max_seq_len = 384
|
| 34 |
+
learning_rate = 3e-5
|
| 35 |
+
lr_schedule = LinearWarmup
|
| 36 |
+
embeds_dropout_prob = 0.1
|
| 37 |
+
temperature = 2
|
| 38 |
+
distillation_loss_weight = 0.75
|
| 39 |
+
```
|
| 40 |
+
## Performance
|
| 41 |
+
We evaluated the extractive question answering performance on the SQuAD v2 dev set.
|
| 42 |
+
Model types and training data are included in the model name.
|
| 43 |
+
For finetuning XLM-Roberta, we use the English SQuAD v2.0 dataset.
|
| 44 |
+
The GELECTRA models are warm started on the German translation of SQuAD v1.1 and finetuned on \\\\germanquad.
|
| 45 |
+
The human baseline was computed for the 3-way test set by taking one answer as prediction and the other two as ground truth.
|
| 46 |
+
```
|
| 47 |
+
"exact": 79.8366040596311%
|
| 48 |
+
"f1": 83.916407079888%
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
## Authors
|
| 53 |
+
- Timo Möller: `timo.moeller [at] deepset.ai`
|
| 54 |
+
- Julian Risch: `julian.risch [at] deepset.ai`
|
| 55 |
+
- Malte Pietsch: `malte.pietsch [at] deepset.ai`
|
| 56 |
+
- Michel Bartels: `michel.bartels [at] deepset.ai`
|
| 57 |
+
## About us
|
| 58 |
+
|
| 59 |
+
We bring NLP to the industry via open source!
|
| 60 |
+
Our focus: Industry specific language models & large scale QA systems.
|
| 61 |
+
|
| 62 |
+
Some of our work:
|
| 63 |
+
- [German BERT (aka "bert-base-german-cased")](https://deepset.ai/german-bert)
|
| 64 |
+
- [GermanQuAD and GermanDPR datasets and models (aka "gelectra-base-germanquad", "gbert-base-germandpr")](https://deepset.ai/germanquad)
|
| 65 |
+
- [FARM](https://github.com/deepset-ai/FARM)
|
| 66 |
+
- [Haystack](https://github.com/deepset-ai/haystack/)
|
| 67 |
+
|
| 68 |
+
Get in touch:
|
| 69 |
+
[Twitter](https://twitter.com/deepset_ai) | [LinkedIn](https://www.linkedin.com/company/deepset-ai/) | [Slack](https://haystack.deepset.ai/community/join) | [GitHub Discussions](https://github.com/deepset-ai/haystack/discussions) | [Website](https://deepset.ai)
|
| 70 |
+
|
| 71 |
+
By the way: [we're hiring!](http://www.deepset.ai/jobs)
|