File size: 5,196 Bytes
19059d3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 |
---
language: fr
license: mit
tags:
- flair
- token-classification
- sequence-tagger-model
base_model: hmbyt5-preliminary/byt5-small-historic-multilingual-span20-flax
inference: false
widget:
- text: Le Moniteur universel fait ressortir les avantages de la situation de l '
Allemagne , sa force militaire , le peu d ' intérêts personnels qu ' elle peut
avoir dans la question d ' Orient .
---
# Fine-tuned Flair Model on French NewsEye NER Dataset (HIPE-2022)
This Flair model was fine-tuned on the
[French NewsEye](https://github.com/hipe-eval/HIPE-2022-data/blob/main/documentation/README-newseye.md)
NER Dataset using hmByT5 as backbone LM.
The NewsEye dataset is comprised of diachronic historical newspaper material published between 1850 and 1950
in French, German, Finnish, and Swedish.
More information can be found [here](https://dl.acm.org/doi/abs/10.1145/3404835.3463255).
The following NEs were annotated: `PER`, `LOC`, `ORG` and `HumanProd`.
# ⚠️ Inference Widget ⚠️
Fine-Tuning ByT5 models in Flair is currently done by implementing an own [`ByT5Embedding`][1] class.
This class needs to be present when running the model with Flair.
Thus, the inference widget is not working with hmByT5 at the moment on the Model Hub and is currently disabled.
This should be fixed in future, when ByT5 fine-tuning is supported in Flair directly.
[1]: https://github.com/stefan-it/hmBench/blob/main/byt5_embeddings.py
# Results
We performed a hyper-parameter search over the following parameters with 5 different seeds per configuration:
* Batch Sizes: `[8, 4]`
* Learning Rates: `[0.00015, 0.00016]`
And report micro F1-score on development set:
| Configuration | Run 1 | Run 2 | Run 3 | Run 4 | Run 5 | Avg. |
|-------------------|--------------|--------------|--------------|--------------|--------------|--------------|
| bs4-e10-lr0.00016 | [0.793][1] | [0.803][2] | [0.8054][3] | [0.8069][4] | [0.8133][5] | 80.43 ± 0.66 |
| bs8-e10-lr0.00016 | [0.7888][6] | [0.8094][7] | [0.8043][8] | [0.8011][9] | [0.8117][10] | 80.31 ± 0.8 |
| bs4-e10-lr0.00015 | [0.7884][11] | [0.8109][12] | [0.8005][13] | [0.8083][14] | [0.8022][15] | 80.21 ± 0.78 |
| bs8-e10-lr0.00015 | [0.788][16] | [0.8003][17] | [0.8067][18] | [0.8035][19] | [0.8064][20] | 80.1 ± 0.69 |
[1]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-1
[2]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-2
[3]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-3
[4]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-4
[5]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-5
[6]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-1
[7]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-2
[8]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-3
[9]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-4
[10]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-5
[11]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-1
[12]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-2
[13]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-3
[14]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-4
[15]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-5
[16]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-1
[17]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-2
[18]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-3
[19]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-4
[20]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-5
The [training log](training.log) and TensorBoard logs are also uploaded to the model hub.
More information about fine-tuning can be found [here](https://github.com/stefan-it/hmBench).
# Acknowledgements
We thank [Luisa März](https://github.com/LuisaMaerz), [Katharina Schmid](https://github.com/schmika) and
[Erion Çano](https://github.com/erionc) for their fruitful discussions about Historic Language Models.
Research supported with Cloud TPUs from Google's [TPU Research Cloud](https://sites.research.google/trc/about/) (TRC).
Many Thanks for providing access to the TPUs ❤️
|