|
--- |
|
language: fi |
|
license: mit |
|
tags: |
|
- flair |
|
- token-classification |
|
- sequence-tagger-model |
|
base_model: hmbyt5-preliminary/byt5-small-historic-multilingual-span20-flax |
|
inference: false |
|
widget: |
|
- text: Rooseveltin sihteeri ilmoittaa perättö - mäksi tiedon , että Rooseveltia olisi |
|
kehotettu käymään Englannissa , Saksassa ja Venäjällä puhumassa San Franciscon |
|
näyttelyn puolesta . |
|
--- |
|
|
|
# Fine-tuned Flair Model on Finnish NewsEye NER Dataset (HIPE-2022) |
|
|
|
This Flair model was fine-tuned on the |
|
[Finnish 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 | [0.8017][1] | [0.7406][2] | [0.7706][3] | [0.7759][4] | [0.7983][5] | 77.74 ± 2.2 | |
|
| bs4-e10-lr0.00015 | [0.7729][6] | [0.7553][7] | [0.7526][8] | [0.7547][9] | [0.7913][10] | 76.54 ± 1.49 | |
|
| bs8-e10-lr0.00016 | [0.6638][11] | [0.5875][12] | [0.78][13] | [0.7804][14] | [0.7176][15] | 70.59 ± 7.34 | |
|
| bs8-e10-lr0.00015 | [0.6783][16] | [0.5867][17] | [0.7229][18] | [0.7761][19] | [0.697][20] | 69.22 ± 6.22 | |
|
|
|
[1]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-1 |
|
[2]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-2 |
|
[3]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-3 |
|
[4]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-4 |
|
[5]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-5 |
|
[6]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-1 |
|
[7]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-2 |
|
[8]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-3 |
|
[9]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-4 |
|
[10]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-5 |
|
[11]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-1 |
|
[12]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-2 |
|
[13]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-3 |
|
[14]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-4 |
|
[15]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-5 |
|
[16]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-1 |
|
[17]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-2 |
|
[18]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-3 |
|
[19]: https://hf.co/hmbench/hmbench-newseye-fi-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-4 |
|
[20]: https://hf.co/hmbench/hmbench-newseye-fi-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 ❤️ |
|
|
