File size: 4,772 Bytes
413a53b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
language: fr
license: mit
tags:
- flair
- token-classification
- sequence-tagger-model
base_model: dbmdz/bert-tiny-historic-multilingual-cased
widget:
- text:  469 . Πεδία . Les tribraques formés par un seul mot sont rares chez les
    tragiques , partont ailleurs qu  au premier pied .  . cependant QEd , Roi ,
    719 , 826 , 4496 .
---

# Fine-tuned Flair Model on AjMC French NER Dataset (HIPE-2022)

This Flair model was fine-tuned on the
[AjMC French](https://github.com/hipe-eval/HIPE-2022-data/blob/main/documentation/README-ajmc.md)
NER Dataset using hmBERT Tiny as backbone LM.

The AjMC dataset consists of NE-annotated historical commentaries in the field of Classics,
and was created in the context of the [Ajax MultiCommentary](https://mromanello.github.io/ajax-multi-commentary/)
project.

The following NEs were annotated: `pers`, `work`, `loc`, `object`, `date` and `scope`.

# Results

We performed a hyper-parameter search over the following parameters with 5 different seeds per configuration:

* Batch Sizes: `[4, 8]`
* Learning Rates: `[5e-05, 3e-05]`

And report micro F1-score on development set:

| Configuration     | Seed 1       | Seed 2          | Seed 3       | Seed 4       | Seed 5       | Average         |
|-------------------|--------------|-----------------|--------------|--------------|--------------|-----------------|
| `bs4-e10-lr5e-05` | [0.5346][1]  | [**0.5752**][2] | [0.543][3]   | [0.5043][4]  | [0.5531][5]  | 0.542 ± 0.026   |
| `bs8-e10-lr5e-05` | [0.4919][6]  | [0.5134][7]     | [0.473][8]   | [0.4909][9]  | [0.5082][10] | 0.4955 ± 0.016  |
| `bs4-e10-lr3e-05` | [0.4902][11] | [0.4944][12]    | [0.4676][13] | [0.494][14]  | [0.4986][15] | 0.489 ± 0.0123  |
| `bs8-e10-lr3e-05` | [0.4695][16] | [0.4836][17]    | [0.4372][18] | [0.4922][19] | [0.5015][20] | 0.4768 ± 0.0251 |

[1]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs4-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-1
[2]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs4-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-2
[3]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs4-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-3
[4]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs4-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-4
[5]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs4-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-5
[6]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs8-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-1
[7]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs8-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-2
[8]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs8-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-3
[9]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs8-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-4
[10]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs8-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-5
[11]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs4-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-1
[12]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs4-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-2
[13]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs4-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-3
[14]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs4-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-4
[15]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs4-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-5
[16]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs8-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-1
[17]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs8-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-2
[18]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs8-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-3
[19]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs8-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-4
[20]: https://hf.co/stefan-it/hmbench-ajmc-fr-hmbert_tiny-bs8-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-5

The [training log](training.log) and TensorBoard logs (not available for hmBERT Base model) 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 ❤️