File size: 2,693 Bytes
da6f75d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
language:
- en
- ka
license: mit
tags:
- flair
- token-classification
- sequence-tagger-model
base_model: xlm-roberta-large
widget:
- text: ამით თავისი ქადაგება დაასრულა და დაბრუნდა იერუსალიმში . ერთ-ერთ გარე კედელზე
    არსებობს ერნესტო ჩე გევარას პორტრეტი . შაკოსკა“ ინახება ბრაზილიაში , სან-პაულუს
    ხელოვნების მუზეუმში .
---

# Fine-tuned English-Georgian NER Model with Flair

This Flair NER model was fine-tuned on the WikiANN dataset
([Rahimi et al.](https://www.aclweb.org/anthology/P19-1015) splits)
using XLM-R Large as backbone LM.

**Notice**: The dataset is very problematic, because it was automatically constructed.

We did manually inspect the development split of the Georgian data and found
a lot of bad labeled examples, e.g. DVD ( 💿 ) as `ORG`.

## Fine-Tuning

The latest
[Flair version](https://github.com/flairNLP/flair/tree/f30f5801df3f9e105ed078ec058b4e1152dd9159)
is used for fine-tuning.

We use English and Georgian training splits for fine-tuning and the
development set of Georgian for evaluation.

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

* Batch Sizes: [`4`]
* Learning Rates: [`5e-06`]

More details can be found in this [repository](https://github.com/stefan-it/georgian-ner).

## Results

A hyper-parameter search with 5 different seeds per configuration is performed and micro F1-score on development set
is reported:

| Configuration     | Seed 1          | Seed 2      | Seed 3      | Seed 4     | Seed 5      | Average         |
|-------------------|-----------------|-------------|-------------|------------|-------------|-----------------|
| `bs4-e10-lr5e-06` | [**0.9005**][1] | [0.9012][2] | [0.9069][3] | [0.905][4] | [0.9048][5] | 0.9037 ± 0.0027 |

[1]: https://hf.co/stefan-it/autotrain-flair-georgian-ner-xlm_r_large-bs4-e10-lr5e-06-1
[2]: https://hf.co/stefan-it/autotrain-flair-georgian-ner-xlm_r_large-bs4-e10-lr5e-06-2
[3]: https://hf.co/stefan-it/autotrain-flair-georgian-ner-xlm_r_large-bs4-e10-lr5e-06-3
[4]: https://hf.co/stefan-it/autotrain-flair-georgian-ner-xlm_r_large-bs4-e10-lr5e-06-4
[5]: https://hf.co/stefan-it/autotrain-flair-georgian-ner-xlm_r_large-bs4-e10-lr5e-06-5

The result in bold shows the performance of this model.

Additionally, the Flair [training log](training.log) and [TensorBoard logs](tensorboard) are also uploaded to the model
hub.