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. 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 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.
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 | 0.9012 | 0.9069 | 0.905 | 0.9048 | 0.9037 ± 0.0027 |
The result in bold shows the performance of this model.
Additionally, the Flair training log and TensorBoard logs are also uploaded to the model hub.