metadata

tags:
  - flair
  - token-classification
  - sequence-tagger-model
language: uk
datasets:
  - ner-uk
model-index:
  - name: flair-uk-ner
    results:
      - task:
          name: NER
          type: token-classification
        metrics:
          - name: NER Precision
            type: precision
            value: 0.8616
          - name: NER Recall
            type: recall
            value: 0.8593
          - name: NER F Score
            type: f_score
            value: 0.8605
widget:
  - text: >-
      Президент Володимир Зеленський пояснив, що наразі діалог із режимом
      Володимира путіна неможливий, адже агресор обрав курс на знищення
      українського народу. За словами Зеленського цей режим РФ виявляє неповагу
      до суверенітету і територіальної цілісності України.
license: mit

flair-uk-ner

Model description

flair-uk-ner is a Flair model that is ready to use for Named Entity Recognition. It is based on flair embeddings, that I've trained for Ukrainian language (available here and here) and has nice performance and a very small size (just 72mb!).

It has been trained to recognize four types of entities: location (LOC), organizations (ORG), person (PERS) and Miscellaneous (MISC).

Results:

F-score (micro) 0.8605
F-score (macro) 0.7472
Accuracy 0.8033

by class	precision	recall	f1-score	support
PERS	0.9305	0.9422	0.9363	1678
LOC	0.8150	0.8678	0.8406	401
ORG	0.6653	0.6092	0.6360	261
MISC	0.6202	0.5375	0.5759	240
micro avg	0.8616	0.8593	0.8605	2580
macro avg	0.7577	0.7392	0.7472	2580
weighted avg	0.8569	0.8593	0.8575	2580

The model was fine-tuned on the NER-UK dataset, released by the lang-uk. Training code is also available here.