README.md

metadata

language:
  - en
inference: false
pipeline_tag: false
datasets:
  - conll2003
  - wnut_17
  - jnlpba
  - conll2012
  - BTC
  - dfki-nlp/few-nerd
tags:
  - PyTorch
model-index:
  - name: bert-base-NER-reptile-5-datasets
    results:
      - task:
          name: few-shot-ner
          type: named-entity-recognition
        dataset:
          name: few-nerd-inter
          type: named-entity-recognition
        metrics:
          - name: 5 way 1~2 shot
            type: f1
            value: 56.12
          - name: 5-way 5~10-shot
            type: f1
            value: 62.7
          - name: 10-way 1~2-shot
            type: f1
            value: 50.3
          - name: 10-way 5~10-shot
            type: f1
            value: 58.82

BERT base uncased model pre-trained on 5 NER datasets

Model was trained by SberIDP. The pretraining process and technical details are described in this article.

• Task: Named Entity Recognition
• Base model: bert-base-uncased
• Training Data is 5 datasets: CoNLL-2003, WNUT17, JNLPBA, CoNLL-2012 (OntoNotes), BTC
• Testing was made in Few-Shot scenario on Few-NERD dataset using the model as a backbone for StructShot

The model is pretrained for NER task using Reptile and can be finetuned for new entities with only a small amount of samples.