ScandiNER - Named Entity Recognition model for Scandinavian Languages

Check out a demo of the model here.

This model is a fine-tuned version of NbAiLab/nb-bert-base for Named Entity Recognition for Danish, Norwegian (both Bokmål and Nynorsk), Swedish, Icelandic and Faroese. It has been fine-tuned on the concatenation of DaNE, NorNE, SUC 3.0 and the Icelandic and Faroese parts of the WikiANN dataset. It also works reasonably well on English sentences, given the fact that the pretrained model is also trained on English data along with Scandinavian languages.

The model will predict the following four entities:

Quick start

You can use this model in your scripts as follows:

>>> from transformers import pipeline
>>> import pandas as pd
>>> ner = pipeline(task='ner', 
...                model='saattrupdan/nbailab-base-ner-scandi', 
...                aggregation_strategy='first')
>>> result = ner('Borghild kjøper seg inn i Bunnpris')
>>> pd.DataFrame.from_records(result)
  entity_group     score      word  start  end
0          PER  0.981257  Borghild      0    8
1          ORG  0.974099  Bunnpris     26   34

Performance

The following is the Micro-F1 NER performance on Scandinavian NER test datasets, compared with the current state-of-the-art. The models have been evaluated on the test set along with 9 bootstrapped versions of it, with the mean and 95% confidence interval shown here:

Aside from its high accuracy, it's also substantially smaller and faster than the previous state-of-the-art:

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 90135.90000000001
• num_epochs: 1000

Training results

Framework versions

• Transformers 4.10.3
• Pytorch 1.9.0+cu102
• Datasets 1.12.1
• Tokenizers 0.10.3