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:

Tag	Name	Description
`PER`	Person	The name of a person (e.g., Birgitte and Mohammed)
`LOC`	Location	The name of a location (e.g., Tyskland and Djurgården)
`ORG`	Organisation	The name of an organisation (e.g., Bunnpris and Landsbankinn)
`MISC`	Miscellaneous	A named entity of a different kind (e.g., Ūjķnustu pund and Mona Lisa)

Quick start

You can use this model in your scripts as follows:

python

>>> 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:

Model ID	DaNE	NorNE-NB	NorNE-NN	SUC 3.0	WikiANN-IS	WikiANN-FO	Average
saattrupdan/nbailab-base-ner-scandi	87.44 ± 0.81	91.06 ± 0.26	90.42 ± 0.61	88.37 ± 0.17	88.61 ± 0.41	90.22 ± 0.46	89.08 ± 0.46
chcaa/da_dacy_large_trf	83.61 ± 1.18	78.90 ± 0.49	72.62 ± 0.58	53.35 ± 0.17	50.57 ± 0.46	51.72 ± 0.52	63.00 ± 0.57
RecordedFuture/Swedish-NER	64.09 ± 0.97	61.74 ± 0.50	56.67 ± 0.79	66.60 ± 0.27	34.54 ± 0.73	42.16 ± 0.83	53.32 ± 0.69
Maltehb/danish-bert-botxo-ner-dane	69.25 ± 1.17	60.57 ± 0.27	35.60 ± 1.19	38.37 ± 0.26	21.00 ± 0.57	27.88 ± 0.48	40.92 ± 0.64
Maltehb/-l-ctra-danish-electra-small-uncased-ner-dane	70.41 ± 1.19	48.76 ± 0.70	27.58 ± 0.61	35.39 ± 0.38	26.22 ± 0.52	28.30 ± 0.29	39.70 ± 0.61
radbrt/nb_nocy_trf	56.82 ± 1.63	68.20 ± 0.75	69.22 ± 1.04	31.63 ± 0.29	20.32 ± 0.45	12.91 ± 0.50	38.08 ± 0.75

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

Model ID	Samples/second	Model size
saattrupdan/nbailab-base-ner-scandi	4.16 ± 0.18	676 MB
chcaa/da_dacy_large_trf	0.65 ± 0.01	2,090 MB

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

Training Loss	Epoch	Step	Validation Loss	Micro F1	Micro F1 No Misc
0.6682	1.0	2816	0.0872	0.6916	0.7306
0.0684	2.0	5632	0.0464	0.8167	0.8538
0.0444	3.0	8448	0.0367	0.8485	0.8783
0.0349	4.0	11264	0.0316	0.8684	0.8920
0.0282	5.0	14080	0.0290	0.8820	0.9033
0.0231	6.0	16896	0.0283	0.8854	0.9060
0.0189	7.0	19712	0.0253	0.8964	0.9156
0.0155	8.0	22528	0.0260	0.9016	0.9201
0.0123	9.0	25344	0.0266	0.9059	0.9233
0.0098	10.0	28160	0.0280	0.9091	0.9279
0.008	11.0	30976	0.0309	0.9093	0.9287
0.0065	12.0	33792	0.0313	0.9103	0.9284
0.0053	13.0	36608	0.0322	0.9078	0.9257
0.0046	14.0	39424	0.0343	0.9075	0.9256

Framework versions

Transformers 4.10.3
Pytorch 1.9.0+cu102
Datasets 1.12.1
Tokenizers 0.10.3