thivy/norwegian-ner-combined

Norwegian NER Combined Dataset

Dataset Description

This dataset combines the NorNE (Norwegian Named Entities) and WikiANN Norwegian datasets for Named Entity Recognition (NER) in Norwegian (Bokmål and Nynorsk).

Key Features

• ✅ 49,870 training samples (NorNE + WikiANN combined)
• ✅ 14,289 validation samples
• ✅ 13,450 test samples
• ✅ 4 entity types: PER, ORG, LOC, MISC
• ✅ Quality filtered: 12 problematic samples removed from NorNE
• ✅ Entity remapping: 9 original types → 4 core types

Languages

• Norwegian Bokmål (nob)
• Norwegian Nynorsk (nno)

Entity Types

Type	Description	Examples
PER	Person names	Erna Solberg, Ibsen
ORG	Organizations	Stortinget, NATO
LOC	Locations	Oslo, Norge, Europa
MISC	Miscellaneous	Nobels fredspris

Dataset Sources

1. NorNE (Norwegian Named Entities)

• Repository: https://github.com/ltgoslo/norne
• Format: Universal Dependencies (CoNLL-U)
• Variants: Bokmål and Nynorsk
• Annotation: Manual annotation by Language Technology Group, University of Oslo

2. WikiANN (Pan-lingual NER)

• Source: HuggingFace Datasets (wikiann, Norwegian subset)
• Origin: Wikipedia articles
• Annotation: Automatic annotation from Wikipedia markup

Data Processing

Quality Filtering

12 problematic samples removed from NorNE:

Category	Count	Reason
Single-word samples	6	No context for NER
Annotation errors	3	Missing/incorrect labels
Very short samples	3	High loss, no entities

Examples of filtered samples:

• "Sharia" - Single word, missing annotation
• "Ombudsmannen" - Single word, no context
• "Verdensarv ." - 2 tokens, no entities, critical loss increase

Entity Type Remapping

Original NorNE had 9 entity types, remapped to 4 standard types:

GPE_ORG → ORG  (Geo-political organizations)
GPE_LOC → LOC  (Geo-political locations)
PROD    → O    (Dropped: Products)
DRV     → O    (Dropped: Derived terms)
EVT     → O    (Dropped: Events)

Kept:
- PER (Person)
- ORG (Organization)
- LOC (Location)
- MISC (Miscellaneous)

Why? These 4 types align with standard NER benchmarks and had sufficient training data from WikiANN.

Dataset Structure

Data Splits

Split	Samples	Bytes	Tokens (avg)
Train	49,870	8.0 MB	~15
Validation	14,289	2.3 MB	~15
Test	13,450	2.2 MB	~15

Data Fields

• id (string): Unique identifier (e.g., "train_0")
• tokens (list[string]): Tokenized sentence
• ner_tags (list[int]): NER tags in IOB2 format

Label Schema (IOB2)

0: O       (Outside any entity)
1: B-LOC   (Beginning of Location)
2: B-MISC  (Beginning of Miscellaneous)
3: B-ORG   (Beginning of Organization)
4: B-PER   (Beginning of Person)
5: I-LOC   (Inside Location)
6: I-MISC  (Inside Miscellaneous)
7: I-ORG   (Inside Organization)
8: I-PER   (Inside Person)

Usage

Load Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("thivy/norwegian-ner-combined")

# Access splits
train_data = dataset["train"]
val_data = dataset["validation"]
test_data = dataset["test"]

# Example
print(train_data[0])

Output:

{
    'id': 'train_0',
    'tokens': ['Oslo', 'er', 'hovedstaden', 'i', 'Norge', '.'],
    'ner_tags': [1, 0, 0, 0, 1, 0]  # B-LOC, O, O, O, B-LOC, O
}

Training Example

from transformers import AutoTokenizer, AutoModelForTokenClassification, TrainingArguments, Trainer

# Load model and tokenizer
model_name = "NbAiLab/nb-bert-base"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(
    model_name, 
    num_labels=9  # 9 labels in IOB2 format
)

# Load dataset
dataset = load_dataset("thivy/norwegian-ner-combined")

# Tokenize and align labels
def tokenize_and_align_labels(examples):
    tokenized_inputs = tokenizer(
        examples["tokens"],
        truncation=True,
        is_split_into_words=True
    )
    
    labels = []
    for i, label in enumerate(examples["ner_tags"]):
        word_ids = tokenized_inputs.word_ids(batch_index=i)
        label_ids = []
        for word_idx in word_ids:
            if word_idx is None:
                label_ids.append(-100)
            else:
                label_ids.append(label[word_idx])
        labels.append(label_ids)
    
    tokenized_inputs["labels"] = labels
    return tokenized_inputs

tokenized_datasets = dataset.map(tokenize_and_align_labels, batched=True)

# Train
training_args = TrainingArguments(
    output_dir="./norwegian-ner-model",
    evaluation_strategy="epoch",
    learning_rate=2e-5,
    num_train_epochs=3,
    per_device_train_batch_size=8,
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_datasets["train"],
    eval_dataset=tokenized_datasets["validation"],
)

trainer.train()

Training Results

A model trained on this dataset achieved:

Metric	Score
F1	0.9329
Precision	0.93
Recall	0.93
Best Epoch	5

Model: Coming soon
Base Model: NbAiLab/nb-bert-base

Training Configuration

• Learning Rate: 3.5e-5 (cosine with restarts)
• Epochs: 20 (early stopped at 11)
• Batch Size: 8
• Weight Decay: 0.15
• Optimizer: AdamW

Dataset Statistics

Entity Distribution (Training Set)

Entity	Count	Percentage
PER	~12,000	40%
ORG	~9,000	30%
LOC	~7,500	25%
MISC	~1,500	5%

Sample Length Distribution

Min tokens: 1
Max tokens: 512
Mean tokens: 15.2
Median tokens: 12

Limitations

1. Domain: Primarily news text and Wikipedia
2. Formality: More formal Norwegian, less conversational
3. Entity coverage: MISC category is underrepresented
4. Temporal: Data may not reflect recent entities/organizations

Ethical Considerations

• News articles may contain biased language
• Person names in dataset are public figures
• Geographic and organizational names may be politically sensitive

Citation

NorNE Dataset

@inproceedings{jorgensen-etal-2020-norne,
    title = "{N}or{NE}: Annotating Named Entities for {N}orwegian",
    author = "J{\o}rgensen, Fredrik  and
      Aasen, Tobias  and
      R{\o}nningstad, Anne-Katrine  and
      Indreb{\o}, Kristin Hagen  and
      Lyse, Gunn Inger",
    booktitle = "Proceedings of the 12th Language Resources and Evaluation Conference",
    month = may,
    year = "2020",
    address = "Marseille, France",
    publisher = "European Language Resources Association",
    pages = "4547--4556",
}

WikiANN Dataset

@inproceedings{pan-etal-2017-cross,
    title = "Cross-lingual Name Tagging and Linking for 282 Languages",
    author = "Pan, Xiaoman  and
      Zhang, Boliang  and
      May, Jonathan  and
      Nothman, Joel  and
      Knight, Kevin  and
      Ji, Heng",
    booktitle = "Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    year = "2017",
}

nb-bert-base Model

@misc{kummervold2021operationalizing,
    title={Operationalizing a National Digital Library: The Case for a Norwegian Transformer Model},
    author={Per E Kummervold and Javier de la Rosa and Freddy Wetjen and Svein Arne Brygfjeld},
    year={2021},
    eprint={2104.09617},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

License

• NorNE: CC-BY 4.0
• WikiANN: CC-BY-SA 3.0
• This Dataset: CC-BY 4.0 (inheriting from NorNE)

Contact

For questions or issues with this dataset:

• GitHub: finetune-ner-norne-wikian
• HuggingFace: @thivy

Acknowledgments

• Language Technology Group (LTG) at University of Oslo for the NorNE dataset
• National Library of Norway AI Lab (NBAiLab) for nb-bert-base model
• HuggingFace for the Transformers library and dataset infrastructure