bert-base-finnish-uncased-ner

Info

This is a fine-tuned model on the NER task. The original model is Turku NLP's bert-base-finnish-uncased-v1, and the fine-tuning dataset is Turku NLP's turku_ner_corpus.

The model is released under Apache 2.0.

Please mention the training dataset if you use this model:

@inproceedings{luoma-etal-2020-broad,
    title = "A Broad-coverage Corpus for {F}innish Named Entity Recognition",
    author = {Luoma, Jouni and Oinonen, Miika and Pyyk{\"o}nen, Maria and Laippala, Veronika and Pyysalo, Sampo},
    booktitle = "Proceedings of The 12th Language Resources and Evaluation Conference",
    year = "2020",
    url = "https://www.aclweb.org/anthology/2020.lrec-1.567",
    pages = "4615--4624",
}

Validation Metrics

• Loss: 0.075
• Accuracy: 0.982
• Precision: 0.879
• Recall: 0.868
• F1: 0.873

Test Metrics

Overall Metrics

• Accuracy: 0.986
• Precision: 0.857
• Recall: 0.872
• F1: 0.864

Per-entity metrics

{
    "DATE": {
        "precision": 0.925,
        "recall": 0.9736842105263158,
        "f1": 0.9487179487179489,
        "number": "114"
    },
    "EVENT": {
        "precision": 0.3,
        "recall": 0.42857142857142855,
        "f1": 0.3529411764705882,
        "number": "7"
    },
    "LOC": {
        "precision": 0.9057239057239057,
        "recall": 0.9372822299651568,
        "f1": 0.9212328767123287,
        "number": "287"
    },
    "ORG": {
        "precision": 0.8274111675126904,
        "recall": 0.7836538461538461,
        "f1": 0.8049382716049382,
        "number": "208"
    },
    "PER": {
        "precision": 0.88,
        "recall": 0.9225806451612903,
        "f1": 0.9007874015748031,
        "number": "310"
    },
    "PRO": {
        "precision": 0.6081081081081081,
        "recall": 0.569620253164557,
        "f1": 0.5882352941176471,
        "number": "79"
    }
}

Usage

You can use cURL to access this model:

$ curl -X POST -H "Authorization: Bearer YOUR_API_KEY" -H "Content-Type: application/json" -d '{"inputs": "Asun Brysselissä, Euroopan pääkaupungissa."}' https://api-inference.huggingface.co/models/iguanodon-ai/bert-base-finnish-uncased-ner

Or Python API:

from transformers import AutoModelForTokenClassification, AutoTokenizer

model = AutoModelForTokenClassification.from_pretrained("iguanodon-ai/bert-base-finnish-uncased-ner")
tokenizer = AutoTokenizer.from_pretrained("iguanodon-ai/bert-base-finnish-uncased-ner")

inputs = tokenizer("Asun Brysselissä, Euroopan pääkaupungissa.", return_tensors="pt")
outputs = model(**inputs)