Bert Base Swedish Cased

🚀 Swedish BERT Models

The National Library of Sweden / KBLab releases three pre - trained language models based on BERT and ALBERT, aiming to provide a representative BERT model for Swedish text.

The National Library of Sweden / KBLab has released three pre - trained language models based on BERT and ALBERT. These models are trained on approximately 15 - 20GB of text (200M sentences, 3000M tokens) from diverse sources such as books, news, government publications, the Swedish Wikipedia, and internet forums. The goal is to offer a representative BERT model for Swedish text. A more comprehensive description will be published later.

The following three models are currently available:

• bert - base - swedish - cased (v1) - A BERT model trained with the same hyperparameters as the one initially published by Google.
• bert - base - swedish - cased - ner (experimental) - A BERT model fine - tuned for Named - Entity Recognition (NER) using the SUC 3.0 dataset.
• albert - base - swedish - cased - alpha (alpha) - An initial attempt at creating an ALBERT model for Swedish.

All models are case - sensitive and trained with whole - word masking.

📦 Files

TensorFlow model weights will be released soon.

📦 Installation

The examples below require Huggingface Transformers 2.4.1 and Pytorch 1.3.1 or greater. For Transformers < 2.4.0, the tokenizer must be instantiated manually, and the do_lower_case flag parameter should be set to False and keep_accents to True (for ALBERT).

To create an environment where the examples can be run, execute the following commands in a terminal on your preferred operating system:

# git clone https://github.com/Kungbib/swedish - bert - models
# cd swedish - bert - models
# python3 -m venv venv
# source venv/bin/activate
# pip install --upgrade pip
# pip install -r requirements.txt

💻 Usage Examples

Basic Usage

BERT Base Swedish

A standard BERT base for Swedish trained on a variety of sources. The vocabulary size is approximately 50k. Using Huggingface Transformers, the model can be loaded in Python as follows:

from transformers import AutoModel,AutoTokenizer

tok = AutoTokenizer.from_pretrained('KB/bert-base-swedish-cased')
model = AutoModel.from_pretrained('KB/bert-base-swedish-cased')

BERT base fine - tuned for Swedish NER

This model is fine - tuned on the SUC 3.0 dataset. Using the Huggingface pipeline, the model can be easily instantiated. For Transformer < 2.4.1, it seems the tokenizer must be loaded separately to disable lower - casing of input strings:

from transformers import pipeline

nlp = pipeline('ner', model='KB/bert-base-swedish-cased-ner', tokenizer='KB/bert-base-swedish-cased-ner')

nlp('Idag släpper KB tre språkmodeller.')

Running the Python code above should produce a result similar to the following. The entity types used are TME for time, PRS for personal names, LOC for locations, EVN for events, and ORG for organizations. These labels are subject to change.

[ { 'word': 'Idag', 'score': 0.9998126029968262, 'entity': 'TME' },
  { 'word': 'KB',   'score': 0.9814832210540771, 'entity': 'ORG' } ]

The BERT tokenizer often splits words into multiple tokens, with the sub - parts starting with ##. For example, the string Engelbert kör Volvo till Herrängens fotbollsklubb gets tokenized as Engel ##bert kör Volvo till Herr ##ängens fotbolls ##klubb. To glue the parts back together, you can use the following code:

text = 'Engelbert tar Volvon till Tele2 Arena för att titta på Djurgården IF ' +\
       'som spelar fotboll i VM klockan två på kvällen.'

l = []
for token in nlp(text):
    if token['word'].startswith('##'):
        l[-1]['word'] += token['word'][2:]
    else:
        l += [ token ]

print(l)

This should result in something like the following (though less neatly formatted):

[ { 'word': 'Engelbert',     'score': 0.99..., 'entity': 'PRS'},
  { 'word': 'Volvon',        'score': 0.99..., 'entity': 'OBJ'},
  { 'word': 'Tele2',         'score': 0.99..., 'entity': 'LOC'},
  { 'word': 'Arena',         'score': 0.99..., 'entity': 'LOC'},
  { 'word': 'Djurgården',    'score': 0.99..., 'entity': 'ORG'},
  { 'word': 'IF',            'score': 0.99..., 'entity': 'ORG'},
  { 'word': 'VM',            'score': 0.99..., 'entity': 'EVN'},
  { 'word': 'klockan',       'score': 0.99..., 'entity': 'TME'},
  { 'word': 'två',           'score': 0.99..., 'entity': 'TME'},
  { 'word': 'på',            'score': 0.99..., 'entity': 'TME'},
  { 'word': 'kvällen',       'score': 0.54..., 'entity': 'TME'} ]

ALBERT base

The easiest way to use this model is also with Huggingface Transformers:

from transformers import AutoModel,AutoTokenizer

tok = AutoTokenizer.from_pretrained('KB/albert-base-swedish-cased-alpha')
model = AutoModel.from_pretrained('KB/albert-base-swedish-cased-alpha')

📄 Acknowledgements

• Resources from Stockholms University, Umeå University, and the Swedish Language Bank at Gothenburg University were used when fine - tuning BERT for NER.
• Model pre - training was partly conducted in - house at the KBLab and partly (for material without active copyright) with the support of Cloud TPUs from Google's TensorFlow Research Cloud (TFRC).
• The models are hosted on S3 by Huggingface 🤗