metadata

license: mit
datasets:
  - Riksarkivet/mini_cleaned_diachronic_swe
language:
  - sv
metrics:
  - perplexity
pipeline_tag: fill-mask
widget:
  - text: >-
      Det vore [MASK] häller nödvändigt att bita af tungan än berättat hvad jag
      varit med om.
train-eval-index:
  - config: Riksarkivet/mini_cleaned_diachronic_swe
    task: fill-mask
    task_id: fill-mask
    splits:
      eval_split: test
    col_mapping:
      text: text
model-index:
  - name: bert-base-cased-swe-historical
    results:
      - task:
          type: fill-mask
          name: fill-mask
        dataset:
          name: Riksarkivet/mini_cleaned_diachronic_swe
          type: Riksarkivet/mini_cleaned_diachronic_swe
          split: test
        metrics:
          - type: perplexity
            value: 3.42
            name: Perplexity (WIP)

Historical Swedish Bert Model

** WORK IN PROGRESS ** (Will be updated with bigger datasets soon + new OCR is coming to extend the dataset even further)

A historical Swedish Bert model is released from the National Swedish Archives to better generalise to Swedish historical text. Researches are well-aware that the Swedish language has been subject to change over time which means that present-day point-of-view models less ideal candidates for the job. However, this model can be used to interpret and analyse historical textual material and be fine-tuned for different downstream tasks.

Intended uses & limitations

This model should primarly be used to fine-tune further on and downstream tasks.

Inference for fill-mask with Huggingface Transformers in python:

from transformers import pipeline

summarizer = pipeline("fill-mask", model="Riksarkivet/bert-base-cased-swe-historical")
historical_text = """Det vore [MASK] häller nödvändigt att bita af tungan än berättat hvad jag varit med om."""
print(summarizer(historical_text))

Model Description

The training procedure can be recreated from here: Src_code. The preprocessing procedure can be recreated from here: Src_code.

Model: The following hyperparameters were used during training:

learning_rate: 3e-05
train_batch_size: 8
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 0
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 6
fp16: False

Dataset (WIP):

Khubist2, which has been cleaned and chunked. (will be further extended)

Acknowledgements

We gratefully acknowledge EuroHPC for funding this research by providing computing resources of the HPC system Vega and SWE-clarin for the datasets.

Citation Information

Eva Pettersson and Lars Borin (2022) Swedish Diachronic Corpus In Darja Fišer & Andreas Witt (eds.), CLARIN. The Infrastructure for Language Resources. Berlin: deGruyter. https://degruyter.com/document/doi/10.1515/9783110767377-022/html