• Description

  ---

• Corpora

  ---

• Taggers

  ---

• Models
  • TnT
    • SUC
    • Bootstrapped
  • HunPos
    • SUC
    • Bootstrapped

  ---

• Requirements

  ---

• References

  ---

• News

  ---

Tagger models

TnT models

TnT models were trained with the setting -c, which gives better results for Swedish (particularly for unknown words). With this setting TnT internally adds a capitalisation label to each tag.

The models were evaluated with default settings.

SUC models

Both the cross-validated (10x = 7.87±0.21% unknown) and the full (ideal = no unknown words) models were evaluated. Just to get an idea of how much better the full model would do, I performed a leave-one-out evaluation, where each of the 500 2000-word files is left out in turn, leaving 7.54±3.03% unknown words.

Bootstrapped models

For small models, such as those trained on SUC only, the default tagger settings (-n3 -a10 -u3 -d4) give the best performance in general. For large models, such as the bootstrapped models, the default settings, except for sparse data mode -d1 (replace 0 by 0.5) instead of -d4 (automatic linear interpolation), give the best performance. (I have only tested the default subsettings for sparse data mode.)

Therefore, for bootstrapping, the default settings are used when tagging the bootstrap corpus with the SUC model, but the -d1 setting should be used when tagging (or evaluating) with the bootstrapped models.

The following models were compared to the corresponding models for HunPos in Forsbom (2008b).