Added v0.2.0

[KennethEnevoldsen]

2.7.0 (15/05/23)

• Updated the DaCy models to version 0.2.0, including a small, medium and large

  • Added beta support for Coreference Resolution! 🤩
  • Added beta support for Named Entity Linking!
  • Updated dependency parsing and part-of-speech tagging to use the latest version of the DDT treebank 🌳
  • Added a trainable lemmatizer, notably improving the lemmatization
  • All model are trained using the intersection between the CDT and the DDT treebanks (so actually trained on less data than before) 🤯
  • This includes the annotations from DaNED, DaCoref and DaNE

• Large model:

  • obtained state-of-the-art performance on:
  • Dependency parsing
  • Part-of-speech tagging
  • Morphological tagging
  • lemmatization (from 84.91 to 95.89!)
  • Improved performance on:
  • Reduced performance for NER down to 87.38 but we recommend either using the :code:nlp.add_pipe("dacy/ner") to add the SotA ScandiNER model to your pipeline or use one of the new fine-grained NER models.
  • Added support
  • Coreference Resolution, performance isn't great yet, but it's a start!
  • Named entity linking, with a precision of 0.86 but recall is still low due to a lacking knowledge base

• Medium model:

  • Consistent improvements across all tasks:
  • Notable performance gain for NER from an F1 of 81.79 to 85.82
  • Notable performance gain for lemmatization from an ACC 84.91 to 94
  • Added support
  • Coreference Resolution
  • Named entity linking

• Small model:

  • Consistent improvements across all tasks
  • Added support:
  • Coreference Resolution
  • Named entity linking

• Fixes a variety of issues:

  • Removed custom requirements for the large model
  • Removing warning when loading the models
  • Fixed annotatations in error in the DDT treebank where "'" was never followed by a space

• Removed support for DaCy model version 0.1.0, if you need to use these models you will have to use DaCy <= 2.0.0

• What is next?

  • A coreference resolution only model
  • Better named entity linking by
  • Improving the annotations of DaNED which currently annotate PERSONS using the QID reference to the name among other things
  • Improving the knowledge base which currently is the main source of low recall
  • Examine model generalization using DANSK and whether we can improve the generalization
  • and more!

[github-actions[bot]]

Tests	Skipped	Failures	Errors	Time
16	0 :zzz:	0 :x:	0 :fire:	54.148s :stopwatch: