Skip to content

Version 3.9
Compare
Choose a tag to compare
@johann-petrak johann-petrak released this 20 Apr 11:28
· 379 commits to master since this release
V3_9
892d89c

NOTE/IMPORT: some PRs in this version are not drop-in compatible with the previous version since their parameters have changed (supporting multiple classes for sequence tagging needed the class annotation type parameter to be a list instead of a single value). Also, models may not be compatible (a model trained with a previous version may not work with this version) since there have been chances of which features get generated and how they are named.

CHANGES (In reverse order of time):

  • Fix issue #32
  • Fix problem with handling the exception correctly when the trainer class cannot be instantiated
  • Fix bug: incorrected trainer class for MALLET_CL_NAIVE_BAYES
  • Better handling of escaping/cleaning strings when exporting to tsv/csv format
  • Better handling of missing values for number-coded nominal features when exporting to ARFF
  • Properly escape empty string when exporting to ARFF, treat null like an empty string
  • More runtime parameters are optional
  • Add JSON exporter
  • Implement START/STOP symbols
  • In many places, move from assuming files to using URLs
  • Major refactoring: the Engine instance now knows which CorpusRepresentation to use and returns it
  • Added support to use previous target(s) as a feature
  • Make classification using a sequence algorithm work
  • Implement PRs for word shape features and affix features generation
  • Implement multi-class sequence tagging
Assets 3
Loading