Version 4.2

v4.2

Release v4.2

Version 3.9

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

Version 3.8

V3_8

Bump to next release version.

Version 3.7

Avoid using ivy and include all jars for all dependencies for now. Adapt License to updated Mallet license.

Version 3.6

Avoid clashes with a pre-loaded commons-cli library, avoid training after an exception during processing the documents in the training set.

Version 3.5.4

The -prune parameter for Mallet C45 now requires a boolean value (true / false).

Version 3.5.3

Small change to how ngrams are treated: now, if an ngram occurs multiple times for an instance, we always accumulate the score of the ngram. This is done if we just count ngrams or if we get a score for the ngram from the featurename4value feature, or if we calculate a score by multiplication for n-grams with n>1 in either of the two cases.

Version 3.5.2

Fix in the cross-validation evalutation for regression.

Version 3.5.1

• Initial support for server-based application
• Initial support for Keras
• Support for using dense vectors as attributes (e.g. embeddings)
• Support for using tf*idf or other scores for attributes and ngrams

Version 3.5

Important changes:

• Support instance weights
• initial code to support per instance classification costs and addition of experimental costcla support, however this does not actually work yet
• add support to get the actual feature values for one-of-k coded nominal attributes from another feature. This should make it easier to use (externally precalculated) TF/IDF values for this.