2.2.11b - January 31 2014
- An armload of bug-fixes related to numeric classes and features.
- Added widget to disable parallel extraction and learning, for memory's sake
- Fixed occasional save-model dialogue deadlock
- Improved cancellation behavior for long model training tasks
- Document lists that are part of feature tables or trained models are now decoupled from the document list displayed in the first tab.
- Random Swing errors (breaking all or part of the workbench UI) should be less frequent, and should display warning messages to the user when they do occur.
2.2.9b - December 11 2013
- Feature Extraction and Machine Learning plugins can now take advantage of multiple processors.
- BasicFeatures, CharacterNGrams, and all WekaCore learning plugins now inherit from a Parallel(Feature|Learning)Plugin superclass.
- Fixed a feature extraction bug when "Include Punctuation" was unchecked.
- Fixed several bugs relating to supplied test sets - in particular, with regard to "Differentiate Text Columns."
2.2.7b - October 14 2013
- Directories now appear in open/save dialogs
- Feature evaluation tables now sort correctly and display "?" for missing entries.
- Feature selection no longer prevents influence evaluation metric from displaying.
- DocumentList will not guess single-value columns as the default class label.
- Now reads and writes UTF-8 files - for CSV, side.xml, and ARFF
- Feature Table and Model save/load should be more consistent
- behind-the scenes support for saving/loading XML-formatted models and feature tables - this is not available in the UI yet.
- Split the "Remove Stopwords" option for Basic Features into "Skip Stopwords Within N-Grams" and "Ignore All-Stopword N-Grams". The first option builds N-grams that pretend stopwords aren't there (so the first part of this sentence would contain the N-grams "BOL_first", "first_part", and "part_sentence"), the second ignores any N-grams that are made entirely of stopwords ("the_first", "part_of", "this_sentence", but not "BOL_so", so_the" or "of_this") - this is what "Remove Stopwords" does in earlier versions of LightSide.
2.2.3b through 2.2.6b - October 2 2013
- Prediction recipe does not appear to save feature set, label array but does seem to save the document list in its entirety. (in the validation settings - I thought we fixed that.) -- doclist has to be null to trigger prediction-only? -- fixed!
- LocalFeatureHits which are present in a document with no hit locations are broken - fixed!
- OrderedPluginMap ordering can contain duplicates - fixed.
- predict without matching text column fails - disable or guess. - fixed
- Predict Button inconsistently disabled - fixed!
- prediction exceptions are not thrown to be caught by UI message - fixed
- options in save/load dialog are duplicated/inconsistent - esp load model for prediction - fixed
- button text on save model is "Save Feature Table" -- again? fixed
- should not be able to save Predict-only recipe as regular recipe - fixed
- should be able to save regular recipes as both - fixed
- errors when extracting docs without class/text columns chosen - mostly fixed
- new model not selected when loaded from prediciton pane - fixed, but not so that it loads in all the panels. Just the one that it was loaded in.
- numeric featuretables should auto-select a numeric-capable learner? ... no. The button is correctly disabled, but they users should still be able to explore their options.
- linelength feature does not render as a number nicely in metric tables - kappa, etc .- fixed by removing semi-intelligent per-column classtypes in feature(tablemodel). Still a problem in sorting the prediction results.
- combineFeatures is not changing the table - fixed!
- combineFeatures gets confused when nesting combofeatures - fixed!
- combineFeatures: should de-select features in big table after combo - fixed!
- force-canceling an actionbartask doesn't actually stop the thread - fixed, with brutal deprecated thread-stop duct tape.