Overall

• Major rewrite of NLP pipeline
• Many miscellaneous improvements and fixes

NLP

• Many performance and speed improvments improvements to NER
• Rewritten, ~2x faster dependency parser
• Rewritten hierarchical cross-document coreference
• Rewritten universal schema relation extraction model and epistemological database
• Faster tokenization with JFlex (50x faster, ~500k tokens/second on modest machine)
• Improvements to model, lexicon, and other resource finding and loading
• New name parser

Linear Algebra

• Several bug fixes to sparse tensors

Learning

• Support for simple constraints in optimization (projected gradient)
• Support for different constraints, regularization, and optimizers over different feature template subsets
• Added support for grid-searching subsets when optimizing hyperparameters
• Bug fixes to Exponentiated Gradient

Inference

• LiteChainModel class for simple chain CRFs

Miscellaneous

• Improved unit- and integration-testing coverage
• Improved serialization