Reusable

[duncanka]

• Improved code structure that is easier to adapt to other neural transition tagging tasks
• Made it possible for CNN library to release memory (necessary for repeatedly creating networks in the same run)
• Fixed command line documentation

I've confirmed that the pretrained model still gets a UAS of 93.55% / LAS of 92.16% on the test set.