Deep learning for NER
We reuse the CommandLineUtils.java in PA2 to simplify parameter tuning process.
Here are the command line options we support:
-window: window size for the model, default is set to 13.
-layers: hidden layers structure, default is set to 300, you can also set multi-level parameter, e.g.: 100, 80 this means two hidden layer neural network, the first one is 100 and the second one is 80.
-alpha: learning rate, default is set to 0.001.
-regularize: regularizing constanct, default is set to 0.0001.
-epoch: number of iteration during the train, default it set to 10.
-data: folder contains necessary train, dev and test data.
-train: train filename.
-test: test filename.
-dump: dump the trained word vectors to filename if you specified, default is turned off.
-v: print the objective function value after every checkpoint.
A example command with our best settings: window size = 13, only one hidden layer with size 250, default alpha and regulaization constant is like:
cd $PROJECT
java -Xmx2g -cp classes:extlib/ejml.jar cs224n.deep.NER -window 13 -layer 250 -epoch 10 -train train -test test -data $DATA