This repository implements the parser described in the paper Deep Biaffine Attention for Neural Dependency Parsing.

I tried my best to exactly copy every detail in the tensorflow code (which is the original code from the paper's authors), including the weight initialization, choice of activation fucntion (and dropout), data preprocess, batch operation, MST algorithm, and, of course, hyperparameters.

For one who is interested, please tell me if you can find a difference.

On some version of WSJ data. This code only acheived UAS 95.56%, though the original code acheived UAS 95.59%.

  cd run
  python train.py --config_file ../configs/default.cfg --save_dir ../ckpt/default

  cd run
  python test.py --config_file ../configs/default.cfg --load_dir ../ckpt/default --output_file here

All configuration options (see in run/config.py) can be specified on the command line, but it's much easier to instead store them in a configuration file like configs/default.cfg.