Taggerflow

This repository contains the code for training the supertagging model from LSTM CCG Parsing at NAACL (Lewis et al., 2016).

Dependencies

• Tensorflow (r0.11 or above)
• https://www.tensorflow.org/versions/r0.11/get_started/os_setup.html#pip-installation

Training and Evaluation

• python taggerflow.py grid.json
  • Trains a supertagging model.
  • Logs evaluation results.
  • Writes checkpoints to the log directory.

Exporting to EasySRL

• python taggerflow.py grid.json -c <checkpoint_path>
  • Evaluates the checkpoint on the dev set as a sanity check.
  • Exports the model information to a temporary directory.
  • Prints the temporary directory with the exported model.
• The temporary directory should contain graph.pb and various .txt files.
• Download and extract http://lil.cs.washington.edu/resources/model_tritrain_finetune.tgz, which provides the correct file structure.
• Remove the existing taggerflow directory and replace it with the temporary directory.

Running EasySRL

• Clone the EasySRL repository: https://github.com/uwnlp/EasySRL.
• Download http://lil.cs.washington.edu/resources/libtaggerflow.so and move it to the lib directory.
• EasySRL will use the trained supertagger for parsing when given the modified model_tritrain_finetune directory.