Imitation Learning for Unsupervised Parsing

This release contains the code used for paper An Imitation Learning Approach to Unsupervised Parsing
Code for PRPN and Gumbel Tree-LSTM is borrowed from PRPN codebase and NYU's implementation respectively.

Preparation

Requirements:

• Python 2.7.5
• Pytorch 0.3.1

Data to download:

• 840B word vectors 300D: GloVe
• All-NLI: MultiNLI and SNLI

Running

1. Train PRPN using its original code base
2. Forward PRPN model on All-NLI data to get predicted trees: ./python/prpn_util/generate_distance.sh
3. Shuffle the generated training set and then split it to train/dev set
   cat train.json | shuf > train_shuffled.json
head -n TRAIN_NUM train_shuffled.json > train_shuffled_train.json
tail -n VALID_NUM train_shuffled.json > train_shuffled_valid.json
4. Step-by-step supervised learning
   ./python/sl_rl.sh
5. Policy refinement
   ./python/rl_ft.sh