Generative Constituency Parser with (linear-time) Discriminative Recognition Algorithm

The code implements a neural sequence-to-tree model in the context of constituency parsing, provides a unifination of discriminative and generative RNNG, and is capable of doing constituency parsing and language modeling.

Data

• English PTB, to generate oracles find scripts in the data folder

Dependencies

• Dynet (2.0)
• Numpy

Instructions

• Training (session_supervised_enc.py, session_supervised_dec.py, session_unsupervised.py)
• Testing (session_lm.py, session_parsing.py)
• For parsing, there are two choices: 1) find the argmax tree from the approximated posterior q(a|x); 2) find the sampled tree from q(a|x) which maximizes the joint p(a,x)
• For language modeling, there are three choices: 1) lower bound approximation; 2) importance sampling using variational distribution as proposal; and 3) directly sampling from prior
• Focused training depending on the final objective: if parsing is the goal, we focus on maximizing q(a|x) and p(a,x); if language modeling is the goal, we focus on maximizing p(x)

Extra feature

To experiment the encoder without look-ahead attention feature, replace encoder.py with encoder_no_attention.py.

Citation

@InProceedings{cheng2017virnng, 
  author = {Cheng, Jianpeng and Lopez, Adam and Lapata, Mirella}, 
  title = {A Generative Parser with a Discriminative Recognition Algorithm}, 
  booktitle = {Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Short Papers)}, 
  year = {2017}, 
  address = {Vancouver, Canada}, 
  publisher = {Association for Computational Linguistics} 
 }

Contact

[email protected]