README.md

Augmenting Unsupevised Constituency Parsing with Rules

Code release for the paper

Augmenting Unsupevised Constituency Parsing with Rules

Atul Sahay*, Anshul Nasery*, Ayush Maheshwari, Ganesh Ramakrishnan, Rishabh Iyer

appearing in the Findings of ACL 2021.

This repo forks the official repo for DIORA and builds on it. Follow the steps in the repo to setup dependencies.

Evaluation

run bash run_all_models.sh /path/to/data to reproduce F1 score results.

Rules

Our rules can be found in the directory rulesets/. The folder rulesets/data_preprocessing contains our code to prepare the datasets by augmenting them with rules for training

Training

Follow instructions in the DIORA repo for preliminary instructions for training the model. Run the command

python3 diora/scripts/train.py \
          --arch mlp-shared \
          --batch_size 32 \
          --data_type nli\
          --elmo_cache_dir ~/path_to_dir\
          --emb elmo\
          --hidden_dim 400\
          --log_every_batch 500\
          --lr 3e-3\
          --normalize unit\
          --reconstruct_mode softmax\
          --save_after 1000\
          --train_filter_length 20\
          --train_path /path_to_data\
          --validation_path ~/path_to_data\
          --cuda\
          --use_reconstruction\
          --rule_based\
          --load_model_path ./checkpoints/diora-checkpoints/mlp-softmax-shared/model.pt

to start a training job

Citation

If you find the code useful for your research, consider citing us -

@inproceedings{sahay-etal-2021-rule,
    title = "Rule Augmented Unsupervised Constituency Parsing",
    author = "Sahay, Atul  and
      Nasery, Anshul  and
      Maheshwari, Ayush  and
      Ramakrishnan, Ganesh  and
      Iyer, Rishabh",
    booktitle = "Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021",
    month = aug,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.findings-acl.436",
    doi = "10.18653/v1/2021.findings-acl.436",
    pages = "4923--4932",
}