Holographic CCG (Hol-CCG)

This is the Python implementation of Holographic CCG Parsing at ACL2023.

Prerequisites

• Python >= 3.8.8

• Install libraries

  pip install -r requirement.txt
  

• Prepare CCGbank

  • Obtain CCGbank and place it at dataset/ directory.
  • The name of the directory containing CCGbank data should be ccgbank_1_1/

• Preprocess CCGbank

  • Convert CCGbank format to Hol-CCG for training and evaluation.

    cd src/
    python preprocessing.py
    

  • Converted data is stored in dataset/ directory.
    • dataset/converted/ -> intermediate txt file generated for conversion
    • dataset/grammar/ -> .pickle files for word and phrase categories and other grammar-related information
    • dataset/tree_list/ -> .pickle file converted from CCG's constituency tree

Usage

Training of Hol-CCG

• Train Hol-CCG by CCGbank.

  python train_holccg.py
  

• The trained model is stored in model/ directory.

Supertagging by trained Hol-CCG

• Supertagging for a list of sentences by a trained Hol-CCG.

  python supertagging.py --path_to_sentence [path to sentence to be supertagged] --path_to_model [path to trained Hol-CCG]
  

Span-based Parsing using trained Hol-CCG

• Parsing a list of sentences by a Span-based Parsing algorithm incorporating a trained Hol-CCG.

  python span_parser.py --path_to_sentence [path to sentence to be parsed] --path_to_model [path to trained Hol-CCG]
  

• Use C&C Parser's generate program and evaluation scripts to evaluate parsing results.
• Convert parsing results (.auto format) to .html for visualization.

  python auto2html.py --path_to_autos [path to .auto file to be converted]
  

Phrase-level infilling experiment

• Infilling text at the level of phrase using Hol-CCG and RoBERTa.

• See Section 5.4 of the paper for details.

  Training of RoBERTa for phrase-level infilling

  • Train Hol-CCG by CCGbank for phrase-level infilling

    python train_infilling_roberta.py
    

  • The trained model is stored in model/ directory.

  Run the infilling experiment

  • Infilling on CCGbank data using trained Hol-CCG and RoBERTa.
  • Compute and compare syntactic agreement before and after infilling using Berkeley Parser.

    python infilling.py --path_to_holccg [path to trained Hol-CCG] --path_to_roberta [path to trained RoBERTa]
    

  Calculate PMI for Infilling results

  • Calculate and compare PMI for candidates infilled by Hol-CCG and RoBERTa.

    python calculate_pmi.py