Coreference and Ellipsis as QA

Code to reproduce the experiments in A Simple Transfer Learning Baseline for Ellipsis Resolution

Requires Python >= 3.5.0

Recommended: Create a conda environment with conda create -n myenv python=3.7

Conversion and Utilities

The repository contains conversion scripts for converting different datasets into the SQuAD 1.1 format.

• vpe2squad.py: Convert VP ellipsis dataset into SQuAD format
• conll2squad.py: Convert coreference data from C0NLL-2012 to SQuAD format
  • First convert .conll files to .jsonlines using this
  • Set ONTONOTES_DIR (ontonotes folder path) and set2fmt (filename to convert to SQuAD format)
  • Run script
• sluice2squad.py: Convert sluice ellipsis dataset into SQuAD format
• wikicoref2conll.py: Convert WikiCoref dataset into CoNLL-2012 format
• squad2conll.py: Convert the prediction files produced by bert/run_squad.py into CONLL format for evaluation

Miscellaneous

• annotate_qwords.py: Adds <ref> and </ref> tags to interrogation words in SQuAD files
• evaluate-v1.1.py: Standard SQuAD v1.1 evaluation script (for evaluating ellipsis)

For coreference resolution, use the standard CoNLL-2012 script after converting the predictions into the CoNLL-2012 format using squad2conll.py.

Training Details

Each model folder contains pre-processing, configuration, training and evaluation scripts for Sluice Ellipsis. To run on other datasets, just replace the data paths appropriately.

DrQA

• Code based on Facebook's DrQA
• Scripts for preprocessing, training and prediction

QAnet

• Code based on AllenNLP
• AllenNLP configuration file
• Scripts for training and prediction

BERT

• Uses Huggingface's Transformers
• Scripts for training and evaluation