joint-coref-srl

Code to reproduce experiments from "Joint Semantic Analysis with Document-Level Cross-Task Coherence Rewards"

Recommended: Python 3.7 with a new conda environment

Installation

Install all dependencies with pip install -r requirements.txt

Then navigate to the allennlp folder and run the following commands

$ pip install --editable .
$ pip install -r dev-requirements.txt

Install pytorc-geometric by following this post

Training

• Download Ontonotes 5.0 and run prepare_ontonotes.sh to convert it into the CoNLL-2012 format
• Make sure that the concatenated files are in a folder named data/conll-2012_single_file with names like test.english.gold_conll (without the version number) and the original split files are in a folder named data/conll-2012
• Run mkdir graphs && python combine_data.py to combine the gold (train and development) SRL tags and corefernce clusters into a single JSON object. This should create 2 files, one for training and development each
• To further convert them into processable graphs, run python graph_util/output_to_graph.py <JSON-path-of-graphs> (once for train and again for dev.)
• Finally, train the coherence classifiers by running dgi.py

Now you are all set to train the coreference and SRL models!

All the configuration files can be found in the configs folder:

• single: contains the configs for the single-task baselines
• mtl: contains the configs for the multi-task baselines
• ft: contains the configs to finetune the different models (each encoder type has a different config file)

To train the baseline models, run

$ python train.py --config_file_path <path-to-config> --serialization_dir <path-to-save>

To finetune the models, run

$ python predict.py <path-to-ft-config> <reward-type> <graph-encoder-type> <ft-task-name> <ft-dataset> <path-to-model-dir>

In all our experiments we use reward-type=sep and graph-encoder-type=GCN.

Evaluation

To evaluate a model on a dataset, run

$ python evaluate.py -s <path-to-model-dir> -t <task-name> -d <dataset-name> -m <type-of-model>

type-of-model can be pt or ft for evaluating pre-trained and fine-tuned models respectively.

For all files, running it with --help flag prints the various options and their help strings.

Thanks

This code is built on HMTL

Citation

@misc{aralikatte2020joint,
      title={Joint Semantic Analysis with Document-Level Cross-Task Coherence Rewards}, 
      author={Rahul Aralikatte and Mostafa Abdou and Heather Lent and Daniel Hershcovich and Anders Søgaard},
      year={2020},
      eprint={2010.05567},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}