Skip to content

AlonEirew/cross-doc-event-coref

Repository files navigation

Cross Document Event Coreferecne Model

This project model code was used in the paper ״WEC: Deriving a Large-scale Cross-document Event Coreference dataset from Wikipedia״ for the cross document event coreference baseline model for WEC-Eng.

WEC-Eng Pre-trained Model

Can be downloaded from huggingface hub: https://huggingface.co/Alon/wec

WEC-Eng Dataset

WEC-Eng can be download from huggingface hub: https://huggingface.co/datasets/biu-nlp/WEC-Eng

See the Dataset card, for instructions on how to read and use WEC-Eng

Prerequisites

  • Python 3.6 or above
  • #>pip install -r requirements.txt
  • #>export PYTHONPATH=<ROOT_PROJECT_FOLDER>

Preprocessing

The main train process require the mentions pairs and embeddings from each set.

Generate Mentions Pairs

ECB+

Project contains datasets/ecb.zip already in the needed input format running the scripts. ECB+ pairs generation for ECB+ test/dev/train sets is straight forward, for example to generate just run:

#>python src/preprocess_gen_pairs.py resources/ecb/dev/Event_gold_mentions.json --dataset=ecb --topic=subtopic
#>python src/preprocess_gen_pairs.py resources/ecb/test/Event_gold_mentions.json --dataset=ecb --topic=subtopic
#>python src/preprocess_gen_pairs.py resources/ecb/train/Event_gold_mentions.json --dataset=ecb --topic=subtopic

WEC-Eng

Since WEC-Eng train set contains many mentions, generating all negative pairs is very resource and time consuming. To that end, we added a control for the negative:positive ratio.
For the Dev and Test sets, as they are much smaller in size,pairs generation is similar to ECB+ (all).

#>python src/preprocess_gen_pairs.py resources/wec/dev/Event_gold_mentions_validated.json --dataset=wec --split=dev
#>python src/preprocess_gen_pairs.py resources/wec/test/Event_gold_mentions_validated.json --dataset=wec --split=test
#>python src/preprocess_gen_pairs.py resources/wec/train/Event_gold_mentions.json --dataset=wec --split=train --ratio=10

Generate Embeddings

To generate the embeddings for ECB+/WEC-Eng run the following script and provide the slit files location, for example:

#>python src/preprocess_embed.py resources/wec/dev/Event_gold_mentions.json resources/wec/test/Event_gold_mentions.json resources/wec/train/Event_gold_mentions.json --cuda=True

Training

See train.py file header for the complete set of script parameters. Model file will be saved at output folder (for each iteration that improves).

  • For training over ECB+:
#> python src/train.py --tpf=resources/ecb/train/Event_gold_mentions_PosPairs.pickle --tnf=resources/ecb/train/Event_gold_mentions_NegPairs.pickle --dpf=resources/ecb/dev/Event_gold_mentions_PosPairs.pickle --dnf=resources/ecb/dev/Event_gold_mentions_NegPairs.pickle --te=resources/ecb/train/Event_gold_mentions_roberta_large.pickle --de=resources/ecb/dev/Event_gold_mentions_roberta_large.pickle --mf=ecb_pairwise_model --dataset=ecb --cuda=True
  • For training over WEC-Eng:
#> python src/train.py --tpf=resources/wec/train/Event_gold_mentions_PosPairs.pickle --tnf=resources/wec/train/Event_gold_mentions_NegPairs.pickle --dpf=resources/wec/dev/Event_gold_mentions_PosPairs.pickle --dnf=resources/wec/dev/Event_gold_mentions_NegPairs.pickle --te=resources/wec/train/Event_gold_mentions_roberta_large.pickle --de=resources/wec/dev/Event_gold_mentions_roberta_large.pickle --mf=wec_pairwise_model --dataset=wec --cuda=True --ratio=10

Inference

See inference.py file header for the complete set of script parameters. Running pairwize evaluation example:

python src/inference.py --tpf=resources/ecb/test/Event_gold_mentions_PosPairs.pickle --tnf=resources/ecb/test/Event_gold_mentions_NegPairs.pickle --te=resources/ecb/test/Event_gold_mentions_roberta_large.pickle --mf=<checkpoint>/ecb_pairwise_modeliter_6 --cuda=True

CD Coreference

Generate Pairs Predictions

Generate the pairs predictions (distance) before running the agglomerative clustering script for final results
See generate_pairs_predictions.py file header for the complete set of script parameters.
Running the pairs prediction algorithm:

python src/generate_pairs_predictions.py --tmf=resources/ecb/test/Event_gold_mentions.json --tef=resources/ecb/test/Event_gold_mentions_roberta_large.pickle --mf=<checkpoint>/ecb_pairwise_modeliter_6 --out=<checkpoint>/ecb_predictions --cuda=True

Clustering

Running agglomerative clustering to get the final cluster configuration on the pairwise predictions. See cluster.py file header for the complete set of script parameters.
Running the pairs prediction algorithm:

python src/cluster.py --tmf=resources/ecb/test/Event_gold_mentions.json --predictions=<checkpoint>/ecb_predictions --alt=0.7

Calculating the CoNLL clustering score

To score our model we used the official CoNLL coreference scorer.
Gold scorer files are at gold_socrer/ecb/* folder.
Usage Example:

#>perl scorer/scorer.pl all gold_scorer/ecb/CD_test_event_mention_dataset.txt <checkpoint>/ecb_pairwise_modeliter_6_0.7 none

Additional Scripts (helper_scripts)

stats.py

Calculate the dataset files statistics (mentions, singleton mentions, clusters...)

python helper_scripts/stats_calculation.py resources/ecb/dev/Event_gold_mentions.json`

visualize.py

Create an HTML page to visualize clusters and mentions from the given set

python helper_scripts/visualize.py resources/ecb/dev/Event_gold_mentions.json --present=cluster`<br/>

Page will be accessed via http://localhost:5000