DyGIE++

Implements the model described in the paper Entity, Relation, and Event Extraction with Contextualized Span Representations.

This repository is under construction and we're in the process of adding support for more datasets.

• Dependencies
• Model training
• Model evaluation
• Pretrained models
• Making predictions

This code was developed using Python 3.7. To create a new Conda environment using Python 3.7, do conda create --name dygiepp python=3.7.

The necessary dependencies can be installed with pip install -r requirements.txt.

The only dependencies for the modeling code are AllenNLP 0.9.0 and PyTorch 1.2.0. It may run with newer versions, but this is not guarenteed. For PyTorch GPU support, follow the instructions on the PyTorch.

For data preprocessing a few additional data and string processing libraries are required including, Pandas and Beautiful Soup 4.

Finally, you'll need SciBERT for the scientific datasets. Run python scripts/pretrained/get_scibert.py to download and extract the SciBERT model to ./pretrained.

To train a model for named entity recognition, relation extraction, and coreference resolution on the SciERC dataset:

• Download the data. From the top-level folder for this repo, enter bash ./scripts/data/get_scierc.sh. This will download the scierc dataset into a folder ./data/scierc
• Train the model. Enter bash ./scripts/train/train_scierc.sh [gpu-id]. The gpu-id should be an integer like 1, or -1 to train on CPU. The program will train a model and save a model at ./models/scierc.
• To train a "lightweight" version of the model that doesn't do coreference propagation, do bash ./scripts/train/train_scierc_lightweight.sh [gpu-id] instead. The result will go in ./models/scierc-lightweight. More info on why you'd want to do this in the section on making predictions.

The steps are similar to SciERC.

• Download the data. From the top-level folder for this repo, enter bash ./scripts/data/get_genia.sh.
• Train the model. Enter bash ./scripts/train/train_genia.sh [gpu-id]. The program will train a model and save a model at ./models/genia.

For more information on ACE relation and event preprocessing, see DATA.md and this issue.

We use preprocessing code adapted from the DyGIE repo, which is in turn adapted from the LSTM-ER repo. The following software is required:

• Java, to run CoreNLP.
• Perl.
• zsh. If this isn't available on your system, you can create a conda environment and install zsh.

First, we need to download Stanford CoreNLP:

bash scripts/data/ace05/get_corenlp.sh

Then, run the driver script to preprocess the data:

bash scripts/data/get_ace05.sh [path-to-ACE-data]

The results will go in ./data/ace05/processed-data. The intermediate files will go in ./data/ace05/raw-data.

In progress.

The preprocessing code I wrote breaks with the newest version of Spacy. So unfortunately, we need to create a separate virtualenv that uses an old version of Spacy and use that for preprocessing.

conda deactivate
conda create --name ace-event-preprocess python=3.7
conda activate ace-event-preprocess
pip install -r scripts/data/ace-event/requirements.txt
python -m spacy download en

Then, collect the relevant files from the ACE data distribution with

bash ./scripts/data/ace-event/collect_ace_event.sh [path-to-ACE-data].

The results will go in ./data/ace-event/raw-data.

Now, run the script

python ./scripts/data/ace-event/parse_ace_event.py [output-name] [optional-flags]

You can see the available flags by calling parse_ace_event.py -h. For detailed descriptions, see DATA.md. The results will go in ./data/ace-event/processed-data/[output-name]. We require an output name because you may want to preprocess the ACE data multiple times using different flags. For default preprocessing settings, you could do:

python ./scripts/data/ace-event/parse_ace_event.py default-settings

When finished, you should conda deactivate the ace-event-preprocess environment and re-activate your modeling environment.

In progress.

To check the performance of one of your models or a pretrained model, you can use the allennlp evaluate command.

Note that allennlp commands will only be able to discover the code in this package if:

• You run the commands from the root folder of this project, dygiepp, or:
• You add the code to your Python path by running conda develop . from the root folder of this project.

Otherwise, you will get an error ModuleNotFoundError: No module named 'dygie'.

In general, you can make evaluate a model like this:

allennlp evaluate \
  [model-file] \
  [data-path] \
  --cuda-device [cuda-device] \
  --include-package dygie \
  --output-file [output-file] # Optional; if not given, prints metrics to console.

For example, to evaluate the pretrained SciERC model, you could do

allennlp evaluate \
  pretrained/scierc.tar.gz \
  data/scierc/processed_data/json/test.json \
  --cuda-device 2 \
  --include-package dygie

To evaluate a model you trained on the SciERC data, you could do

allennlp evaluate \
  models/scierc/model.tar.gz \
  data/scierc/processed_data/json/test.json \
  --cuda-device 2  \
  --include-package dygie \
  --output-file models/scierc/metrics_test.json

We have versions of DyGIE++ trained on SciERC and GENIA available. We also have a "lightweight" version of SciERC trained without coreference propagation. See details on this in the caveat below.

Run ./scripts/pretrained/get_dygiepp_pretrained.sh to download all the available pretrained models to the pretrained directory. If you only want one model, here are the download links:

• SciERC
• SciERC lightweight (no coref)
• GENIA

The SciERC model gives slightly better test set performance than reported in the paper:

2019-11-20 16:03:12,692 - INFO - allennlp.commands.evaluate - Finished evaluating.
...
2019-11-20 16:03:12,693 - INFO - allennlp.commands.evaluate - _ner_f1: 0.6855290303565666
...
2019-11-20 16:03:12,693 - INFO - allennlp.commands.evaluate - rel_f1: 0.4867781975175391

The lightweight version of SciERC does slightly worse, but should be more memory-friendly.

2020-03-31 21:23:34,708 - INFO - allennlp.commands.evaluate - Finished evaluating.
...
2020-03-31 21:23:34,709 - INFO - allennlp.commands.evaluate - _ner_f1: 0.6778959810874204
...
2020-03-31 21:23:34,709 - INFO - allennlp.commands.evaluate - rel_f1: 0.4638157894736842

Similarly for GENIA:

2019-11-21 14:45:44,505 - INFO - allennlp.commands.evaluate - ner_f1: 0.7818707451272466

To make a prediction, you can use allennlp predict. For example, to make a prediction with the pretrained scierc model, you can do:

allennlp predict pretrained/scierc.tar.gz \
    data/scierc/processed_data/json/test.json \
    --predictor dygie \
    --include-package dygie \
    --use-dataset-reader \
    --output-file predictions/scierc-test.jsonl \
    --cuda-device 0

Caveat: Models trained to predict coreference clusters need to make predictions on a whole document at once. This can cause memory issues. To get around this there are two options:

• Make predictions using a model that doesn't do coreference propagation. These models predict a sentence at a time, and shouldn't run into memory issues. For SciERC, use scierc-lightweight. To train your own coref-free model, set coref loss weight to 0 in the relevant training config.
• Split documents up into smaller chunks (5 sentences should be safe), make predictions using a model with coref prop, and stitch things back together.

See the docs for more prediction options.

Following Li and Ji (2014), we consider a predicted relation to be correct if "its relation type is correct, and the head offsets of two entity mention arguments are both correct".

In particular, we do not require the types of the entity mention arguments to be correct, as is done in some work (e.g. Zhang et al. (2017)). We welcome a pull request that implements this alternative evaluation metric. Please open an issue if you're interested in this.