USABILITY: This branch is an ongoing update of the DyGIE++ code to play nicely with AllenNLP V1 (and by extension, Huggingface Transformers). The relation, named entity, and coreference modules should work, but event extraction will break. The performance may be a point or two lower on relation extraction due to tuning issues. I will update this status as changes are made.

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 on existing datasets
• Working with new datasets
• Contact

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.

This library relies on AllenNLP and uses AllenNLP shell commands to kick off training, evaluation, and testing.

Warning about coreference resolution: The coreference code will break on sentences with only a single token. If you have these in your dataset, either get rid of them or deactivate the coreference resolution part of the model.

We rely on Allennlp train to handle model training. The train command takes a configuration file as an argument, and initializes a model based on the configuration, and serializes the traing model. More details on the configuration process for DyGIE can be found in doc/config.md.

To train a model, enter bash scripts/train.sh [config_name] at the command line, where the config_name is the name of a file in the training_config directory. For instance, to train a model using the scierc.jsonnet config, you'd enter

bash scripts/train.sh scierc

The resulting model will go in models/scierc. For more information on how to modify training configs (e.g. to change the GPU used for training), see config.md.

Information on preparing specific training datasets is below. For more information on how to create training batches that utilize GPU resources efficiently, see model.md

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.sh scierc.
• To train a "lightweight" version of the model that doesn't do coreference propagation and uses a context width of 1, do bash scripts/train.sh scierc_lightweight instead. 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 genia.
• As with SciERC, we also offer a "lightweight" version with a context width of 1 and no coreference propagation.

The ChemProt corpus contains entity and relation annotations for drug / protein interaction. The ChemProt preprocessing requires a separate environment:

conda deactivate
conda create --name chemprot-preprocess python=3.7
conda activate chemprot-preprocess
pip install -r scripts/data/chemprot/requirements.txt

Then, follow these steps:

• Get the data. Run bash ./scripts/data/get_chemprot.sh. This will download the data and process it into the DyGIE input format.
  • NOTE: This is a quick-and-dirty script that skips entities whose character offsets don't align exactly with the tokenization produced by SciSpacy. We lose about 10% of the named entities and 20% of the relations in the dataset as a result.
• Train the model. Enter bash scripts/train chemprot.

For more information on ACE relation and event preprocessing, see doc/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.

Enter bash scripts/train ace05_relation. A model trained this way will not reproduce the numbers in the paper. We're in the process of debugging and will update.

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.

Enter bash scripts/train ace05_event. A model trained in this fashion will reproduce (within 0.1 F1 or so) the results in Table 4 of the paper. To reproduce the results in Table 1 requires training an ensemble model of 4 trigger detectors. The basic process is as follows:

• Merge the ACE event train + dev data, then create 4 new train / dev splits.
• Train a separate trigger detection model on each split. To do this, modify training-config/ace05_event.jsonnet by setting

  model +: {
    modules +: {
      events +: {
        loss_weights: {
          trigger: 1.0,
          arguments: 0.5
        }
      }
    }
  }

• Make trigger predictions using a majority vote of the 4 ensemble models.
• Use these predicted triggers when making event argument predictions based on the event argument scores output by the model saved at models/ace05_event.

If you need more details, email me.

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/normalized_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/normalized_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. There are two versions:

• The "lightweight" versions don't use coreference propagation, and use a context window of 1. If you've got a new dataset and you just want to get some reasonable predictions, use these.
• The "full" versions use coreference propagatation and a context window of 3. Use these if you need to squeeze out another F1 point or two. These models take longer to run, and they may break if they're given inputs that are too long.

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
• GENIA
• GENIA lightweight
• ChemProt (lightweight only)
• ACE05 event (uses BERT large)

• SciERC

  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
  

• SciERC lightweight

  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
  

• GENIA

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

• GENIA lightweight And the lightweight version:

  2020-05-08 11:18:59,761 - INFO - allennlp.commands.evaluate - ner_f1: 0.7671077504725398
  

• ChemProt

  2020-05-08 23:20:59,648 - INFO - allennlp.commands.evaluate - _ner_f1: 0.8850947021684925
  2020-05-08 23:20:59,648 - INFO - allennlp.commands.evaluate - rel_f1: 0.35027598896044154
  

  Note that we're doing span-level evaluation using predicted entities. We're also evaluating on all ChemProt relation classes, while the official task only evaluates on a subset (see Liu et al. for details). Thus, our relation extraction performance is lower than, for instance, Verga et al., where they use gold entities as inputs for relation prediction.

• ACE05-Event

  2020-05-25 17:05:14,044 - INFO - allennlp.commands.evaluate - _ner_f1: 0.906369532679145
  2020-05-25 17:05:14,044 - INFO - allennlp.commands.evaluate - _trig_id_f1: 0.735042735042735
  2020-05-25 17:05:14,044 - INFO - allennlp.commands.evaluate - trig_class_f1: 0.7029914529914529
  2020-05-25 17:05:14,044 - INFO - allennlp.commands.evaluate - _arg_id_f1: 0.5414364640883979
  2020-05-25 17:05:14,044 - INFO - allennlp.commands.evaluate - arg_class_f1: 0.5130228887134964
  

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 \
    --silent

The predictions include the predict labels, as well as logits and softmax scores. For more information see, docs/data.md.

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. Use the "lightweight" models to avoid this. 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.

Follow the instructions as described in Formatting a new dataset.

To make predictions on a new, unlabeled dataset:

1. Download the pretrained model that most closely matches your text domain.
2. Make predictions the same way as with the existing datasets:

allennlp predict pretrained/[name-of-pretrained-model].tar.gz \
    [input-path] \
    --predictor dygie \
    --include-package dygie \
    --use-dataset-reader \
    --output-file [output-path] \
    --cuda-device [cuda-device]

Follow the process described in Training a model, but adjusting the input and output file paths as appropriate.

For questions or problems with the code, create a GitHub issue (preferred) or email [email protected].