Jointly extracts multiple events: event detection and argument extraction for multiple events in one pass Loosely based on the paper "Jointly Multiple Events Extraction via Attention-based Graph Information Aggregation" (EMNLP 2018)

conda create -n eventx python=3.7
conda activate eventx

pip install -r requirements.txt

Download the preprocssed SD4M+Daystream data from https://dfkide-my.sharepoint.com/:u:/g/personal/lehe02_dfki_de/Ecl1cWZlZVZNg844cZPUdOgB3MGjIapyeCGyDfpx_kylUw. And extract it into the data directory.

In order to train the different models you can use the scripts in the repository. You may need to adjust the configuration file, the training & development data paths and the save paths.

• train_eventx_snorkel.sh: Train a single model.
• random_repeats.sh: Random repeats with different seeds.
• increasing_train_data.sh: Random repeats with different seeds for increasing training data.
• mlv_run.sh: Random repeats with different seeds for Daystream data labeled with Majority Label Voter.

E.g. to train a single model:

./scripts/train_eventx_snorkel.sh data/training_run_1

In order to recreate the models in our main experiments, you need to run random_repeats.sh.

For our other experiments, you need to run increasing_train_data.sh (Increasing Daystream training data) and mlv_run.sh (Majority Label Voter).

For convenience we provide the trained event extraction model for the setup with the merged training set (manual annotation + annotation with the Snorkel approach): https://dfkide-my.sharepoint.com/:u:/g/personal/lehe02_dfki_de/ESEeQWkkzRlKpPl95e8HSeEBJdl1W9N2e4d246onXQmJiQ