Home

Awesome

Lifelong Event Detection with Knowledge Transfer

Repo for the EMNLP2021 paper: Lifelong Event Detection with Knowledge Transfer

Data Preprocessing

We are unable to include entire datasets of MAVEN and ACE due to large size of data. We show examples of processed data in the data folded. prepare_inputs.py preprocesses the original MAVEN data. For ACE 2005 the data needs to be processed similarly.

For the subsets that forms incremental tasks, please refer to data/*/streams.json for partitions (integers to label names is in data/*/label2id.json), and run_train.py: PERM for order permutations. Since splits collection includes manual check to make sure all types are covered in development/test data, we don't include scripts for this process. For Silver Negative settings, instances are collected with prepare_stream_instances.py.

Training and Testing

Run the cmd python run_train.py with proper arguments in order to train/test. Valid arguments can be found in utils/options.py. We will release more details in later updates of this repo. By default the script runs evaluation periodically during training. But it is also possible to run testing only using --test-only.

Requirements: