Dependencies: numpy Python 3.6+ ELMo Pytorch 1.0.1 (with CUDA support if you want to use a GPU) sklearn-crfsuite (for CRF sequence tagger)

Setup:

./setupy.py
pip install .

1. Convert data to a csv file according to the header in data_tools/data_scheme OR - extend data_adapter.py and create your own adapter to load and write to your desired data format.

2. Create a config file including parameters for the models you want to run. See config_example_* for examples.

   Stage options include:

   sequence (time and event tagging using BIO labels): model options: crf

   ordering (listwise temporal ordering): model options: neurallinear (RNN), neural (Set2Seq)

   classification

3. python chrononet.py --config [CONFIG].ini

Temporal ordering models from the 2019 paper (see below) are in models/ordering/pytorch_models.py. The 'linear' model is called GRU_GRU, and the set2sequence model is called Set2SequenceGroup.

If you use this code, please cite the following paper:

Serena Jeblee, Frank Rudzicz, Graeme Hirst. 2019. Neural temporal ordering of events in medical narratives with a set-to-sequence model. In Proceedings of Machine Learning for Health Workshop (ML4H 2019). (In press).

Previous papers

Serena Jeblee, Graeme Hirst. 2018. Listwise temporal ordering of events in clinical notes. In Proceedings of the Ninth International Workshop on Health Text Mining (LOUHI 2018). (In press).