This is a reimplementation in Pytorch of the neural mention-ranking system by Sam Wiseman et al. (ACL 2015).

Here's how to use it:

1. For corpus preprocessing, you need the tools from Wiseman's original implementation, which you can in the modifiedBCS subdirectory of his Github repository. Follow the instructions there to generate the txt features files.

2. Convert the txt files into our own HDF5 format (which is not the same as Wiseman's) by running features.py. Note that the main function in features.py contains some hardcoded paths that you need to adapt to your system.

3. Train the model with

   python mention_ranking.py --train training.h5 --dev dev.h5 --checkpoint OUTPREFIX --train-config TRAIN_CONFIG --net-config NET_CONFIG
   

   Here, OUTPREFIX is the file name prefix of the model files that will be saved after each epoch. TRAIN_CONFIG and NET_CONFIG are JSON files to set up the training process and network configuration. If you leave out these options, the defaults in mention_ranking.py will be used. There you can also find the options that can be set in these files.

4. Create predictions with

   python mention_ranking.py --predict test.h5 --model MODEL_FILE
   

   MODEL_FILE is one of the checkpoint files created during the training run. The predictions will be output to stdout in the form of a backpointer file that can be processed with modifiedBCS/WriteCoNLLPreds.sh in Wiseman's repository.