This repository contains the implementation of the model presented at the "Zero-shot transfer for implicit discourse relation classification" paper.
You can prepare the data for the model via "prepare_data.sh" script. This scripts only needs the directory of PDTB3 annotations. ./prepare_data.sh pdtb3_annotations # where pdtb3 annotations are saved under the directory "pdtb3_annotations"
As a result, the extracted sentences as well as their laser embeddings will be saved to data/text data/embed respectively. Note: Since Ted-MDB annotations are publicly available, that part of the script is self-contained
To train the model, you can simply run the "run.sh" script as follows:
./run.sh pdtb3 saved_models
This script will
- train a separate "one vs. all" classifier for each sense using the PDTB3 annotations
- save the models under "saved_models/" dir
- report the performance of the model on all Ted-MDB languages as well as PDTB3 test set.