Code for " A Structure Self-Aware Model for Discourse Parsing on Multi-Party Dialogues"(IJCAI2021)
- Molweni. This dataset can be directly preprocessed with our code.
- STAC. This dataset should be formatted first using data_pre.py provided by Shi and Huang in their work "A Deep Sequential Model for Discourse Parsing on Multi-Party Dialogues".
For pre-trained word vectors, this work used GloVe (200d).
Run the model without auxiliary losses, please run
sh script/train.sh
Run our full model with auxiliary losses, please run
sh script/full.sh
Remember to train the teacher first before adding the ''distill'' command
sh script/teacher.sh
For the experiments about STAC and ELECTRA, you can find in this repository