Structure Self-Aware

Code for " A Structure Self-Aware Model for Discourse Parsing on Multi-Party Dialogues"(IJCAI2021)

Dataset

• 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".

Word Vectors

For pre-trained word vectors, this work used GloVe (200d).

Run

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

Others

For the experiments about STAC and ELECTRA, you can find in this repository